Masayoshi Son

1. Snapshot and testable edge thesis

Masayoshi Son's defensible edge is a repeatable sequence, not clairvoyance: hold the mission constant; name the next technology regime; locate the scarce distribution, infrastructure, intellectual-property, or capital layer; select an ambitious leader; then choose minority influence, operating control, or a scaled fund vehicle to match the work. The pattern is visible from Internet portals and broadband through mobile, Arm, the Vision Funds, and the current AI stack. [researcher inference · strong] FY1998 report, printed pp. 4–5 FY2001 report, printed pp. 2–5 Arm rationale, paragraphs 2–7 FY2026 message, “Securing the four corners”

The thesis is falsifiable. It predicts decision-date regime maps, access to scarce assets before consensus, ownership rights matched to the intervention, independently measured operating transfer, and attractive outlier-removed cash returns after price, financing and dilution. The public record only partly passes. Alibaba is a dominant company-reported SBG outcome; Arm is strategically central but materially marked and retained; Sprint and WeWork show that a correct regime can coexist with poor price, governance, financing, or execution. Only 11 of 534 company-family keys are sufficiently reconstructed: each has chronology plus material ownership, structure, governance, financing or reported-economics evidence. That threshold is not a full decision/economics reconstruction, so no career hit rate or personal IRR is supportable. [researcher inference · strong] denominator reconciliation dominant-outlier tests FY2022 held-position disclosures, printed pp. 98, 101–103 and 109–110

The strongest test is therefore prospective: freeze the regime, security, price, ownership, team, financing, milestones, falsifiers, and stop rule before approval; then compare actual cash and remaining marked value with the memo and a matched opportunity set. If the method disappears outside famous winners, or if ex-Alibaba and outlier-removed cohorts do not beat comparable alternatives, the edge thesis fails. [researcher inference · strong] operational rules Sprint ex-ante transcript, lines 701–787 and 1,065–1,264 FY2022 Q1 transcript, A9

2. Core philosophy

Keep purpose fixed and technology flexible

Son's most stable doctrine is “information revolution—happiness for everyone,” while products, technologies, and organizational forms must be cannibalized before they become obsolete. The 2010 vision explicitly separates a durable purpose from any thirty-year technology forecast and favors a self-renewing federation over attachment to one product. [investor-stated · strong] Next 30-Year Vision, PDF pp. 89–99 WIRED/Reuters report, paragraphs 74–100

Backcast from a future state, then identify the control point

The historical method starts with a long-horizon state—Internet traffic, broadband adoption, mobile data, connected devices, AI—and works backward to the layer that constrains participation or captures coordination value. Portals aggregate attention; broadband and mobile networks control access; Arm licenses foundational compute IP; the 2026 AI map links models, chips, data centers, agents, and robotics. [researcher inference · strong] FY1999 report, printed pp. 3–5 FY2026 report, printed pp. 23–25

That is a search heuristic, not valuation evidence. A regime can arrive later, value can accrue elsewhere, and a strategically central layer can still earn poor returns after capital intensity, competition, regulation, or an excessive entry price. [researcher inference · strong] historical context FY2015 selected scorecard, printed p. 26 / PDF p. 28 Japan Telecom impairment, sections 1–2

Prefer category leaders and ambitious founders—but use a second key

Son repeatedly describes selecting founders with conviction, recruiting force, and category ambition, then preserving operating autonomy. The Alibaba story is the iconic retrospective example; the Vision Fund later formalized the category-leader screen. [retrospective witness · provisional] Tokyo Forum, 03:14–17:11 Milken, 24:42–29:27

The second key is price, security, governance, and survivability. Son's 2020 WeWork admission and 2022 defense-mode postmortem make explicit what the winner stories omit: the best company can be a poor investment if conviction overwhelms valuation or control. [investor-stated · strong] Forbes, WeWork admission FY2022 Q1 transcript, A9

Protect the right to wait

Long horizons are usable only while the owner can finance them. The 2009–2010 record treats free cash flow and debt reduction as option value; the 2022 record adds normal and emergency LTV limits, a cash buffer, and slower deployment. These are not side constraints on the philosophy. They determine whether a thesis can survive a closed exit window. [documented behavior · strong] FY2009 report, printed pp. 3–7 FY2022 message, “Defense Mode” and LTV policy

3. Sourcing

Son's visible sourcing begins top-down. He selects a technological regime, maps its leaders and bottlenecks, and seeks assets that can change SoftBank's strategic position. Yahoo U.S. combined observed usage growth with an attention gateway; Yahoo Japan added a local operating partner; Vodafone Japan and Sprint supplied controlled distribution; Arm supplied foundational IP; the Vision Funds expanded access to late-stage category leaders. [researcher inference · strong] Yahoo U.S. lab Sprint SEC-filed transcript

The second channel is relational and team-mediated. The 1999 Forbes account attributes SoftBank Technology Ventures' operating work to Gary Rieschel and describes a network of alliances and follow-ons; Yahoo Japan depended on Jerry Yang/Yahoo and Masahiro Inoue's local execution; later fund teams, partners, founders, banks, and LPs widened the funnel. This record does not justify calling every opportunity “Son-sourced.” [contemporaneous record · moderate] Forbes, Yahoo/STV and alliance passages people map

The public gap is a sourcing denominator. No disclosed ledger shows inbound versus outbound, introducer, rejected themes, first meeting, competing opportunities, source-to-IC conversion, or performance by channel. A usable reconstruction should record those fields, conflicts and which entity had lawful capacity to invest. [researcher inference · strong] retrieval ledger FY2022 position lists, printed pp. 98, 101–103 and 109–110

Sourcing checklist

1. Name the regime change and the dated evidence that makes it investable now.
2. Map the scarce layer, leading substitutes, regulators, customers, suppliers, and capital providers.
3. Record who introduced the company, who owns the relationship, and any economic interest.
4. Separate the local operator's edge from SoftBank's brand, balance sheet, and later narration.
5. Log every serious pass and rejected alternative before the outcome is known.

This is an analyst-created implementation of the evidence, not a recovered SoftBank template. [researcher inference · strong] writings playbook

4. Picking

Regime, company, and security are three decisions

The regime must have independent adoption evidence; the company must show durable capture; the security must offer an acceptable price, ownership path, governance, financing burden, and exit. Son's public record is strongest on the first question, mixed on the second, and thinnest on the third. [researcher inference · strong] writings playbook failure postmortems WeWork S-1, risk factors, printed pp. 49–51 FY2022 Q1 transcript, A9

Founder and team

Test ambition, speed, recruiting force, integrity, learning rate, and willingness to surface bad news. The later Alibaba accounts support Son's founder heuristic as testimony, but not as a base rate: the rejected-founder denominator and original memo are unavailable. WeWork demonstrates why magnetism cannot replace related-party, voting, board, and capital-dependence analysis. [researcher inference · strong] Alibaba lab WeWork S-1, printed pp. 49–51

Product, market, and control point

Require product evidence that the company—not merely the category—captures the scarce layer. Measure retention, willingness to pay, unit economics, technical performance, switching costs, supplier/customer concentration, and regulatory permission. “No. 1” must be defined by a dated metric and geography; local network effects can justify multiple leaders, but also reveal factor concentration. [researcher inference · strong] Vision Fund screen, Milken 24:42–29:27 portfolio factor analysis

Price, ownership, governance, and financing

Write expected return at entry across base, upside, and downside cases; identify the legal owner and security; model dilution and follow-ons; name votes, vetoes, board seats, information rights, related parties, founder control, covenants, capital calls, and refinancing. The corpus supports no universal Son hurdle rate, ownership target, position ceiling, or loss ratio. [researcher inference · strong] vehicle map FY2026 report, IC and risk, printed pp. 54 and 101–102

Pre-IC decision card

Before approval, freeze:

1. regime claim, observed evidence, forecast horizon, probability, and falsifier;
2. bottleneck, substitutes, value-accrual mechanism, and synergy counterfactual;
3. founder/team references, attribution, governance risks, and bad-news channel;
4. security, cost, ownership, dilution, valuation bridge, and expected return;
5. vehicle, decision rights, geographic authority, financing, and stakeholder conflicts;
6. first-check size, reserve, next milestone, follow-on gate, and stop rule; and
7. exit routes, duration, regulatory path, cash downside, and benchmark.

[researcher inference · strong] operational rules

5. Deal and portfolio mechanics

Early Yahoo evidence supports small entry followed by larger commitment as evidence improved. That is not a universal historical rule: Vodafone Japan, Sprint, Arm, SVF late-stage rounds, and OpenAI use much larger and less reversible commitments. Staging is valid only when the milestone changes risk and the vehicle retains the legal and financial ability to stop. [researcher inference · strong] Forbes, Yahoo/STV passages OpenAI 2026 release, sections 1–5

Ownership should follow the work:

• use a small minority stake for learning and optionality;
• use a material minority plus information/board rights when founder autonomy and influence are sufficient;
• buy control only when operating transfer or ownership of a bottleneck is necessary, jurisdictionally feasible, and financed through downside;
• use a fund only when its duration, LP rights, fees/carry, reserves, concentration limits, and valuation process fit the strategy.

[researcher inference · strong] Yahoo Japan lab Vodafone Japan lab SVF1 terms, FY2022 printed pp. 91–93

Portfolio construction must be factor-aware rather than company-count-aware. Mobility across Uber, DiDi, Grab and Ola is geographic breadth but a shared category/exit factor; the 2025–2026 Arm–OpenAI–Ampere–Stargate program connects several assets but concentrates AI adoption, compute, power, financing, execution, and regulatory risk. [researcher inference · strong] portfolio forensics FY2026 report, printed pp. 23–25 and 80–83

Every portfolio review needs separate operating, mark, and cash ledgers. Report security-level contributions, follow-ons, realized proceeds, retained fair value, transfers, debt, fees/carry, tax, currency, and outside interests. Then show vintage cohorts, loss ratio, concentration, reserves, and results excluding the largest one and two outcomes. The 830 public observations and 534 canonical families are a discovery universe, not an investable or economic denominator; only 11 families are sufficiently reconstructed. [researcher inference · strong] denominator reconciliation reported snapshots FY2015 selected scorecard, printed p. 26 / PDF p. 28 FY2025 gross public-company tables, pp. 78–79

6. Vehicle and incentive fit

“SoftBank invested” can mean Son personally, SBG parent, SBGC or another subsidiary, a telecom operating company, Yahoo Japan/LY, SVF1/SBIA, SVF2/SBGA, the LatAm funds, SB Northstar, MASA USA/MgmtCo, or an external LP/co-investor. These are different owners, beneficiaries, decision systems, durations, and return perimeters. [contemporaneous record · strong] vehicle and economics map FY2021 Northstar disclosure, printed p. 82 / PDF p. 86 FY2022 fund and co-investment terms, printed pp. 68–69 and 91–93

7. Team attribution and work with founders

Son can be credited with his stated mission, public regime thesis, sponsorship of named transactions, and formal votes where disclosed. He cannot automatically be credited with sourcing, diligence, negotiation, board labor, operating execution, or company value creation. SBG and its vehicles supplied capital; founders and employees built companies; operating executives and fund teams performed distinct work. [researcher inference · strong] people and influence map FY2017 organization and fund roles, printed pp. 14–15 FY2026 Investment Committee, printed p. 54

The dated map matters. Gary Rieschel led STV's operating model, while Masahiro Inoue led Yahoo Japan. [contemporaneous record · moderate] Forbes, STV operating passages Yahoo 1999 10-K, Note 6 Masayoshi Son and Yoshimitsu Goto occupy different current sponsorship/financing roles; Nikesh Arora carried a dated global-strategy and succession remit; Ken Miyauchi and Marcelo Claure had telecom operating roles. [contemporaneous record · strong] OpenAI follow-on release, sections 1–5 2015 Arora appointment FY2015 telecom roles, printed pp. 6–7 FY2016 Sprint roles, printed pp. 9–11 Rajeev Misra had dated Vision Fund roles; SBIA/SBGA teams managed, valued, and supported investments; external LPs supplied capital and contractual governance; the current SBG IC has five members. [contemporaneous record · strong] FY2017 fund roles, printed pp. 14–15 FY2022 manager and operating-team roles, printed pp. 68–69 and 91–93 SVF1 first close, investor and governance terms FY2026 Investment Committee, printed p. 54

The founder-working model is conditional autonomy: supply capital, strategic access, recruiting, customers, financing, and sometimes operating transfer while leaving the entrepreneur room to lead. The model works only when autonomy is paired with information rights, bad-news channels, related-party controls, milestone review, and a precommitted intervention/stop rule. Yahoo Japan is the positive localization case; WeWork is the warning that economic exposure without voting control can leave responsibility and authority misaligned. [researcher inference · strong] Yahoo Japan lab WeWork lab Yahoo 1999 10-K, Note 6 WeWork S-1, risk factors, printed pp. 49–51

For every deal, maintain a role ledger: sourced; introduced; sponsored; diligenced; voted; negotiated; funded; served on board; recruited; provided customers/technical help; approved follow-ons; managed crisis; influenced exit; narrated later. Unknown stays unknown, and legal decision rights do not prove labor. [researcher inference · strong] people and influence map Sprint ex-ante role record FY2026 delegated approval record, printed p. 54

8. Views on the craft

Son treats investing as future design rather than passive security selection. The craft combines technology forecasting, founder judgment, corporate strategy, financing, operating intervention, and institution building. This broader remit explains why an Internet portal, a telecom acquisition, an IP platform, a fund, and an AI-infrastructure consortium can belong to one narrative—but it also makes a single venture-return score misleading. [researcher inference · strong] writings synthesis Next 30-Year Vision, PDF pp. 89–99 FY2026 stack map, printed pp. 23–25

Power-law economics make access to rare companies valuable, but fund size does not manufacture selection skill. Larger pools widen the check-size frontier and can change company runway; they also raise the amount each winner must return, increase price pressure, consume governance bandwidth, and correlate marks and exits. A great investor must therefore show both access and disciplined conversion of access into security-level, outlier-tested cash economics. [researcher inference · strong] Pareto and outlier views SVF1 terms, FY2022 printed pp. 91–93

Governance and exit are parts of picking, not aftercare. Rights must match the work; continued ownership must beat selling after financing, fund life, regulation, and opportunity cost; and company success must be separated from who sourced, governed, funded, and received the return. On this record, greatness means recognizing a durable regime early, securing the value-capture layer on defensible terms, helping the institution execute, surviving long enough to realize value, and producing repeatable ex-outlier economics—not merely naming the future. [researcher inference · strong] team and authority map exit modes and duration Arm prospectus, “Principal and Selling Shareholder” FY2025 gross public-company tables, pp. 78–79

Speed is valuable after preparation. The 2016 Arm record describes a decade-watched asset and rapid action once available; the 1990s Yahoo pattern shows learning before expansion. Speed without a frozen price, rights, financing, and downside memo is merely compressed error. [researcher inference · strong] SoftBank World 2016 official English video, 02:03–05:38 Yahoo U.S. lab

Ambition is an organizational technology, not a probability. Backcasting can align resources, while extreme forecasts and round-number NAV goals require explicit baselines, scenarios, confidence intervals, and failure dates. The movement from a roughly thirty-year singularity horizon to near-term AGI and ten-year ASI is a reason to strengthen forecast scoring, not proof of acceleration. [researcher inference · strong] talks, forecast timing FY2026 message

Humility has to be procedural. Admissions after WeWork and 2022 defense mode are valuable, but learning is credible only when it changes position limits, price discipline, follow-on gates, governance, risk budgets, and stop authority in records created before the next outcome. [researcher inference · strong] documented process changes FY2022 Q1 transcript, A4–A5 and A9 FY2026 Investment Committee, printed p. 54

9. Signature bets and decision process

Yahoo U.S. and Yahoo Japan

Yahoo U.S. is the cleanest early example of observing an attention gateway, making a small commitment, and increasing exposure; Yahoo Japan adds local partnership, meaningful ownership, and operating autonomy. These cases support staged conviction and localization, but not a general formula: complete cash-flow ledgers and rejected-opportunity cohorts remain absent. [researcher inference · strong] Yahoo U.S. lab Yahoo Japan lab 1995 announcement, investment and users 1996 follow-on announcement, stake and traffic Yahoo 1999 10-K, Notes 6–7

Alibaba

Alibaba is the strongest founder-selection story and the dominant company-reported SBG outcome. SBG reported ¥9.7T total return and 57% IRR through 2023-05-11, but those figures belong to the disclosed SBG perimeter, not Son personally or a complete fund-net ledger. The original memo, exact initial security, rejected-founder denominator, and full cash-flow attribution remain incomplete. [contemporaneous record · strong] Alibaba lab Pareto economic view FY2023 CFO message, Alibaba return perimeter

Yahoo! BB and Vodafone Japan

Yahoo! BB built the broadband access layer through subsidized acquisition and losses; Vodafone Japan bought national mobile distribution through a layered, leveraged control transaction. Together they show why a bottleneck thesis becomes an operating and financing problem. Their lesson is not “always buy control,” but “buy control only when the transfer plan and downside financing are explicit.” [researcher inference · strong] Yahoo! BB lab Vodafone Japan lab FY2003 broadband operations Vodafone acquisition, sections 1–5

Sprint to T-Mobile

The 2012 SEC-filed materials freeze the ex-ante mobile-data, procurement, network, and turnaround logic. Later T-Mobile consideration is economically meaningful but includes securities, contingent rights, financing, and a long regulatory path; a later company-calculated IRR cannot replace the transaction ledger. Sprint tests the portability of an operating playbook across countries and the value of an exit route not controlled at entry. [contemporaneous record · strong] Sprint lab SEC transcript

Arm

Arm is the clearest strategic-control case: foundational IP across computing waves, acquired outright, partly transferred to SVF1, subjected to a failed NVIDIA sale, relisted, and largely retained. It is both a powerful thesis result and a reminder that marked retained value, intra-group transfers, IPO proceeds, deposits, tax, and financing cannot be collapsed into cash return. [contemporaneous record · strong] Arm lab 2024 AGM presentation, p. 15

Vision Fund and WeWork

The Vision Fund made capital scale and speed part of the offering to late-stage AI/category leaders. WeWork is the decisive countercase: high economic exposure, weak voting control, repeated follow-ons, and a severe valuation reset. Son's admission that valuation and founder trust were wrong is important, but no public precommitment shows when SoftBank should have stopped. [researcher inference · strong] Vision Fund rationale, Milken 24:42–29:27 WeWork lab

NVIDIA

NVIDIA supports early recognition of a compute platform and a profitable SBG/SVF1 transaction, not the counterfactual that a longer hold was feasible. FY2019 discloses a $2.834B SBG purchase, September 2017 SVF1 transfer, collars, January 2019 exit, and a reported $2.915B gain; later regret conflicts with the documented sale chronology and omits authorization, vehicle, and liquidity constraints. [contemporaneous record · strong] NVIDIA lab FY2019 report, printed p. 66

OpenAI, Stargate, and the ASI stack

The current program applies the full architecture—model partnership, chips, data centers, agents, and robotics—but combines minority preferred equity, a control acquisition, corporate bridge capacity, and consortium infrastructure. It has no realized return, and announced maximums, funded tranches, project capacity, partner economics, power, utilization, and stop rules must remain separate. [contemporaneous record · strong] OpenAI/Stargate lab FY2026 report, printed pp. 23–25 and 80–83

10. Mental models and maxims

The 16 promoted models below are the complete reconciled inventory. Their detailed provenance, exceptions, falsifiers, and 2026 translations belong in mental-models.md; none is a claim that Son published this exact wording.

1. Keep the mission fixed; cannibalize the method — origin: FY1998, printed pp. 2–5 and WIRED/Reuters, paragraphs 74–79.
2. Backcast from a future state, but make forecasts falsifiable — origin: 2010 vision, PDF pp. 1, 30–32 and 125–127.
3. Name the regime before the company — origin: FY1999, printed pp. 2–5.
4. Own or influence the bottleneck where value can accrue — origin: FY2001, printed pp. 2–5.
5. Use the time machine through local operators, not copy-paste — origin: Yahoo 1999 10-K, Note 6.
6. Back ambitious category leaders with a separate price-and-governance key — origin: 2018 message, “Creating a unique synergy group structure”.
7. Stage conviction; re-underwrite every follow-on and precommit the stop — origin: Forbes, Yahoo/STV passages.
8. Match ownership and control to the work required — origin: FY1999, printed pp. 3–5.
9. Write the operating-transfer plan before a control acquisition — origin: 2006 acquisition release, sections 1–5 and FY2007, pp. 13–20.
10. Treat capital scale and speed as a product inside vehicle and price limits — origin: SVF1 first close.
11. Build ecosystems around scarce infrastructure, but make synergy earn a number — origin: Forbes, alliance passages.
12. Separate adoption from market price, but never from financing survival — origin: FY2002, printed pp. 2–3.
13. Attach every return numerator and denominator — origin: FY2017, Note 48.
14. Underwrite duration and exits separately from entry conviction — origin: 2001 Yahoo sale release and Alibaba prospectus.
15. Treat internal data and experiments as evidence only after outcome measurement — origin: SoftBank World 2013 deck, pp. 33–42 and 47–58.
16. Institutionalize attribution, decision rights, and succession — origin: Forbes, STV passages and FY2017 governance, printed pp. 70–73.

[researcher inference · strong] Regret minimization is not promoted because the visible examples are isolated and outcome-aware. Probability trees are a missing corrective folded into falsifiable backcasting, not a Son-stated standalone model. writings decision playbook Milken Japan, Amazon regret at 08:48–10:37 Next 30-Year Vision, backcasting at PDF pp. 30–32 and 125–127

11. Evolution and contradictions

The recurring tensions should travel with the method: autonomy versus intervention; category leadership versus valuation; long horizon versus forced liquidity; operating earnings versus NAV; debt elimination versus leverage; ecosystem breadth versus factor concentration; humane AI aspiration versus operational safeguards; and a self-renewing institution versus Son-centered succession. [researcher inference · strong] writings contradiction ledger talks counterevidence FY2022 Q1 transcript, A4–A5 and A9 46th AGM presentation, pp. 6–8 and 62

12. Failures and limits

• WeWork: category/founder conviction overwhelmed valuation, governance, related-party, lease, and financing evidence; later support blurred rescue and new investment. FY2025 reports $6.71B of reclassified losses across SVF1/SVF2, not a Son-personal loss. [contemporaneous record · strong] case lab FY2025 report, printed p. 91
• DiDi: category leadership did not neutralize data, listing, governance, and regulatory risk; the FY2025 public table showed 0.3x on the disclosed gross perimeter. [contemporaneous record · strong] case lab DiDi 20-F, Risk Factors and principal shareholders FY2025 results, p. 78
• Greensill/Katerra and OneWeb: ecosystem adjacency and internal relationships did not substitute for counterparty, financing, collateral, or independent survival evidence. Complete security-level loss and personal-attribution ledgers remain unavailable. [researcher inference · strong] Greensill lab OneWeb lab UK Treasury evidence, Q189–Q199 and Q293 OneWeb loss, printed p. 14 / PDF p. 18
• Sprint: the strategic logic was contemporaneous, but turnaround, leverage, regulation, and exit took a different path from the Vodafone precedent. Later consideration should not be read as clean validation of the entry plan. [researcher inference · strong] Sprint lab Sprint ex-ante transcript, lines 701–787 and 1,065–1,264 2020 AGM result perimeter, pp. 34–35
• NVIDIA: a profitable exit can still be a duration regret, but later counterfactual storytelling cannot establish that holding was feasible. [researcher inference · strong] NVIDIA lab FY2019 NVIDIA note, printed p. 66 / PDF p. 68
• Ordinary outcomes: FY2025 public rows such as Uber 1.2x, Grab 0.8x, and Paytm 0.7x prevent the winner/failure binary from hiding the middle of the distribution. They are gross disclosed snapshots, not fund-net returns. [contemporaneous record · strong] FY2025 results, p. 78
• Amazon and other misses: later regret can identify a blind spot, but does not prove a feasible security, allocation, or return. [retrospective witness · provisional] Amazon lab Milken Japan, 08:48–10:37

The public process changes—leadership intervention, overseas-team consolidation, cost reduction, lower deployment, cash buffers, and LTV policy—are documented. A company-by-company revised IC template, position-limit schedule, price rule, reserve rule, and causal before/after cohort are not. [researcher inference · strong] documented later changes FY2019 Q2 deck, slides 20–29 FY2022 Q1 transcript, A4–A5 and A9

13. Historical operating environment

1995–2000: Internet distribution was scarce, public adoption rose rapidly, and small stakes/alliances offered cheap strategic optionality before bubble valuations and failures exposed selection and realization risk. The correct comparator is the period's portal, publishing, trade-show, and venture alternatives—not today's platform maturity. [researcher inference · strong] 1995 Yahoo announcement FY2015 scorecard, printed p. 26

2001–2006: broadband and mobile distribution required operating losses, infrastructure, spectrum, financing, and control. Yahoo! BB and Vodafone Japan should be compared with telecom builds and leveraged acquisitions, not seed venture. [researcher inference · strong] FY2006 report, printed pp. 6–10

2012–2016: Sprint and Arm moved the same control-point heuristic into cross-border control at much larger scale. Competition, capital expenditure, debt, merger review, customer neutrality, currency, and public-shareholder premiums became first-order. [researcher inference · strong] Sprint materials Arm transaction announcement, sections 1–2, 4, 7, and 11

2017–2022: large late-stage private rounds, outside LP capital, cheap financing, and rising marks expanded the opportunity set while correlating valuation, listing, subsidy, and governance risks. The FY2025 disclosed public-company tail shows that category leadership was not sufficient. [researcher inference · strong] FY2025 results, p. 78

2023–2026: the program shifts from defense to concentrated AI equity, semiconductor control, and infrastructure. Base rates now span late-stage AI securities, chip acquisitions, data-center construction, energy/power procurement, model governance, and consortium execution; Alibaba is not a sufficient analogy. [researcher inference · strong] OpenAI/Stargate lab FY2026 stack and investment states, printed pp. 23–25 and 80–83

Across eras, public company counts are non-additive and outcome-selected. FY1998 “more than 60,” FY2000 “more than 400,” the 23-position FY2015 table, a 2021 count including unclosed projects, a later “nearly 500,” 337 live SVF cards, and FY2022 fund counts use different entities and rules. Adding them would manufacture a base rate. [contemporaneous record · strong] public-count boundary FY1998 report, printed pp. 4–5 FY2000 report, printed pp. 4–5 FY2015 selected table, printed p. 26 / PDF p. 28 FY2022 position lists, printed pp. 98, 101–103 and 109–110

14. Modern VC translation (as of 2026-08-22)

Every row is an analyst translation, not a claim that Son stated the rule in this wording. “Current” means grounded in the frozen record through 2026-08-22; announced, conditional, funded, operating, marked, and realized states remain distinct.

15. Comparison with completed peers

Gary Rieschel, Jim Breyer, and Scott Sandell are completed canon peers with underlying evidence and useful matched dimensions. Rieschel is the direct former-SoftBank/local-authority comparison; Breyer tests prepared themes, cross-border networks, minority governance, and outlier dependence; Sandell tests technical/customer proof, distributed team credit, and an LP-backed multi-stage institution. This is a comparison of disclosed doctrine and boundaries, not a return ranking. [researcher inference · moderate] Rieschel Breyer Sandell Rieschel, 2005 USCC statement Breyer, Facebook S-1/A Sandell, CHM oral history, PDF pp. 1–75

Underlying controls anchor the comparison: Rieschel's 2005 USCC statement and Qiming formation/process account; Breyer's Facebook S-1/A; and Sandell's Computer History Museum oral history, PDF pp. 1–75, official NEA biography, and Cloudflare founder account. The shared result is that access, institution building, and company success do not by themselves prove personal alpha. [researcher inference · strong]

16. The strongest case against greatness

The skeptical case is that Son is an exceptional storyteller and corporate financier whose reputation is dominated by one extraordinary Alibaba outlier, selected survivor histories, and company-reported marks. The reconstructed universe has 534 canonical company families but only 11 sufficiently reconstructed with chronology plus material ownership, structure, governance, financing or reported-economics evidence; the subset is not a full decision/economics reconstruction, and selection and career-performance coverage are not computable. In the FY2015 selected scorecard, Alibaba supplied about 75% of reported value/return; removing Alibaba and Yahoo Japan reduced the researcher-calculated selected-table multiple to roughly 3.5x, before fees, tax, leverage, and selection-bias correction. [researcher inference · strong] dominant-outlier tests FY2015 selected scorecard, printed p. 26 / PDF p. 28 FY2022 position disclosures, printed pp. 98, 101–103 and 109–110

Many achievements are institution- and team-produced. Founders and employees created Yahoo, Alibaba, Arm, Coupang, DoorDash, and other company value; operating executives ran telecom assets; fund professionals sourced and managed positions; LPs supplied most SVF1 commitments. Son's public role is often thesis sponsor and narrator, while private sourcing, dissent, security design, board work, and exit authority remain unavailable. [researcher inference · strong] people map FY2017 organization and fund roles, printed pp. 14–15 SVF1 first close, investors and commitments FY2026 Investment Committee, printed p. 54

The philosophy's weakest operational link is between strategic vision and security economics. The record is rich in regime, category, mission, scale, and synergy, but thin in entry expected return, competing valuation, ownership target, reserve schedule, probability, negative diligence, position cap, and stop rule. WeWork, DiDi, Sprint, OneWeb, and ordinary outcomes show these are not clerical omissions. [researcher inference · strong] failure packet WeWork S-1, risk factors, printed pp. 49–51 FY2022 Q1 transcript, A9 FY2025 public-company tables, pp. 78–79

The current AI stack renews the strongest and weakest parts simultaneously: a coherent bottleneck map and extraordinary access, paired with concentrated factor exposure, large conditional financing, incomplete project economics, and founder/key-person dependence. Until a full security-level cash ledger, decision/pass archive, outlier-removed cohort, team attribution, and post-Son authority test are available, the most defensible conclusion is exceptional regime recognition and strategic asset assembly with unproved career-wide personal investment alpha. [researcher inference · strong] OpenAI/Stargate lab uncertainty ledger FY2026 stack and financing states, printed pp. 23–25 and 101–103

17. Analyst study guide, glossary, and go deeper

Recommended order

1. Start with the profile for chronology, entity boundaries, and the falsifiable edge.
2. Read the writings synthesis and talks operating sequence before using retrospective anecdotes.
3. Read Yahoo U.S., Alibaba, Vodafone Japan, Sprint, Arm, and OpenAI/Stargate in decision-date order.
4. Read WeWork, DiDi, OneWeb, and the ordinary public tail before accepting a winner-led rule.
5. Reconcile vehicles, people, and the denominator.
6. Apply all 16 mental models to a live opportunity and score every forecast, follow-on, and exit without hindsight.

Decision labs

Lab A — regime without a company. Write the regime map, bottlenecks, substitutes, adoption falsifiers, and value-accrual paths before naming candidates. Add three ways the regime can be right while every available security is wrong. [researcher inference · strong] Model 3

Lab B — founder with a second key. Evaluate ambition, integrity, learning, and recruiting; then independently price the security, map votes/related parties, model dilution, and assign stop authority. Compare Alibaba and WeWork without importing either outcome into the decision-date memo. [researcher inference · strong] Model 6

Lab C — ownership mode. For the same opportunity, write small-option, influential-minority, and control cases. State the work each permits, capital each obligates, and evidence that makes the extra ownership worthwhile. [researcher inference · strong] Model 8

Lab D — follow-on under stress. Hide the eventual outcome. Re-IC a down round using only fresh probability, security, expected return, remaining reserve, governance leverage, and portfolio liquidity. Force a named stop. [researcher inference · strong] WeWork failure lab

Lab E — return reconstruction. Build separate entity/security cash ledgers for Alibaba, Arm, and NVIDIA; show realized cash, remaining marks, transfers, fees, tax, leverage, and attribution unknowns. Then remove the top one and two outcomes. [researcher inference · strong] dominant-outlier tests

Lab F — 2026 AI stack. Treat OpenAI, Ampere, Stargate, Arm, agents, and robotics as both linked projects and correlated risk factors. Require bilateral synergy numbers, financing/downside cases, operational milestones, and decision rights. [researcher inference · strong] OpenAI/Stargate lab

Reusable checklists

Pre-IC: regime evidence; company capture; founder references; price/security; ownership/rights; vehicle/beneficiary; team attribution; financing; regulation; synergy math; milestone/falsifier; reserve/stop; duration/exit; comparable alternatives. [researcher inference · strong] Sprint ex-ante transcript, lines 701–787 and 1,065–1,264 FY2026 approval process, printed p. 54

Follow-on: original prediction versus actual; new security and price; ownership/dilution; fresh expected return; remaining reserves; governance change; correlated portfolio exposure; cash/LTV stress; stop authority; no sunk-cost language. [researcher inference · strong] WeWork S-1, risk factors, printed pp. 49–51 FY2022 Q1 transcript, A9

Portfolio review: complete active/exited/written-off/pass denominator; vintage and stage; cost/proceeds/mark; DPI/TVPI/IRR definitions; fees/tax/leverage/currency; vehicle and third-party interests; factor concentration; outlier-one and outlier-two exclusions; ordinary-outcome controls. [researcher inference · strong] FY2015 selected scorecard, printed p. 26 / PDF p. 28 FY2025 public-company tables, pp. 78–79

Postmortem: decision-date context; alternatives; original thesis/risks; warning chronology; each follow-on; governance and intervention; exit; cash/mark outcome; founder/team/vehicle attribution; violated rule; documented process change; prospective test of that change. [researcher inference · strong] FY2019 Q2 WeWork response, slides 20–29 FY2022 Q1 transcript, A4–A5 and A9

Ranked primary sources

1. 2012 Sprint SEC-filed transaction transcript — strongest ex-ante major-deal rationale, with market, operating-transfer, financing, competition, regulation, and alternatives visible. [contemporaneous record · strong]
2. SoftBank Next 30-Year Vision, deck PDF — best primary statement of mission, backcasting, federation, values, and succession; it is a vision document, not a return method. [investor-stated · strong]
3. FY2022 Q1 earnings transcript and FY2022 Son message — strongest public failure/defense record for price regret, selection, cash, LTV, costs, and slower deployment. [investor-stated · strong]
4. FY2026 annual report, pp. 23–25, 54, 80–83, and 101–102 — best current control for the ASI stack, IC, OpenAI/Ampere states, financing, and risk; announced strategy is not realized outcome. [contemporaneous record · strong]
5. FY2015 annual report, printed pp. 26–29 — best bounded historic selected-position scorecard and outlier test; it is thresholded, company-selected, and not a complete portfolio. [contemporaneous record · strong]
6. Arm acquisition terms and rationale, sections 1–2, 4, 7, and 11 and Arm sale termination — decision and exit-route controls for the strategic-IP thesis; neither alone supplies complete return economics. [contemporaneous record · strong]
7. Tokyo Forum with Jack Ma, 03:14–17:11 — the fullest direct founder-selection recollection in the processed corpus; it remains retrospective and supplies no original term sheet or rejected-founder denominator. [retrospective witness · provisional]

Ranked secondary and edited sources

1. Forbes, Master of the Internet — richest near-contemporaneous early report on Yahoo, STV, alliances, team, follow-ons, and misses; it is edited reporting, not Son-authored or a full denominator. [contemporaneous record · moderate]
2. Bloomberg/Rubenstein interview — high-density biography, Alibaba, Vodafone Japan, leverage, and succession testimony; the route is edited and the deal memories are retrospective. [retrospective witness · provisional]
3. TIME, 2026 profile — strongest current external chronology check on NVIDIA regret and the AI allocation thesis; reporter framing and later memory cannot replace transaction records. [retrospective witness · moderate]
4. Forbes, 2020 WeWork interview — rare direct admission on valuation and founder trust under pressure; it does not disclose the rescue decision expected return or full loss ledger. [investor-stated · strong]

Glossary

• ASI / AGI: Son's forecast categories for intelligence beyond or near broad human capability; public benchmarks and failure dates are incomplete.
• Backcasting: define a future state, then work backward to present actions; useful for coordination, not a probability estimate.
• Bottleneck / control point: scarce layer that constrains a system; strategic importance does not guarantee value capture.
• Category leader: company judged first in a defined market/geography; not a valuation, governance, or return conclusion.
• Decision-date freeze: evidence, alternatives, terms, risks, and forecasts recorded before the outcome.
• DPI / TVPI / IRR / MOIC: distributions, total value, time-weighted return, and value multiple; each requires a named vehicle, numerator, denominator, dates, and cash/mark boundary.
• LTV: holding-company loan-to-value; a financing-survival constraint, not a company-selection metric.
• Mark: estimated or public value still held; not cash distribution.
• Operating transfer: specific people, processes, assets, or procurement moved into a controlled company against a baseline.
• Power law: a small number of outcomes drive a large share of value; it does not excuse price, concentration, or missing denominator data.
• Sufficient reconstruction: this project's threshold requires chronology plus material ownership, structure, governance, financing or reported-economics evidence; it is not a full decision/economics reconstruction. Eleven of 534 company-family keys meet it, not the whole career. FY2022 position disclosures, printed pp. 98, 101–103 and 109–110
• Vehicle: legal/economic entity that owns the security and receives or bears cash flows; names below must not be pooled. vehicle map FY2022 fund terms, printed pp. 91–93
• SBG: SoftBank Group Corp., the listed strategic-holding parent. Its board, corporate financing, and Investment Committee govern the parent perimeter; its economics accrue to SBG and its shareholders, not automatically to Son personally or to fund LPs. vehicle map FY2026 group/IC disclosures, printed pp. 54 and 98
• SBGC / direct holding subsidiaries / operating companies: SBG subsidiaries can acquire, hold, transfer, finance, or operate assets; SoftBank Corp. and other opcos have their own management, boards, operating capital, and results. They are neither fund managers nor Son-personal accounts; operating revenue is not investment proceeds. vehicle map FY2026 group structure, printed p. 98
• Yahoo Japan / Z Holdings / LY Corp.: the Japanese portal/platform operating-family and its successors, with local management, boards, and joint-ownership history. It is not Yahoo U.S., an SBG fund, or Son personally; current LY operating results cannot be assigned to the 1996 investment without a security-level ledger. vehicle map FY2026 ownership map, printed p. 98
• SVF1 / SBIA: Vision Fund 1 limited-partnership/AIV perimeter and its manager. FY2022 reports $98.6B commitments—$33.1B SBG and $65.5B third parties—so SBG LP value, outside-LP value, and SBIA fee/performance economics must remain separate. vehicle map FY2022 terms, printed pp. 91–92
• SVF2 / SBGA: Vision Fund 2 limited-partnership/AIV perimeter and its SBG-owned manager. SBG and MgmtCo used different preferred/program-equity forms and distribution priorities; fund assets, manager economics, and participant economics are not interchangeable. vehicle map FY2022 terms, printed pp. 92–93
• SoftBank Latin America funds: regional and parallel investment vehicles managed by SBGA from FY2022 onward, with a separate portfolio/team history and economics. The FY2022 101-position disclosure is outside the live 337-card Vision Fund denominator; its gross fair value is not cash performance. vehicle map FY2022 LatAm disclosures, printed pp. 109–110 and 124
• SB Northstar: the listed-stock, derivative, and credit partnership disclosed as 67% SBG / 33% Son at FY2021, including Son's pro-rata unfunded obligation. It is not a venture fund; its gains, losses, liabilities, and instruments must retain the disclosed split. vehicle map FY2021 Northstar disclosure, printed p. 82 / PDF p. 86
• MASA USA LLC / MgmtCo: Son-controlled co-investment entity for specified SVF2/LatAm program investments. Its 17.25% applies only to scoped program equity, with restricted distributions and SBG preferred priority—not to SVF2, LatAm, each company, or SBG as a whole; it is related-party management capital, not an external LP. vehicle map FY2022 co-investment terms, printed pp. 92–93 and 124

Go-deeper questions

• Which private record would most likely falsify the regime-and-control edge?
• What was known, forecast, and unavailable at each decision date?
• Which security and vehicle funded the decision, and who ultimately received the economics?
• Who sourced, diligenced, voted, operated, governed, financed, and exited?
• What would results look like without Alibaba and without the top two outcomes?
• Which ordinary companies applied the same thesis but produced mediocre results?
• What milestone earned each follow-on, and who had authority to stop?
• What is the operational evidence for each claimed ecosystem synergy?
• What capital/liquidity state makes the investor a forced seller?
• Can the method operate and retain confidence without Son?

Pareto 80/20 — what drove the record

The public record supports concentration, but not a literal claim that 20% of investments produced 80% of all returns. There is no complete all-history position denominator, the available scorecards mix cash and marks, and the legal/economic perimeter changes from SBG corporate capital to controlled operating companies, SVF1, SVF2, LatAm, co-investment programs, and Son's separately disclosed interests. Alibaba is nevertheless the best-supported dominant realized contributor: SBG reported ¥9.7 trillion of total return and a 57% IRR through 11 May 2023, largely monetized, pre-tax and net of monetization costs, while excluding the then-unmonetized stake. That is a company-reported SBG measure—not a Son-personal, fund-net, or independently reconstructed result. [contemporaneous record · strong] SoftBank Group Report 2023, CFO message major-case packet, Alibaba

The rankings below deliberately answer three different questions. “Economic contribution” ranks the best-supported directional contribution, not a currency-converted common return series. “Historical importance” ranks changes to SoftBank's capabilities and strategic path, not money. “Decision-learning value” ranks cases that most sharply test process. Every row states why the rank is bounded. [researcher inference · strong] major-case contribution audit

Ranked by economic contribution

Ranked by historical importance

Ranked by decision-learning value

What remains after Alibaba. Removing Alibaba eliminates the best-supported dominant largely realized contributor. The remaining record still includes a historically valuable Yahoo Japan position, operating-platform builds, the marked Arm stake, Coupang and DoorDash gains, and a profitable NVIDIA exit—but no source permits their aggregation into one clean corporate/fund cash return. In the disclosed SVF1 public-company table alone, removing Coupang leaves about 1.03x; removing Coupang and DoorDash leaves about 0.87x. This mechanical test is not an adjusted fund return, yet it demonstrates how quickly the public-book conclusion changes when two contributors are excluded. [researcher inference · strong] FY2025 results presentation, SVF1 table, p. 78 failure/ordinary packet, dominant-outlier test

Evidence boundary and denominator reconciliation

This synthesis is built from three frozen evidence packets: the denominator, vehicle, people, and full-ledger packet; the nine major-case packet; and the failures, ordinary deals, and anti-portfolio packet. Writings and talks are used to establish dated thesis evolution and admissions, not to replace transaction records. The cutoff is 2026-08-22. [researcher inference · strong]

Exact reconciliation: 830 observations become 534 keys, not a complete population

The 24 unique deep/partial case labs are a different analytical set from the eleven strict sufficient ledger families. They comprise nine major labs plus fifteen adverse/ordinary labs after deduplicating NVIDIA across packets, treating Greensill/Katerra and Ziff-Davis/COMDEX as paired transaction systems, and treating Wag and Zume as distinct underlying adverse companies. Yahoo! BB is an organic operating lab and Amazon is a missed-investment lab, so neither is an external-position denominator key. Several labs remain partial because terms, team attribution or complete cash flows are missing. [researcher inference · strong] major packet, scope failure packet, scope

Eight public portfolio counts are visible but non-additive: FY1998 “more than 60” indirect fund companies; FY2000 “more than 400” group companies and fund investees; 23 thresholded FY2015 scorecard rows; 264 Vision Fund “investments” in 2021 including 44 IC-approved but unclosed projects; “nearly 500” invoked in 2023; 337 live SVF cards on 2026-08-22; 101 FY2022 LatAm held positions; and 332 FY2022 held SVF positions. Their dates, entities, pipeline states and inclusion rules differ. Adding them would manufacture a denominator. [contemporaneous record · strong] FY2022 report, printed pp. 98, 101–103 and 109–110 denominator packet, public claims

Selection bias runs in both directions. Corporate histories emphasize defining wins and major crises; the FY2015 scorecard admits only large positions; the live SVF page omits economics and may omit historical failures; adverse press overweights spectacular losses. This report therefore uses ordinary public rows—Uber, Grab, Paytm, low-value SVF1/SVF2 positions—as controls, while refusing to invent decision narratives for thin rows. [researcher inference · strong] FY2015 report, printed p. 26 / PDF p. 28 failure packet, public tail

Historical context, contemporaneous alternatives, and base rates

1995–2000: scarce Internet distribution and cheap optionality—followed by bubble selection

SoftBank entered the public Internet from a technology-media and trade-show base. Yahoo's 800,000 daily users in November 1995 and roughly 15 million daily users by April 1996 were contemporaneous traction, not a story retrofitted after the IPO. The decision set included a small syndicate check, a larger follow-on, internally built portal/distribution, regional licensing, or continued ownership of publishing and events. SoftBank used several at once. The defensible skill claim is staged recognition of an attention gateway and localization optionality; it is not that Internet affiliation reliably produced returns. The FY2015 threshold ledger itself includes Betfair at 0.8x, Renren at 0.9x, Key3Media at 0.6x, Zynga at 0.7x, CNET at 0.1x and Asia Global Crossing at 0x. [researcher inference · strong] 1995 Yahoo announcement FY2015 report, p. 26

Alibaba's 2000 decision belongs to the same post-portal search for control points but in a much thinner China Internet market. Alternatives included smaller exposure, regional portals, incumbent commerce companies, a fund basket, or passing until monetization was established. Public records do not supply a 2000 valuation memo or base-rate analysis. That missing counterfactual matters: a 761x selected-scorecard outcome can make sparse ex-ante evidence appear obvious. [researcher inference · strong] Alibaba 2014 prospectus, history denominator packet, FY2015 ledger

2001–2006: capital-intensive distribution and control

After the technology crash, Yahoo! BB and the Vodafone Japan acquisition changed the problem from minority optionality to operating execution. The alternatives were to remain an investor/content distributor, lease networks, partner with incumbents, buy a smaller carrier, or acquire national control with leverage. Yahoo! BB required subsidized customer acquisition and ongoing operating losses before scale; Vodafone combined roughly ¥1.75T of equity value with a layered financing structure that included SBG cash, Yahoo Japan preferred securities, nonrecourse acquisition debt and Vodafone instruments. These cases should be evaluated against telecom turnarounds and leveraged-control base rates, not early-stage venture checks. [contemporaneous record · strong] FY2006 report, pp. 6–10 major packet, Vodafone

2012–2016: cross-border control at much larger scale

Sprint and Arm applied the same “control point plus operating playbook” method in different industries. At Sprint, the decision alternatives included a minority partnership, roaming/device cooperation, waiting for consolidation, or declining a leveraged U.S. turnaround. The public case asserted purchasing, devices, network and operational synergies, while contemporaneous risks included competitive intensity, capital expenditure, debt, regulation and uncertain consolidation. Arm offered a cleaner control asset with foundational IP and no comparable turnaround, but exposed SBG to sterling acquisition price, public-shareholder premium, semiconductor cyclicality, customer neutrality and the challenge of translating technology importance into cash return. [researcher inference · strong] Sprint SEC-filed transaction materials Arm acquisition announcement, sections 2–7 and 11

2017–2022: scale fund investing, cheap capital, marks, and correlated downside

SVF1 began with outside LP capital and a $98.6B commitment structure; the scale changed both opportunity set and error cost. Category leaders could accept larger primary/secondary checks and extend runway, but late-stage private valuations, founder votes, shared listing windows, leverage and subsidy-dependent unit economics correlated supposedly diverse positions. Mobility exposure across Uber, DiDi, Grab and Ola diversified geography while concentrating the same category; WeWork/OYO/Compass diversified names while retaining physical-network and capital-intensity risk. Against the public-company base rates now visible, “leader” was not sufficient: 2026 rows showed Uber 1.2x, Grab 0.8x, DiDi 0.3x and Paytm 0.7x. [researcher inference · strong] FY2025 results, SVF1 public table, p. 78 failure packet, portfolio patterns

2023–2026: defense gives way to concentrated AI infrastructure

After defense-mode rhetoric, overseas-team consolidation and tighter liquidity/runway focus, SBG returned to very large concentration through Arm, OpenAI, Ampere and Stargate. The historical analogy is not simply Alibaba. OpenAI is a minority preferred-equity position at a mature private valuation, funded through SVF2 and corporate bridge capacity; Stargate is an infrastructure program involving other counterparties; Ampere is a control acquisition. Alternatives include smaller staged exposure, syndicated equity, customer/offtake partnerships, project-level nonrecourse financing, and a diversified model/compute basket. The proper base rate spans late-stage AI equity, semiconductor acquisitions, hyperscale infrastructure construction and technology-platform regulation—not venture-company selection alone. [researcher inference · strong] FY2026 report, pp. 23–25 and 80–83 major packet, OpenAI

Vehicles, economics, and incentives

Instrument type repeatedly changes apparent exposure. WeWork used convertibles, warrants, preferreds and support finance; OneWeb used common/preferred equity, loans, warrants and shareholder loans; Sprint ended in T-Mobile shares, conditional shares, options and sales; Arm involved corporate acquisition, in-kind transfer, failed-sale deposit, IPO and retained equity; OpenAI involves funded preferred equity, syndication, decided follow-ons and bridge finance. A valid result therefore needs a security-by-security, date-stamped cash-flow ledger before any IRR or multiple. [researcher inference · strong] denominator packet, vehicle map WeWork S-1/A, pp. 118–119 OneWeb instruments, printed p. 14 / PDF p. 18 OpenAI 2026 tranches, sections 1–5

Structural comparison with an LP-backed venture fund. SBG corporate capital is evergreen in legal duration but constrained by listed-company leverage, liquidity, ratings, governance and NAV discount. SVF1 is itself LP-backed: it had a defined investment period, a fund life in principle to 2029 with extension options, outside LPs, adviser fees, realization-linked performance fees and clawback. Thus “patient SoftBank capital” applies most cleanly to retained corporate holdings, not automatically to every fund asset. An LP-backed vehicle can diversify capital sources and formalize manager/advisory rights, while also introducing fund-life, distribution, valuation and manager/LP incentive constraints. This is a structural comparison on disclosed terms, not a claim about a generic venture-fund return benchmark. [researcher inference · strong] FY2022 report, pp. 68–69 and 91–93

Fundraising, geographic authority, reserves, succession, and key-person continuity. SBG's listed-company capital has no closed-end fundraising cycle, but every large commitment competes with capital-market access, liquidity, leverage and asset-monetization capacity; SVF1 instead combined outside-LP fundraising with a defined investment period and fund life. The dated authority map puts SVF1 under SBIA, SVF2 under SBGA, and the LatAm funds under a historically distinct team later managed by SBGA; it does not disclose a complete geography-by-geography approval map, company-level reserve policy, follow-on reserve schedule, or exception log, so those remain unknown. Nikesh Arora's announced 2015 management transition and later reversal, Rajeev Misra's dated fund roles, and the current five-member SBG Investment Committee show changing institutional seats, not a disclosed post-Son succession plan or a tested key-person continuity mechanism. [researcher inference · strong] FY2015 report, management transition, printed p. 7 / PDF p. 9 FY2016 report, reversal, printed pp. 10–11 / PDF pp. 12–13 FY2017 report, fund leadership, printed pp. 14–15 / PDF pp. 16–17 FY2022 report, advisers and LatAm management, printed pp. 91–93 and 124 FY2026 report, Investment Committee, printed p. 54

People, capital, and influence map

Influence and economics do not travel together. Coupang illustrates large economic ownership with founder supervotes; WeWork could reach approximately 80% economic ownership without majority shareholder/board votes; DiDi disclosed 20.0% economics but 10.7% voting power for the SVF entity; PPBL's regulatory obligations remained distinct from listed One97; Stargate partners do not become SBG investees merely through project participation. Every case below therefore maps capital provider, legal holder, manager, board/vote rights, operating team and regulator separately. [researcher inference · strong] Coupang 424B4, principal shareholders FY2019 Q2 WeWork deck, accounting slides 4–5 DiDi 20-F

Case-specific decision-role map

“Unknown” means the scoped public record does not assign that function; it does not mean no one performed it. [researcher inference · strong] denominator packet, ledger controls

Comprehensive case ledger

This ledger indexes the 24 unique deep/partial labs. It does not replace the full 534-key union: every FY2015, FY2014, annual-report, live-SVF, historical-SVF and LatAm row remains in the full denominator packet. “Outcome” below always retains the cited source's date and perimeter. [researcher inference · strong]

Nine major decision labs

For each lab, the order is fixed: contemporaneous information and alternatives first; an analyst-created four-prompt investment-committee worksheet second; actual outcome only third. The worksheet is explicitly a research inference—not an allegation about SoftBank's unreleased internal memo. [researcher inference · strong] major-case method

1. Yahoo U.S. — attention gateway and staged conviction

Ex ante (November 1995–July 1998). Yahoo was a young portal with about 800,000 daily users when SoftBank Holdings and Ziff-Davis together invested $2M for about 5%; SBH's exact portion is not disclosed. The thesis joined traffic and navigation technology to SoftBank's publishing/event distribution and already contemplated Japan localization. By April 1996 Yahoo reported about 15M daily users in more than 90 countries, and SoftBank disclosed $108.25M cumulative investment for 37.02% of expected post-offering shares, subject to option/underwriter dilution. July 1998 added a $250M private placement under a contractual 35% adjusted ownership ceiling. Alternatives were to keep the original small syndicate position, invest through diversified funds, rely on distribution partnerships, build a portal, or stage follow-ons. Missing were a public revenue model, preference schedule, board-rights map, downside case, reserve policy and named dissent. [contemporaneous record · strong] 1995 announcement, company/purpose/investment 12 April 1996 announcement Yahoo 1999 10-K, Note 7

No-hindsight IC worksheet. (1) Invest now? Make a staged minority investment because traffic leadership and regional rights create asymmetric optionality, but do not size from a future IPO. (2) Kill the deal? Persistent traffic-share loss, failure to monetize repeat use, dilution below strategic relevance, or inability to obtain Japan/localization rights. (3) Ownership/terms? Information and pro-rata rights, a board/observer path, explicit territorial/JV rights, and a hard ownership/valuation ceiling; only the stake and ceiling are publicly proven. (4) Next check? Retention, advertiser conversion, public-market financing, and executed Yahoo Japan rights. [researcher inference · moderate] 1996 announcement, traffic and dilution

Actual decision and outcome. SoftBank escalated after observable adoption. Its July 1998 release reported about ¥332.5B of unrealized gain—a mark, not cash. SoftBank America then agreed to sell 17,290,866 shares for about $300M/¥39.4B in December 2001, expecting about ¥27B gain, and sold another 11.5M shares for about $171M/¥22.4B in April 2002, expecting about ¥15.7B gain. Neither transaction closes the full ledger. Son's 2017 recollection—$100M for 35%—is directionally close but conflicts with the exact 1996 $108.25M/37.02% record, which controls. The case supports gateway recognition, staged follow-on and localization; it does not prove that traffic is a sufficient screen or yield a complete cash multiple. [contemporaneous record · strong] 9 July 1998 announcement 28 December 2001 release major packet

2. Yahoo Japan — localization, joint control, and operating autonomy

Ex ante (January–April 1996). SoftBank and Yahoo created a Japanese joint venture in January 1996 with a 60/40 ownership split, combining a proven portal/brand and SoftBank's local distribution and operating capacity. The decision was not merely to buy Yahoo U.S. shares: it was to secure territory, localize the service and create a separate issuer/operator. Alternatives included licensing, a distribution agreement, internal build, a smaller minority, or postponing until Japan's Internet adoption matured. Public records support the JV and partner logic but do not expose the original cash/security schedule, board reserved matters, budget, unit economics, management selection, downside case or dissent. [contemporaneous record · strong] Yahoo 1999 10-K, Yahoo Japan discussion major packet, ex-ante freeze

No-hindsight IC worksheet. (1) Invest now? Form the JV if exclusive/local rights, management accountability and staged funding can be secured. (2) Kill the deal? Nonexclusive brand/technology, partner veto without operating responsibility, weak Japanese adoption, or no path to local monetization. (3) Ownership/terms? Joint board, reserved matters, territorial rights, IP/service commitments, pro-rata funding and explicit dilution/exit rules. (4) Next check? Japanese traffic retention, advertiser demand, local content depth, reliable service and a stand-alone financing plan. [researcher inference · moderate] Yahoo 10-K, JV discussion

Actual decision and outcome. Yahoo Japan listed and diluted over time while remaining a separately managed platform. FY2015 reported SoftBank's voting rights at 43.0% and placed the position in a selected threshold scorecard: ¥7.7B accumulated investment, ¥1,358.9B estimated value/accumulated return, 78% IRR and 176x. Those are management estimates within a size-selected table and include mark/recovered-value concepts, not a current cash-only return. Later Z Holdings/A Holdings/LY structures further change the perimeter. The evidence supports local platform construction and long-duration partnership, but not sole Son execution or a timeless ownership figure. [contemporaneous record · strong] FY2015 report, pp. 26–29 denominator packet, Yahoo Japan

3. Alibaba — founder selection, follow-ons, governance, and monetization

Ex ante (2000–2005). The public chronology says a SoftBank-led group invested $20M in Alibaba in 2000, but does not publish the original security, SoftBank-only allocation, valuation, cap table or IC alternatives. The discernible thesis was a founder-led Chinese commerce platform with network effects and a long market runway. The 2003 structure added joint-venture subscriptions and convertibles: public records distinguish a $50M subscription, $30M convertibles and later exchanges rather than one simple check. In 2005 Yahoo invested $1B and contributed Yahoo China; Yahoo paid SoftBank $360M for shares, while SoftBank exchanged remaining Taobao-related interests and reinvested $180M through convertibles. These separate primary, secondary, contributed-asset and rollover flows. Alternatives included smaller staged exposure, a China Internet basket, regional portals, waiting for monetization, or partnering without ownership. Key missing decision-time evidence is a valuation memo, customer/merchant cohorts, regulatory/currency downside, founder-governance protections, reserve limits and dissent. [contemporaneous record · strong] Alibaba 2014 prospectus, history and principal shareholders major packet, ex ante

No-hindsight IC worksheet. (1) Invest now? Take a staged minority only if merchant liquidity, founder integrity and enforceable information/governance rights compensate for a nascent model and China risk. (2) Kill the deal? Weak repeat merchant activity, unreliable payment/trust infrastructure, opaque related parties, regulatory inability to own/enforce rights, or follow-on price untethered to operating evidence. (3) Ownership/terms? Pro-rata and information rights, board/observer access, founder-transfer limits, related-party approval and explicit protections through reorganizations; do not equate percentage ownership with voting power. (4) Next check? Marketplace liquidity, repeat merchants/buyers, monetization without destroying growth, regulatory structure, qualified finance controls and outside financing. [researcher inference · moderate] Alibaba prospectus

Actual decision and outcome. SoftBank followed on, accepted changing governance and capitalization, held through the 2014 IPO and monetized in stages. The prospectus recorded SoftBank at 32.4% before the offering and described voting/proxy/partnership arrangements; that ownership is not timeless. In 2023 SBG reported ¥7.4B invested and ¥9.7T total return, 57% IRR through 11 May 2023, net of monetization costs, pre-tax and excluding the unmonetized stake. The apparent ¥7.4B cost does not reconcile every $20M/2003/2005 security flow in the public chronology, so the company-reported return may be quoted only with its perimeter. Alibaba's team created the business; Son's supported contribution is early sponsorship, continued backing and later strategic governance. [contemporaneous record · strong] 2023 CFO message major packet, actual

4. Yahoo! BB — organic access layer built through losses

Ex ante (June 2001–FY2003). Yahoo! BB was an SBG operating build, not an external investee. The thesis was that low-price high-speed broadband could break an incumbent bottleneck and give SoftBank a direct customer/distribution layer for Internet services. Decision-time alternatives included remaining a content/portal investor, reselling incumbent access, a network JV, slower regional rollout or abandoning the subsidy. Risks were capital burn, customer support, network quality, incumbent response, financing during the technology crash and uncertain attachment economics. The reports document operating ambition and losses but not a standalone IC memo, contribution-margin cohorts, hard customer-acquisition ceiling, named dissent or shutdown trigger. [investor-stated · strong] FY2003 report, broadband sections major packet

No-hindsight IC worksheet. (1) Invest/build now? Proceed only as milestone-funded infrastructure with explicit group liquidity protection. (2) Kill the build? Churn or service failures above plan, contribution margin that does not improve with scale, financing runway below the network break-even interval, or incumbent pricing that destroys the cost advantage. (3) Ownership/terms? Because this is an opco build, focus on project budget, procurement, financing covenants, accountability and customer-quality metrics rather than venture preference terms. (4) Next check? Net adds, churn, activation cost, support cost, network uptime, ARPU, contribution margin and months to cash break-even. [researcher inference · strong] FY2003 report

Actual decision and outcome. SBG continued through large broadband losses, built scale and used the access relationship as a strategic foundation for later services and mobile entry. The evidence supports concentration, persistence and operating execution by broad teams. It does not support translating subscriber or operating improvement into a venture multiple, and the case is excluded from the 534-key external-position union. Persistence was valuable here, but that fact cannot justify rescuing any capital-intensive portfolio company without comparable milestone evidence. [researcher inference · strong] major packet, outcome FY2003 report, broadband results and strategy

5. Vodafone Japan — leveraged control of mobile distribution

Ex ante (17 March 2006). Vodafone K.K. was Japan's third-largest mobile operator, with more than 15M subscribers and over 99% population coverage; mobile-number portability was expected that November. A wholly owned SoftBank subsidiary agreed to acquire 97.7% at ¥1.75T equity value. The announced financing was ¥200B SoftBank common equity, ¥120B Yahoo Japan preferred shares, ¥1.1–1.2T nonrecourse LBO debt, and ¥400B from Vodafone International through preferred shares/warrants and subordinated loans; Vodafone retained a director appointment while its preferreds remained. The thesis was buy-versus-build: combine inherited coverage with broadband/portal assets, procurement, handsets, sales channels and Yahoo content. Alternatives were organic spectrum pursuit, MVNO entry, a smaller partnership or waiting. The chosen leverage amplified churn, network-quality, integration, price-war and refinancing risk. No public downside valuation, covenant schedule, integration budget, fairness opinion, dissent or leverage stop rule is available. [contemporaneous record · strong] 17 March 2006 release, sections 1–5 FY2006 report, pp. 6–10

No-hindsight IC worksheet. (1) Acquire? Only with control and a ring-fenced debt-service plan because the national network is scarce but the equity cushion is thin. (2) Kill the deal? Portability-driven subscriber loss, persistent network gaps, EBITDA inadequate for nonrecourse debt, poor handset economics, partner conflict or refinancing dependence before repair. (3) Ownership/terms? Operating control, covenants aligned to carrier cash flow, capex authority, preferred-capital clarity and a board-approved integration plan; the release proves the main layers, not detailed covenants. (4) Next check? Coverage/speed, handset depth, churn, net additions, ARPU, EBITDA, free cash flow and debt repayment—not subscriber count alone. [researcher inference · strong] 2006 release, transaction overview

Actual decision and outcome. SoftBank doubled targeted 3G base stations, expanded handsets, introduced Yahoo! Keitai, installments and White Plan pricing. One year later it reported subscriber-addition share up from 3.5% to 14.2% and Mobile Communications operating income of ¥155.7B; it described about ¥1.4T acquisition debt as nonrecourse to other operations and serviced by mobile EBITDA. These are operating/financing milestones, not final return. Son's 2017 recollection of roughly $2B equity/$18B debt is less precise than the announcement structure and cannot replace it. The case supports control plus coordinated operating intervention by network, product, finance and distribution teams. A complete return still requires acquisition funding, capex, debt service, dividends, taxes, listing/stake-sale proceeds and retained SoftBank Corp. value. [contemporaneous record · strong] FY2007 report, pp. 13–20 major packet, outcome

6. Sprint to T-Mobile — control, turnaround, and regulatory exit

Ex ante (15 October 2012). The initial agreement contemplated $20.1B for about 70% fully diluted: $12.1B secondary cash and $8.0B new Sprint capital, beginning with a $3.1B seven-year 1% convertible at $5.25, followed by $4.9B common at $5.25 and cash consideration; SoftBank also received a five-year warrant for about 55M shares at $5.25. Son said funding would combine about ¥700B cash and a ¥1.5T bridge loan, without new SBG equity, and retained Sprint's listing partly to preserve financing access. A ten-member proposed board gave SoftBank majority appointment rights, Son chairmanship and Dan Hesse continued CEO authority. The thesis was U.S. mobile-data scale, new network capital and transfer of procurement/device/network experience. Alternatives included tender offer variants, minority partnerships, other strategic partners or no transaction. Missing are internal spectrum/capex cases, leverage stress, churn downside, integration budget, required-return bridge and dissent. [contemporaneous record · strong] SEC-filed transcript/slides, lines 701–814 and 1794–1928

No-hindsight IC worksheet. (1) Invest? Conditionally, with majority rights and milestone-gated capital; do not treat Vodafone as a sufficient base rate. (2) Kill? Network metrics fail despite capex, churn/ARPU deteriorate, refinancing fails, regulation blocks the strategic path, or required capital exceeds a predefined SBG balance-sheet ceiling. (3) Ownership/terms? Majority board rights, primary/secondary pricing clarity, management-replacement authority, funding limits and a liquid minority float. (4) Next check? Coverage, speed, churn, net adds, post-capex free cash flow, leverage and management execution. [researcher inference · strong] SEC-filed materials, financing and governance

Actual decision and outcome. Final July 2013 terms superseded the proposal: SBG invested about $21.6B and held about 78% fully diluted of New Sprint; about 72% of old shares received $7.65 cash, Sprint had acquired Clearwire, and SBG borrowed ¥1,284.9B under the bridge facility. By 2016 Son admitted the rebuild was harder than expected and credited CEO Marcelo Claure with execution. DOJ/FCC conditions then shaped the 2020 T-Mobile combination. At close, Sprint shares converted at 0.10256 T-Mobile shares; after 48,751,557 contingent shares were surrendered, Deutsche Telekom and SoftBank held about 43.6% and 24.7%. T-Mobile reported $40.8B acquisition-date consideration, including stock, Sprint debt repayment and contingent value. SBG's AGM-reported 25% Sprint equity IRR mixed shares, sales, options and a contingent right while excluding a potential indemnity. It is company-calculated, not a final cash IRR. [contemporaneous record · strong] 2013 completion release T-Mobile 2020 10-K, Note 2 2020 AGM deck, pp. 34–35

7. Arm — foundational IP, control, failed sale, and relisting

Ex ante (18 July 2016). SBG agreed to buy 100% of Arm for 1,700 pence per share—about £24.3B equity value and $31B/¥3.3T total investment—at a 43% prior-close and 69.3% three-month-VWAP premium. Up to ¥1T came from a Mizuho facility and the balance from cash. Son's thesis was a horizontal, energy-efficient processor architecture for IoT and later AI, with Arm remaining independent in Cambridge, retaining talent and customer neutrality while accelerating R&D. Alternatives were minority ownership, strategic partnership or no deal. Risks included the control premium, licensee neutrality, customer design-arounds, RISC-V/x86 competition, cyclicality, royalty economics and financing. The public announcement names advisers and approvals but not a customer-diligence file, royalty downside, R&D budget, independent SBG fairness opinion, dissent or exit rule. [contemporaneous record · strong] 18 July 2016 announcement, sections 2–7 and 11 6 September 2016 statement

No-hindsight IC worksheet. (1) Acquire? Conditionally, after a valuation and ecosystem-neutrality stress test; secular importance does not itself justify the premium. (2) Kill? Licensees design around Arm, ownership impairs neutrality, R&D fails in servers/AI/automotive, open architectures compress royalties, or financing crowds out investment. (3) Ownership/terms? Full control paired with an independent Arm board, neutrality commitments, engineering retention and leverage ceiling; only control and public promises are established. (4) Next check? License renewal, royalty growth, non-mobile design wins, talent retention, R&D milestones and customer evidence that ownership has not damaged the ecosystem. [researcher inference · strong] Arm 2023 prospectus, business model and risks

Actual decision and outcome. SBG completed a £24B cash acquisition in September 2016 and later used about 24.99% of Arm, valued around $8.2B for commitment purposes, as part of its SVF1 in-kind commitment. In 2020 SBGC/SVF1 agreed to sell Arm to NVIDIA for up to $40B across a deposit, cash, stock, earn-out and employee awards. Competition authorities challenged customer-access risks; the parties terminated in 2022, leaving a nonrefundable $1.25B deposit, 24.99% attributable to SVF1. The Arm China governance dispute and Acetone structure show that owning Arm did not mean simple direct control of the China operation. Before the 2023 IPO, an SBG subsidiary repurchased substantially all of SVF1's 24.99% for $16.1B through installments—an intra-group transaction. The IPO sold 102.5M ADSs at $51; Arm received no proceeds and SBG retained control. SBG's later ¥24.6T/10x claim is company-reported and largely mark/retained-value dependent, not a cash IRR. [contemporaneous record · strong] 2022 termination Arm prospectus, cover major packet, actual

8. NVIDIA — profitable exit, duration failure, and recollection conflict

Ex ante (December 2016–September 2017). SBG bought public NVIDIA shares for $2.834B and transferred the same shares to SVF1 at cost in September 2017. The timing beside Arm suggests semiconductor/accelerated-computing exposure, but the public note provides no original thesis, price target, target ownership, board right, alternatives, duration, kill criteria or dissent. This was liquid minority equity, not Arm control. SVF1 used collars, monetized shares into borrowings and later delivered shares against those borrowings, so downside protection, liquidity and surrendered upside were part of the position. [contemporaneous record · strong] FY2019 report, “Investment in NVIDIA,” printed p. 66

No-hindsight IC worksheet. (1) Invest? Yes only as a sized liquid position with an explicit accelerated-computing thesis and horizon; the source is too thin to score the actual memo. (2) Kill? Competitive displacement, thesis failure in data center/AI, valuation far above underwritten cash flows or portfolio liquidity breaching policy. (3) Ownership/terms? Position-size, collar, borrowing/collateral and sell-rule limits matter more than board control; any hedge should price surrendered upside. (4) Next check? Data-center/AI revenue materiality, product leadership and a portfolio liquidity budget that does not force sale simply because the asset is liquid. [researcher inference · provisional] FY2019 report, NVIDIA note

Actual decision and outcome. SVF1 unwound collars and sold/delivered all NVIDIA shares in January 2019, reporting $2.915B total gain before third-party interests: $1.450B share gain and $1.465B derivative gain. The later standardized table gives $2.878B cost, $5.813B return and 2.0x. It was profitable, but missed far larger later upside. Son's 2026 recollection that COVID-era financial constraints forced the sale conflicts with the January 2019 filing chronology and is not used as fact. The useful lesson is to make thesis duration, liquidity, collateral and re-underwriting explicit; it is not “never sell.” The unearned later value is a counterfactual, not a loss to net against realized gain. [contemporaneous record · strong] FY2019 report, NVIDIA table TIME, sale discussion failure packet, NVIDIA

9. OpenAI and Stargate — current concentration with no realized return

Ex ante (September 2024–January 2025). SBG first invested $2.2B through SVF2 from September 2024. A March 2025 agreement then authorized up to $40B at $260B pre-money, with up to $10B third-party syndication and effective SBG exposure up to $30B. The convertible interest was expected to become preferred stock with 1x liquidation preference and automatic IPO conversion; a $10B first close preceded a conditional second close. The thesis linked a leading model platform to Arm and Stargate. Stargate separately announced an intention to invest $500B over four years, with $100B “immediately”; SoftBank, OpenAI, Oracle and MGX were initial equity funders, while SoftBank held financial responsibility, OpenAI operational responsibility and Son chairmanship. These were roles and plans, not a completed project-capital ledger. Alternatives included a smaller syndicate position, a model basket, waiting, customer-backed projects or staged nonrecourse infrastructure. Risks included valuation, minority governance/recapitalization, model competition, safety/regulation, power, construction, demand and correlated Arm/OpenAI/compute exposure. [contemporaneous record · strong] 1 April 2025 transaction release 22 January 2025 Stargate announcement

No-hindsight IC worksheet. (1) Invest? Only in staged tranches, with model equity and infrastructure project finance underwritten separately. (2) Kill? Recapitalization weakens rights, model leadership/retention erodes, regulation blocks deployment, sites lack contracted demand/power, construction fails, or SBG breaches LTV/liquidity policy. (3) Ownership/terms? Preferred liquidation/information rights, tranche conditions, syndication and group exposure cap for OpenAI; nonrecourse vehicles, offtake, completion guarantees and equity-loss caps for sites. (4) Next check? Closed recapitalization, funded syndication, stable rights, revenue/retention, powered sites, contracted customers and genuinely nonrecourse financing. [researcher inference · strong] 2025 transaction release

Actual funding through 2026-08-22. SBG says it completed $7.5B first-close and $22.5B second-close investments through SVF2 in 2025; third parties supplied $11B, producing $41B aggregate commitment and about 11% SBG ownership. In February 2026 SBG agreed to another $30B through SVF2 at $730B pre-money in three $10B preferred tranches. The first two were funded by July; the third remained scheduled for 1 October, so $20B was funded and $10B was still committed at the cutoff. If all closed, guidance was $64.6B cumulative investment and about 13% ownership. The July tranche used $10B bridge borrowing. Planned Stargate capacity, expected jobs and future project investment are not deployed SBG capital or return; current nonrecourse project-finance language is policy, not proof every site closed. OpenAI is FVTPL, so marks can affect profit/NAV without cash. Son owns the thesis and Stargate chair role; boards, Goto, SVF2, OpenAI, co-investors, lenders, Oracle and site teams own distinct approvals and execution. No realized return is claimed. [contemporaneous record · strong] 31 December 2025 completion 27 February 2026 follow-on 1 July 2026 funding FY2026 CFO message

Failure, miss, and ordinary-outcome decision labs

These fifteen analytical units add the adverse and ordinary comparison set; the Greensill unit includes Katerra only as an outcome-linked partial control, not a complete standalone decision lab. NVIDIA's exit/miss appears once, in the major section, to avoid double counting. A “no documented process change” conclusion means no case-specific change was found in the scoped public sources—not that no private lesson or control existed. [researcher inference · strong] failure-packet scope

1. WeWork — conviction, escalation, governance, and stop discipline

Ex ante (2017–14 August 2019). Son described WeWork's workspace network and $890M 2017 sales as technology/community-driven category creation. Before the planned IPO, however, the issuer's S-1 disclosed dependence on continuing capital, large lease liabilities, immature controls, related parties, broad founder discretion and founder financing. The amended filing also documented a $1.0B 2018 SBG convertible, $2.5B paid under a warrant, a further $1.5B commitment and $1.6B Asia-JV preferred liquidation preferences. These were decision-near warnings, not hindsight. Alternatives were a smaller milestone round, governance restructuring, senior rescue terms, an external syndicate or refusal. The public record does not reveal the full IC comparison, reserve cap or dissent. [contemporaneous record · strong] We Company S-1, pp. 49–51 S-1/A, pp. 118–119

No-hindsight IC worksheet. (1) Invest then? Only as a staged restructuring, not a growth top-up. (2) Kill? Unresolved founder related parties, no lease-adjusted path to self-financing, weak controls or votes/board rights inadequate for total exposure. (3) Ownership/terms? Senior preference, independent board, related-party veto, founder-liquidity limits, hard reserve cap and votes matching downside. (4) Next check? Location cohorts, cash burn, lease-adjusted contribution, controls and an IPO feasible without optimistic marks. [researcher inference · strong] S-1 risk/control sections

Actual outcome and process evidence. After the IPO failed, a November 2019 package put total equity exposure at $7.8B and could raise economic ownership to about 80%, yet SBG lacked majority shareholder/board votes and therefore accounting control. New financing, management replacement, halted openings and cost cuts were intervention rather than a clean stop. The 2024 SVF1 public row later showed $3.468B cost and zero return; the broader FY2025 reclassification showed $6.71B losses—$3.18B SVF1 and $3.53B SVF2. Those perimeters differ. Son later admitted paying too high a valuation and believing too much in the entrepreneur; 2022 defense mode and leadership changes are documented, but their exclusive causal origin in WeWork is not. [contemporaneous record · strong] FY2019 Q2 deck, slides 4–5 and 20–29 FY2025 report, p. 91 failure packet

2. DiDi — category leader with regulatory and listing risk

Ex ante (2017–30 June 2021). Son treated ride sharing as a portfolio of geographic leaders, making DiDi consistent with a number-one/network thesis. Yet the public case did not state a China cybersecurity protocol, foreign-listing contingency or country concentration budget. DiDi's later filing showed the SVF entity with 20.0% economic interest and 10.7% voting power; SBIA, not Son personally, was manager. Alternatives included a smaller position, pre-IPO reduction, waiting for written regulatory status, or allocating across regions. [investor-stated · strong] Milken Japan, 24:42–31:45 DiDi 2021 20-F, shareholders

No-hindsight IC worksheet. (1) Invest/hold? Only after counsel documents data, listing and regulator clearance. (2) Kill? Unresolved cybersecurity review, app-removal/new-user risk, or listing downside that cannot be priced. (3) Ownership/terms? Country cap, staged capital, information rights and a board/regulatory committee; avoid financing against assumed listing liquidity. (4) Next check? Written regulator status, privacy remediation, subsidy-normalized unit economics and voting power sufficient to influence compliance. [researcher inference · strong] DiDi 2022 20-F, cybersecurity section

Actual outcome and process evidence. Two days after the NYSE listing, authorities suspended new-user registration, followed with app removals and rectification; the company later delisted, paid an RMB8.026B administrative fine and resumed registrations in 2023. At 2026-03-31 SBG reported $12.073B cost, $3.980B gross return, $8.093B gross loss and 0.3x, including an OTC residual rather than all cash. Son's later “leader but overpaid” admission fits the public selection logic, and his 2026 China-stop statement documents geographic reallocation; no source proves DiDi alone caused it or identifies internal dissent. [contemporaneous record · strong] DiDi 2022 20-F FY2025 results, p. 78

3. Greensill — correlated finance; Katerra as an outcome-only control

Ex ante (2018–October 2019). Lex Greensill later testified that SVF invested more than $1.5B of equity at a $3.5B valuation, about 20 times one-year-forward earnings, and used a 0%, ten-year bridge automatically converting after approvals. His testimony is a failed founder's account, not an SBG memo. By mid-2019 public concern already covered fund liquidity, customer concentration, GAM/Tim Haywood, Gupta-linked instruments and regulatory experience. Greensill said SoftBank introduced fewer than ten portfolio companies as customers and acknowledged a conflict while asserting separate credit/risk committees; that does not independently prove controls worked. Alternatives were pass, asset-level diligence, a capped exposure, unaffiliated funding or commercial introductions without portfolio financing. [retrospective witness · moderate] Treasury Committee evidence, Q85–Q97 and Q163–Q206

No-hindsight IC worksheet. (1) Invest? Pass or cap pending receivable-level proof. (2) Kill? Insurance loss, irreducible borrower concentration, unverifiable future receivables or related-company credit outside demonstrably independent approval. (3) Ownership/terms? No zero-cost bridge before approvals; independent risk chair, loan-level reporting, concentration covenants and restrictions on related-portfolio products. (4) Next check? Insurer renewal, regulatory capital, top obligors, cash collection and third-party receivable audit. [researcher inference · strong] Treasury evidence, Q92–Q98

Actual outcome and process evidence. Greensill conceded concentrations were unacceptably high and connected insurance nonrenewal/funding withdrawal to its March 2021 administration, while the parliamentary report preserved wider regulatory questions. Katerra failed separately, but the scoped record provides no standalone Katerra decision-date thesis, terms, IC record, warning/governance sequence or stop decision; Katerra is excluded from the complete-lab count and retained only as an outcome-linked partial control. SBG disclosed plaintiffs' allegation that SBG/SVF/Greensill orchestrated a Katerra receivable restructuring causing a $440M loss and the SoftBank parties' denial; it is contested, not adjudicated. No public security-level SVF cost/recovery or Son-authored case lesson was found. Later valuation/risk controls and defense posture exist, but no source causally attributes one to this paired analytical unit. [contemporaneous record · strong] Treasury report, chapter 1 FY2024 financial report, Note 46(3)(a), p. 138

4. OneWeb — ecosystem logic versus financing dependence

Ex ante (2016–2017). SBG linked OneWeb's planned low-earth-orbit constellation with Arm chips, Sprint spectrum and connecting unserved populations. This was strategic complementarity and a deployment plan, not operating cash flow. Alternatives were no investment, milestone tranches, vendor partnership or a diversified consortium. Capital intensity, launch/execution risk, approvals and a long pre-cash interval were knowable; no public IC memo, launch-release schedule, named dissent or Son-only instrument decision is available. [investor-stated · strong] FY2017 report, pp. 13–15

No-hindsight IC worksheet. (1) Invest? Only through milestone tranches sized for launch and refinancing failure. (2) Kill? Missed spectrum/launch milestones, no diversified syndicate, unproven terminal economics or revised plan making SBG lender of last resort. (3) Ownership/terms? Senior downside, board/technical-audit rights, a hard follow-on cap and outside-capital conditions. (4) Next check? Satellites built/launched, tested performance, customer contracts, unit cost, insured-launch record, runway and third-party financing. [researcher inference · strong] FY2017 report, OneWeb plan

Actual outcome and process evidence. OneWeb filed Chapter 11 after failing to secure required funding. SBG recorded a $1.593B FY2019 loss and carrying value zero on $2.023B cumulative exposure: $1.000B common/preferred equity, $487M loans, $426M warrants and $110M shareholder loans. That is escalation across instruments, not an LP-net cash return. No OneWeb-specific Son postmortem or causal control change was found; later operating support, valuation review and defense mode are only group-level changes. [contemporaneous record · strong] FY2019 financial report, Note 6(2), printed p. 14 / PDF p. 18

5. OYO — platform ambition, forecast calibration, and unresolved outcome

Ex ante (2018–March 2019). Son framed OYO as a hotel category rebuilt with data, dynamic pricing and AI and predicted it would soon be the world's largest hotel company. Geographic growth was visible; mature cohorts, partner retention, audit quality, governance and downside criteria were not publicly presented. The later DRHP supplies company/securities/risk evidence, not the missing SVF IC memo or dissent. Alternatives were geography-gated tranches, asset-light partnership, smaller minority or pass. [investor-stated · strong] Milken Japan, 31:45–34:16 Oravel DRHP

No-hindsight IC worksheet. (1) Invest? Only with geography-level cohorts and tranches; global room count is not a moat. (2) Kill? Negative mature-market contribution, partner churn/refunds, unreliable reporting, service quality lagging expansion or capital beyond a fixed reserve. (3) Ownership/terms? Audit/information rights, independent directors, related-party controls, founder-liquidity limits and milestone follow-ons. (4) Next check? Retained hotels/customers, revenue per available room, take rate, contribution, complaints/refunds, burn and tested systems. [researcher inference · strong] Oravel DRHP, risks and metrics

Actual outcome and process evidence. OYO shifted from rapid to sustainable growth; its FY2022 report showed revenue INR47,813.62M versus INR39,616.49M and adjusted EBITDA negative INR4,717.21M versus negative INR17,447.26M. These issuer-selected non-GAAP improvements are retrenchment evidence, not fund recovery. SEBI lists a new updated DRHP filed 30 June 2026, so no realized IPO exit can be claimed. Security-level cost, follow-ons/refusals, proceeds, mark, named postmortem and OYO-caused process change remain unavailable. [contemporaneous record · strong] OYO FY2022 report SEBI 2026 updated DRHP page

6. Amazon — a recollected miss, not a return ledger

Ex ante record. Son said in 2019 that he offered $100M for nearly 30% of private Amazon, Bezos requested $130M, and the $30M gap ended the negotiation. No contemporaneous term sheet, date, capitalization table, Bezos confirmation, SBG board record, business evidence or security was found. The decision is therefore a retrospective witness account, not a verified rejected transaction. [retrospective witness · provisional] Milken Japan, 08:48–10:37

No-hindsight IC worksheet. (1) Invest? Insufficient public decision-time record. (2) Kill? Failure to establish date, security, ownership, governance, dilution, operating evidence or funding capacity. (3) Ownership/terms? Never convert “nearly 30%” into an exact stake or value it using later market capitalization. (4) Next check? Original term sheet, cap table, operating metrics, counterparty corroboration and approval record. [researcher inference · strong] failure packet, Amazon Milken Japan, 08:48–10:37

Actual non-decision and lesson. No investment closed, so there is no actual return. Son used Amazon, eBay, Baidu and Google misses to justify preparing a much larger capital pool so cash scarcity would not recur. That is a documented later lesson, but a hypothetical stake could have diluted, been sold, hedged, taxed or never closed. Larger capital removes one constraint while potentially weakening price and concentration discipline. [investor-stated · moderate] Milken Japan, 08:48–10:37

7. Uber — a category leader with an ordinary 1.2x result

Ex ante (2017–2019). Son described a geographic mobility portfolio including Uber, DiDi, Grab and Ola, with autonomy as later optionality. This diversified issuers/geographies but concentrated subsidy, regulatory, labor, listing and category economics. Public sources do not disclose a cross-company scorecard, aggregate category cap, valuation ceiling or dissent. Uber ATG's separate $1B financing—$333M from SVF and a 13.8% post-money stake for the investor group at $7.25B post—must not be treated as the SVF1 listed-Uber check. [investor-stated · strong] Milken Japan, 24:42–31:45 Uber amended S-1, ATG investment terms

No-hindsight IC worksheet. (1) Invest? Only at a price supported by subsidy-normalized cohorts and a defined regulatory/labor case. (2) Kill? No path to contribution after driver incentives, persistent safety/regulatory failures, or autonomous optionality required to justify the base valuation. (3) Ownership/terms? Information, board/observer, pro-rata and orderly-liquidity rights with a category exposure cap. (4) Next check? Cohort contribution, retention, take rate, insurance/safety, free cash flow and economics excluding ATG narrative. [researcher inference · strong] Uber amended S-1, risk factors and ATG terms

Actual outcome and process evidence. At 2026-03-31 SBG classified Uber as fully exited: $7.666B cost, $9.215B gross return, $1.549B gross gain and 1.2x. It is a positive but ordinary gross fund row before taxes, financing liabilities, fees and expenses—not evidence that category leadership delivered power-law performance. No complete follow-on/refusal schedule, case-specific Son postmortem or Uber-caused control change was found. [contemporaneous record · strong] FY2025 results, SVF1 table, p. 78

8. Coupang — a large winner with founder control and retained-mark risk

Ex ante (2015–2017). SBG described Coupang as a fast-growing Korean e-commerce company likely to benefit from the information revolution. The public record does not disclose initial unit economics, valuation, competitors, staged alternatives or dissent. By IPO, the prospectus showed 568,156,413 shares attributable to SVF—39.4% pre-offering and 36.8% post—but only 8.6% post-offering votes because founder Bom Suk Kim held 76.7% of voting power. SBIA UK held exclusive final fund decision authority; Lydia Jett sat on the board but disclaimed voting/dispositive authority over the fund shares. [contemporaneous record · strong] FY2016 report, pp. 10–11 Coupang 424B4, pp. 178–179

No-hindsight IC worksheet. (1) Invest/hold? If repeat purchase, fulfillment density and mature-market contribution support price; discount founder-vote risk. (2) Kill? Cohort degradation, capex without density gains, founder control blocking remediation or IPO assumptions requiring implausible margins. (3) Ownership/terms? Information, board, pro-rata, related-party and transfer protections, with economics assessed separately from votes. (4) Next check? Active-customer retention, revenue/customer, fulfillment utilization, free cash flow and governance incidents. [researcher inference · strong] Coupang prospectus

Actual outcome and process evidence. At 2026-03-31 the public row reported $2.729B cost, $11.821B gross return, $9.092B gain and 4.3x. With no fully-exited designation, return includes residual marked securities/hedges and is not all cash. The case supports patient exposure to a genuine operator-led leader, but not Son-only credit; no full follow-on/refusal schedule, final exit or Coupang-caused process change is public. [contemporaneous record · strong] FY2025 results, p. 78

9. Grab — regional leadership, incentives, and a 0.8x marked outcome

Ex ante (2018–6 March 2019). Son's thesis was Southeast Asian leadership across mobility, delivery and financial-services optionality. It did not disclose profitability thresholds or prove incentive-normalized network effects. Grab's filed subscription agreement records the Series H and amendment chronology involving SVF entities; it does not show Son alone sourced, priced or approved each tranche. Alternatives included a smaller country-gated round, category/country caps, or pass. [investor-stated · strong] Milken Japan, 24:42–31:45 Grab subscription agreement

No-hindsight IC worksheet. (1) Invest? Only after incentive-normalized city cohorts and portfolio category limits. (2) Kill? Growth disappears without subsidies, burn does not decline, financial-services credit losses exceed risk capacity or multi-vertical complexity erodes execution. (3) Ownership/terms? Preferred downside, information/observer rights, follow-on cap and milestones by country/vertical. (4) Next check? Contribution by city, rider/driver retention, incentive share of GMV, burn, credit loss, licenses and runway. [researcher inference · strong] Grab 2021 20-F

Actual outcome and process evidence. Grab's first 20-F disclosed large 2019–2021 losses and preference/listing accounting effects—warning evidence, not a SoftBank intervention. At 2026-03-31 SBG reported $2.993B cost, $2.448B gross return, $545M gross loss and 0.8x. Because the row retained market exposure, the result can change. No complete follow-on/refusal schedule, case-specific postmortem or Grab-caused control change was found. [contemporaneous record · strong] Grab 20-F, p. 64 FY2025 results, p. 78

10. Paytm — payments growth, regulated-affiliate risk, and a 0.7x exit

Ex ante (2017–before 18 November 2021). Son highlighted a 35-fold rise in transaction value from FY2014 to FY2017 as proof of cashless-payment potential. That selected scale metric did not establish take-rate durability, revenue quality, bank compliance or valuation discipline. The prospectus showed SVF India Holdings at 113,262,230 shares/17.3% plus another SVF Panther holding. The pre-IPO shareholder agreement and special rights terminated at IPO, but Part A of the post-listing articles contemplated an SVF director-nomination mechanism while the relevant SVF shareholder held at least 10%, subject to law and a shareholder special resolution. This was neither a guaranteed continuing seat nor a complete governance sunset. One97 and regulated PPBL remain separate entities. [contemporaneous record · strong] 2018 CEO message, Paytm Paytm RHP, pp. 223 and 261

No-hindsight IC worksheet. (1) Invest/hold? Only if payment scale converts to contribution and regulated affiliates pass independent testing. (2) Kill? Repeated restrictions, weak KYC/customer controls, subsidy-dependent monetization, or post-listing influence legally unavailable/falling below threshold without price compensation. (3) Ownership/terms? Audit/information rights, compliance committee, entity separation, orderly sale and confirmation of required approval for any nominee mechanism. (4) Next check? Revenue/take rate, contribution, PPBL audits, complaints/fraud, burn, 10% threshold and actual Part A right status. [researcher inference · strong] Paytm RHP, pp. 223 and 261

Actual outcome and process evidence. RBI's 2022/2024 directions constrained PPBL and later barred new deposits/top-ups and most services while preserving withdrawals of existing balances; this was action against PPBL, not proof listed One97 closed or that Son managed compliance. At 2026-03-31 SBG classified Paytm fully exited at $1.600B cost, $1.056B gross return, $544M loss and 0.7x. No complete follow-on/refusal chronology, Paytm-specific Son lesson or causal control change was found. [contemporaneous record · strong] RBI release, items 1–5 FY2025 results, p. 78

11. Wirecard — strategic intent and disputed financing attribution

Ex ante (24 April–September 2019). Wirecard's offering memorandum described a binding term sheet under which a SoftBank affiliate would invest about €900M through five-year convertibles, initially convertible at €130 into up to 6,923,076 shares, alongside Asian payments/AI/data cooperation. This is issuer advocacy, not independent diligence. The document does not identify a Vision Fund IC decision, Son-personal approval, named dissent or beneficial holders behind the financing. Alternatives were forensic diligence before direct equity, staged conversion, commercial partnership without finance or pass. [contemporaneous record · strong] Wirecard offering memorandum, §6.5.2, pp. 62–63

No-hindsight IC worksheet. (1) Invest? Pass pending independent cash, bank, acquiring and customer confirmation. (2) Kill? Failed confirmations, untraceable cash, regulatory/audit inconsistency, opaque beneficial ownership or conversion economics requiring an unverified narrative. (3) Ownership/terms? Escrowed stages, audit access, independent compliance/board rights, default security and no reputation licensing before financing ownership is known. (4) Next check? Bank confirmations, merchant/acquirer reconciliation, regulator files, concentration, cash conversion and beneficial ownership. [researcher inference · strong] Wirecard memorandum, risks and §6.5.2

Actual outcome and process evidence. Wirecard failed in 2020. The Bundestag record contains testimony/assertions that employees used SoftBank's name for financed speculation rather than SBG placing €900M at risk, while the issuer document called the buyer an affiliate. Without the final financing/recovery chain, this report does not book an SVF loss or Son-personal decision. No case-specific Son postmortem or causal control change was found; the supported rule is to prove beneficial ownership before attributing a branded affiliate transaction. [retrospective witness · moderate] Bundestag report, p. 153

12. Wag — an adverse exit with incomplete terms

Ex ante. FY2019 lists Wag as an SVF1 holding but supplies no thesis, terms, evidence, named decision owner or dissent. The analytical alternatives are staged city exposure, lower-cost partnerships or pass; the public record is insufficient to score what SoftBank actually considered. [contemporaneous record · moderate] FY2019 report, portfolio list, pp. 87–89

No-hindsight IC worksheet. (1) Invest? Only with city-level repeat use and contribution after walker acquisition, insurance, support and promotions. (2) Kill? Weak retention, adverse incidents, no density advantage or capital beyond reserve. (3) Ownership/terms? Downside security, information/board rights and staged funding. (4) Next check? Retained users, fulfilled-walk density, claim rate, contribution and burn. [researcher inference · provisional] FY2019 portfolio list

Actual outcome and process evidence. In December 2019 SVF reportedly agreed to sell its stake back for undisclosed consideration and relinquish its board seat after a reported $300M 2018 investment. This supports an exit/governance stop but not exact loss or MOIC. No Wag-specific Son lesson or process change was found. [contemporaneous record · moderate] Axios, opening four paragraphs

13. Zume — a shutdown without a security-level loss ledger

Ex ante. FY2019 lists Zume but provides no original robotics/food/packaging thesis, security, valuation, team attribution or dissent. Decision alternatives included one-business pilots, customer-funded automation or pass; no primary IC record resolves them. [contemporaneous record · moderate] FY2019 report, portfolio list, pp. 87–89

No-hindsight IC worksheet. (1) Invest? Only by milestones after one repeatable business model is proven. (2) Kill? Pivots before unit economics, capex ahead of demand, excessive food/logistics/robotics complexity or capital beyond reserve. (3) Ownership/terms? Staged preferred capital, board/technical rights and pivot-specific covenants. (4) Next check? Contracts, gross margin, asset utilization, throughput, burn and milestone delivery. [researcher inference · provisional] FY2019 portfolio list

Actual outcome and process evidence. Zume shut down and liquidated in 2023 after reportedly raising $445M from SoftBank and others, including a reported $375M from SoftBank. Round amounts are not security-level cost/recovery or audited fund loss. No Zume-specific Son lesson or causal process change was found. [contemporaneous record · moderate] Axios, opening paragraphs

14. Ziff-Davis and COMDEX — distribution control, leverage, and mixed divestiture

Ex ante (1995–1998). SoftBank argued that COMDEX trade shows plus Ziff-Davis publishing would control technology-promotion channels, share advertising/information and bridge print into electronic content; it projected first-year EPS contribution. The assets became wholly owned consolidated subsidiaries—corporate acquisitions, not fund checks. Alternatives were minority partnerships, selected assets or preserving balance-sheet capacity. The announcement lacked purchase-price allocation, downside, integration gates, dissent and kill rules; by 1999 Forbes reported that acquisitions including these helped leave SoftBank with about $3B debt, a publisher account rather than audited return. [investor-stated · strong] 1995 acquisition announcement Forbes 1999

No-hindsight IC worksheet. (1) Invest? Prefer staged/minority or asset-specific purchases until cross-media synergy is measured. (2) Kill? Leverage breach, faster print decline than digital monetization, weakening events or failure of projected EPS/cross-sell. (3) Ownership/terms? Ring-fenced debt, indemnities, management retention, accountability and divestiture triggers. (4) Next check? Segment cash flow, advertisers, digital revenue/audience, bookings, interest coverage and realized cross-sell. [researcher inference · strong] 1995 announcement

Actual outcome and process evidence. FY2000 recorded a ¥119.1B one-time evaluation loss on Ziff-Davis intangibles/restructuring; later discontinued-operation losses followed. The CNET/ZDNet combination produced a disclosed ¥29.4B share-sale gain, but the exchange itself was noncash and left an 18.1% CNET stake. Management divested publishing, separated exhibitions and shifted architecture. These actions are documented; a complete acquisition/debt/operations/sale/tax return and Son-authored causal postmortem are not. [contemporaneous record · strong] FY2000 report, pp. 35–37 CNET/Ziff completion

15. Japan Telecom — control acquisition, operating pivot, and impairment

Ex ante (May–July 2004). SBG bought 100% to combine roughly 12,000km fiber, enterprise customers and about 10M access lines with its broadband base. The roughly ¥340B announced perimeter comprised ¥143.3B common-stock purchase, estimated ¥164.0B net debt and ¥32.5B preferred stock. The target's FY2004 sales were ¥347.2B with ¥50.6B net loss. Alternatives were wholesale access, partial ownership or organic build. Public materials do not reveal bid comparison, synergy sensitivity, dissent or stop criteria; Hideki Kurashige and operating teams remain separate from Son's sponsorship. [contemporaneous record · strong] May 2004 announcement July completion

No-hindsight IC worksheet. (1) Acquire? Only if enterprise cash flow and replacement value cover leverage before consumer synergies. (2) Kill? Interest-coverage failure, consumer economics below plan, duplicate capex or no measurable cross-sell. (3) Ownership/terms? Full control with ring-fenced debt, integration gates, management accountability and asset-sale options. (4) Next check? EBITDA-to-cash conversion, churn, utilization, Otoku Line contribution, capex, debt service and cross-sell. [researcher inference · strong] May 2004 announcement

Actual outcome and process evidence. Japan Telecom shifted Otoku Line emphasis from consumers toward corporate customers. SBG recognized an ¥18.0B fixed-line impairment within a ¥36.4B group fixed-asset impairment because consumer-related/suspended facilities were not expected to be used. A later acquisition note reported ¥189.959B cost, ¥29.462B loans, ¥60.584B cash acquired and ¥158.837B payments—an accounting perimeter legitimately different from announced enterprise-value components. The operating reset is documented; a standalone return or Son-authored process lesson is not. [contemporaneous record · strong] 2006 impairment release FY2006 report, p. 91

Portfolio forensics — what the distribution says

Stage and check-size evolution

SoftBank's historical “portfolio” is not one venture vintage. Early Yahoo/Alibaba positions were corporate or subsidiary interests with very long duration; Yahoo! BB was an organic build; Vodafone, Japan Telecom, Sprint and Arm were corporate control acquisitions; SVF1 was a late-stage scale fund; SVF2 moved toward earlier stages; OpenAI is again a giant late-stage minority and Ampere a control acquisition. At March 2019 SVF1 reported 69 companies and $60.1B cumulative acquisition cost. By FY2022 SBG described SVF2 as typically investing $100–200M for 10–15% ownership and held more than 250 companies after deploying $46.9B. Smaller checks reduce single-name exposure but can increase selection and monitoring load; 208 investments in one year do not by themselves prove better diversification. [contemporaneous record · strong] FY2019 report, p. 25 FY2022 report, pp. 22–25

The observed check distribution has a long tail plus exceptional concentration. The live portfolio page omits checks; historical filings reveal a few very large positions—DiDi $12.073B cost in the current public table, Uber $7.666B, WeWork's multi-vehicle exposure, Arm's £24B acquisition, and OpenAI's current tens of billions—alongside hundreds of undisclosed/smaller names. A median check, weighted ownership or concentration ratio for the 830 observations cannot be calculated from available rows. [researcher inference · strong] FY2025 results, p. 78 denominator packet, ledger field controls

Sector and factor concentration

The 2019 portfolio taxonomy spread across consumer, enterprise, fintech, frontier tech, health, real estate and transportation/logistics. Economic factors were less diverse. Uber/DiDi/Grab/Ola shared mobility subsidies, regulation and listing windows; WeWork/OYO/Compass shared physical-network execution and private-market marks; Greensill/Katerra linked portfolio financing and counterparty risk; Arm/NVIDIA/Ampere/OpenAI/Stargate now share AI-compute demand, capital markets, power and infrastructure. The correct concentration analysis looks through sector labels to cheap-growth capital, founder votes, private marks, IPO liquidity, regulation and follow-on dependence. [researcher inference · strong] FY2019 report, portfolio map, pp. 87–89 failure packet, portfolio patterns

Geography and regulation

At March 2020, SVF1's 91-company/$81.0B acquisition-cost book was reported as 36% Americas, 18% EMEA and 46% Asia. This distribution did not diversify shared liquidity or regulatory shocks. DiDi exposed PRC data/listing risk; Paytm separated listed-company and regulated-bank risk; Grab operated across multiple Southeast Asian regimes; Sprint's exit required DOJ/FCC remedies; Arm's proposed NVIDIA sale failed under multi-jurisdiction competition review. Son's later statement that SBG stopped new China investing is a documented geographic change, not proof that geographic diversification previously worked or that DiDi alone caused the stop. [contemporaneous record · strong] FY2020 report, pp. 20–21 TIME, China discussion

Ownership, governance, and follow-on behavior

Economic ownership frequently exceeded governance. WeWork's proposed 80% economics did not convey accounting control; DiDi's 20.0% economics carried 10.7% votes; Coupang's 36.8% post-IPO economics carried only 8.6% votes; Paytm's pre-IPO SHA ended while a conditional articles-based nomination path depended on a 10% threshold and approvals. Conversely, corporate purchases of Vodafone, Sprint, Arm and Japan Telecom created control but exposed SBG to operating, debt and regulator obligations. An IC ledger that records only ownership percentage is therefore incomplete. [researcher inference · strong] FY2019 Q2 deck, accounting slides 4–5 Coupang 424B4, pp. 178–179 Paytm RHP, pp. 223 and 261

Follow-on capital shaped losses as much as initial selection. WeWork escalated across convertibles, warrants, preferreds, joint ventures and support finance; OneWeb escalated across equity, loans, warrants and shareholder loans. Both cases argue for a fresh IC, maximum total exposure and outside-financing proof at every follow-on. Coupang is the necessary adverse control to a blanket “never double down” rule: retained exposure preserved substantial upside. The supported rule is conditional re-underwriting, not automatic rescue or automatic abandonment. [researcher inference · strong] WeWork S-1/A, pp. 118–119 OneWeb table, printed p. 14 / PDF p. 18 FY2025 results, Coupang row, p. 78

Exit modes and duration

The disclosed outcomes include full cash exits, partial sales, residual public marks, OTC securities, mergers, in-kind transfers, collars, deposits, impairments and bankruptcy claims. Uber and Paytm were labeled fully exited; Coupang and Grab retained exposure; DiDi remained OTC; NVIDIA was a profitable collar-assisted exit; Sprint became shares, contingent shares and options; Arm combined IPO sale and retained control. Exit policy therefore needs a thesis horizon, liquidity/collateral hierarchy, valuation re-underwriting, governance constraints and security-specific tax/financing treatment. A single “hold winners forever” rule ignores solvency; a single “take profit at 2x” rule ignores power-law duration. [researcher inference · strong] FY2025 results, p. 78 FY2019 report, NVIDIA note

Reported performance snapshots—kept incomparable

The public tail prevents a winners-only reading. SVF1's 2026 table also showed Getaround $348M/zero, Ginkgo $404M/$4M, View $1.175B/zero, OneConnect $100M/$6M and GoTo $841M/$363M. SVF2's public table included Klarna $1.706B/$202M, Better $497M/$44M, JD Logistics $601M/$205M, DingDong $325M/$36M, Ola Electric $556M/$279M, Chime $350M/$95M and Pear $46M/zero. These are comparators, not complete failure labs: the table lacks original thesis, security, dissent, team, warnings and process response. [contemporaneous record · strong] FY2025 results, pp. 78–79

Cross-case failures, misses, and postmortems

Documented later process changes—and causal limits

Operational rules supported by the record

1. Freeze the memo. Record thesis, evidence, base rate, alternatives, valuation, security, ownership/votes, decision owners, dissent, reserve and kill criteria before approval. Founder enthusiasm and market rank are not reproducible process. [researcher inference · strong] WeWork S-1, risks
2. Underwrite company and price separately. Uber 1.2x, Grab 0.8x, DiDi 0.3x and Paytm 0.7x show that a leader can be a middling or losing security. [researcher inference · strong] FY2025 results, p. 78
3. Re-IC every follow-on. Reset the thesis, price, governance, runway, outside-funding proof and maximum total exposure; sunk cost and reputational rescue are not evidence. [researcher inference · strong] OneWeb exposure table, p. 14 / PDF p. 18
4. Treat governance as a security. Map economics, votes, board rights, manager authority, regulated affiliates, conflicts, audit/enforcement and their duration. [researcher inference · strong] Coupang prospectus, pp. 178–179
5. Separate solvency from duration. LTV/cash floors protect the parent; they do not decide whether a liquid winner's thesis has matured. Record re-underwriting and hedge/collar rules. [researcher inference · strong] FY2019 NVIDIA note
6. Report distributions, not stories. Show defined cohorts, gross/net, realized/unrealized, security-level cash flows and outlier-excluded views. [researcher inference · strong] FY2025 results, pp. 78–79

Dominant-outlier exclusion tests

Corporate-history scorecard: remove Alibaba

The FY2015 selected 23-position table reported ¥387.8B accumulated investment and ¥10,653.5B estimated value/accumulated return, summarized as 27x and 45% IRR. Alibaba alone contributed ¥10.5B of the reported investment and ¥7,993.4B of value/return—about 75% of the table's reported value/return, not 80%. Mechanically removing it leaves ¥377.3B against ¥2,660.1B, roughly 7.1x before considering timing. Removing Yahoo Japan as well leaves ¥369.6B against ¥1,301.2B, roughly 3.5x. These are researcher subtractions from an outcome/size-selected management table, not portfolio returns or recomputed IRRs. [researcher inference · strong] FY2015 report, complete table and notes, printed p. 26 / PDF p. 28

That residual still contains large successes—Yahoo U.S., Trend Micro, SBI, GungHo—and severe losses such as CNET at 0.1x and Asia Global Crossing at zero. But because the table's admission rule favored accumulated investment or return of at least ¥10B and current listed affiliates, even the 7.1x residual cannot estimate all-history performance. The valid conclusion is narrower: Alibaba dominates the best published corporate scorecard; Yahoo Japan provides a second large contributor; without both, the selected record remains positive but far less extraordinary. [researcher inference · strong] denominator packet, FY2015 full ledger FY2015 report, selected table and notes, printed p. 26 / PDF p. 28

SVF1 public book: remove current disclosed leaders

At 2026-03-31 the disclosed SVF1 public-company table reported $46.966B cost, $57.572B gross return, $10.606B gross gain and 1.2x. Coupang, DoorDash, NVIDIA and Guardant Health contributed $21.650B of rounded gross gains—more than twice the public book's net gain—so other rows offset more than half of those four gains. Removing Coupang alone leaves $44.237B cost, $45.751B return and $1.514B gain, about 1.03x. Removing DoorDash as well leaves $43.557B cost, $37.807B return and a $5.750B loss, about 0.87x. [researcher inference · strong] FY2025 results, SVF1 table and notes, p. 78

This second test is also bounded. It excludes SVF1's $38.215B private-company/etc. cost and $51.873B gross return bucket, which includes Arm, and it does not adjust fees, tax, financing, timing, LP interests or intra-group Arm transactions. It is not an adjusted fund multiple. It does establish that the disclosed public book's sign and magnitude depend heavily on a few names, so principles repeated in speeches cannot stand in for distributional evidence. [researcher inference · strong] FY2025 results, p. 78 and note 11 talks synthesis

Overall verdict without Alibaba

Without Alibaba, the record still supports several real capabilities: recognizing an attention gateway; localizing it through a JV; financing broadband/mobile access; executing at least one difficult carrier turnaround; acquiring and retaining foundational semiconductor IP; and producing some fund winners. It also becomes materially harder to defend a general selection edge. Yahoo/Vodafone/Arm mix corporate, operating and marked effects; Coupang and DoorDash are gross fund rows; NVIDIA was sold profitably but early; Sprint has a company-calculated mixed-perimeter return; OpenAI remains unrealized. No defensible common denominator or currency/timing-consistent cash series permits an aggregate residual multiple. [researcher inference · strong] major packet, cross-case implications 2023 CFO message, Alibaba perimeter FY2025 public-return table, p. 78 FY2026 report, OpenAI, printed pp. 23–25

The practical answer to “what drove the record?” is therefore: one exceptional, largely monetized Alibaba outcome; a smaller set of durable platforms/control assets and fund winners; and a long tail that contains ordinary returns, marks, write-offs, disputed cases and unknowns. That is a power-law-shaped assessment, not a sourced literal 80/20 law. [researcher inference · strong] major-case contribution audit 2023 CFO message, Alibaba perimeter FY2025 SVF1 public-return distribution, p. 78

Uncertainty and retrieval ledger

Access boundary. Official annual-report PDFs, SEC/issuer filings, regulator/court/parliamentary materials, transaction releases, the live manager index, and selected credible adverse reporting were used. Video evidence was limited to official or already verified timestamp/caption routes in the frozen talks corpus; slides were not treated as spoken words. No audio was downloaded or transcribed, and no local or non-OpenAI transcription was used. [researcher inference · strong] talks access ledger failure packet, final boundary

Final evidence rule. A public mark may be informative without being cash; a commitment may be strategically meaningful without being funded; a famous founder story may explain intent without proving team attribution; and a repeated principle may be consistent without being true. Where a complete cash flow, approval record or denominator is absent, this synthesis preserves unknown rather than converting narrative into performance. [researcher inference · strong] denominator packet, field controls

Coverage and evidence base

This corpus contains 35 underlying works across 36 legitimate routes: 16 ★★★, 11 ★★, and 8 ★. Every work has a source note, a documented disposition, and its own subsection below, yielding 100% disposition for ★★★ (16/16), ★★ (11/11), and ★ (8/8). Thirty-four full-access works—15 ★★★, 11 ★★, and 8 ★—were deeply content-analyzed. The remaining ★★★ work is Alan M. Webber’s 1992 Harvard Business Review interview: its public page exposes the title, byline, issue, and purchase controls but not the body, so it was lawfully inventoried and access-audited, not content-analyzed, and no doctrine is inferred from it. HBR public-page access boundary

By route, the same inventory is 17 ★★★, 11 ★★, and 8 ★. The 36-route count is one greater than the work count because the 2010 Next 30-Year Vision was checked through both the 133-page official deck and its related announcement; those are two routes to one underlying work, not independent corroboration. deck, PDF pp. 1–133 related announcement, paragraphs 7–9

Most of the record is signed or first-person SoftBank annual-report material. The doctrine can therefore be attributed to Son where the text does so, but charts, footnotes, calculations, transaction states, and selected examples remain corporate disclosures, not proof of Son's sole authorship or personal economics. Forbes, Reuters/Wired, HBR, and Asahi are edited publications: speaker-labeled answers remain Son's voice; reporter synthesis and named witnesses remain theirs. Throughout, SBG, its controlled operating companies, SoftBank-managed funds, outside LP capital, portfolio-company teams, and Son's personal holdings are kept separate. Marks, NAV, commitments, announced projects, and conditional transactions are never treated as realized cash.

Source-by-source analysis

Access and attribution convention for every subsection. Unless a subsection states the sole HBR partial boundary, the linked work was accessed in full. Signed or first-person official messages support Son attribution for the doctrine; their charts, notes, translations, and calculations remain corporate-report apparatus. Edited interviews and reports preserve only speaker-labeled Son remarks as direct voice; reporter and witness analysis remains separately attributed.

1. Japanese-Style Entrepreneurship — HBR interview (1992)

• Context, attribution, and access — contemporaneous record · strong. HBR identifies Alan M. Webber as author/interviewer, Son as SoftBank CEO, and January–February 1992 as the issue. The interview body was not publicly visible, so the record establishes an early direct-voice route but no investable proposition. title, byline, issue, and access boundary
• Argument, reasoning, evidence, and assumptions — researcher inference · insufficient public record. None can be reconstructed from the headline or metadata. The title cannot support claims about Son's entrepreneurship, Japanese management, biography beyond the named role, or investment process. public-page boundary
• Rule, implication, and evolution — researcher inference · strong. The only decision-useful rule is evidentiary: retrieve a lawful full copy, separate Webber's framing from Son's answers, and withhold doctrinal comparison with the late-1990s Internet record until page-level text exists. The missing body is the corpus's sole access exception. public-page boundary

2. Annual Report 1998 — To Our Shareholders (1998)

• Context and argument — investor-stated · strong. Against weak Japanese and Asian conditions and lackluster earnings, Son narrows a broad digital-information portfolio toward Internet services. His four-stage history moves from analog technology and services to digital technology and then digital services, which he expects to become the largest layer. printed pp. 2, 4; PDF pp. 4, 6
• Reasoning, evidence, and assumptions — investor-stated · moderate. Yahoo Japan's listing, Yahoo, GeoCities, ZDNet, and the public value of Ziff-Davis are selected evidence that traffic, service ownership, and listings can expose value. The examples are marks, not a complete return denominator, and assume page-view leadership will translate into defensible economics. printed pp. 2–5; PDF pp. 4–7
• Rules, operations, and implications — researcher inference · moderate. Seek Internet leaders by traffic and unique users, use subsidiary-managed venture funds for option-like relationships, localize U.S. models into adoption-lagging Japan, and list mature companies to reveal value. The more-than-60-company count belongs to group funds, not Son personally; modern use requires retention, monetization, localization, price, governance, and a loss denominator. printed p. 5; PDF p. 7

3. Annual Report 1999 — Dear Shareholders (1999)

• Context and central argument — investor-stated · strong. Son dates the Internet-centric turn to a January 1, 1996 email asking the board to focus resources on Internet opportunities. The strategy is sequential: secure information flow through portals, place digital finance on that traffic, then add goods and services. printed pp. 2–4; PDF pp. 3–5
• Reasoning and evidence — investor-stated · moderate. Yahoo is treated as digital prime real estate: despite about $1.6 million of sales and a roughly $1 million loss when visited, search could aggregate exploding information. Finance should move first because it is database-heavy and physically light. The logic is coherent, but the choice is also described as intuitive and gives no valuation ceiling or losing-deal denominator. printed pp. 2–4; PDF pp. 3–5
• Rules, organization, and evolution — contemporaneous record · strong. Secure the aggregation point, prioritize low-friction digital categories, take meaningful minority positions, localize through joint ventures, and preserve operating autonomy while centralizing cash allocation. The approved holding-company structure, majority-external board, and monthly subsidiary forecasts show an attempt to institutionalize the method; they do not establish security-level diligence. printed pp. 3–5; PDF pp. 4–5

4. Master of the Internet — Forbes profile/interview (1999)

• Context, attribution, and argument — contemporaneous record · moderate. Neil Weinberg reconstructs a method of early, unusually large bets followed by managerial autonomy, while preserving the skeptical alternatives of breadth, luck, a rising Internet market, and bubble-valued public holdings. That thesis is Forbes's synthesis; only speaker-labeled passages are Son's own words. paragraphs beginning “Venture capitalists,” “That's vintage,” and “Just about anyone”
• Reasoning and evidence — contemporaneous record · moderate. The Yahoo case includes a team model of distribution-cost removal and 30% pretax margins, followed by a $2 million 1995 test, a $105 million offer for 33% in 1996, and later purchases by SoftBank. The E-Trade case records rapid $9 million and later $400 million decisions and a contemporaneous $2.4 billion mark. These are corporate or venture-vehicle commitments and public marks, not Son-personal cash returns. Yahoo and E-Trade passages
• Portfolio, team, and counterevidence — contemporaneous record · strong. Gary Rieschel ran the $230 million first SoftBank Technology Ventures pool across 55 companies; American Cybercast and OnLive failed. Forbes also reports a mixed more-than-100-company denominator and separates SoftBank, its venture arm, and Son's 42% corporate holding. Category mapping and staged top-ups therefore reflect team and vehicle machinery, not lone intuition. paragraphs beginning “All told,” “In 1995 Son discovered,” and “In a frenetic first year”
• Decision rule and implications — researcher inference · moderate. Enter a secular service shift early, start with a falsifiable commercial model, top up when operating evidence improves, preserve management autonomy, and measure portfolio alliances in transactions or revenue. The rule fails if top-ups merely chase prices, breadth buys market beta, or network adjacency never creates economics. Yahoo, STV, alliance, and failure passages

5. Annual Report 2000 — Message from the President (2000)

• Argument and context — investor-stated · strong. Son formalizes the fourth-stage focus on digital information services and cites the sale of Kingston and most Ziff-Davis assets as evidence that management resources should leave digital hardware and legacy publishing. printed pp. 2–4; PDF pp. 35–37
• Reasoning, examples, and assumptions — investor-stated · moderate. U.S. venture operations discover models, geographic and business vehicles transplant them, and joint ventures scale faster than internal invention. Yahoo leads the examples and the report counts more than 400 group and fund-investee companies; the count mixes controlled and noncontrolled entities and says nothing about realized portfolio economics. printed pp. 4–5; PDF pp. 37–38
• Operating rule and evolution — documented behavior · moderate. A three-tier structure assigns group capital and strategy to SoftBank, coordination to operating holding companies, and cash generation to operators, with a CEO Council, board supervision, and monthly corporate-value review. It is a more explicit operating system than 1999, but still lacks security-level authority, ownership returns, reserves, and named accountability. printed pp. 5–6; PDF pp. 38–39

6. Softbank's Son: “Cannibalize!” — Reuters/Wired report (2000)

• Context and attribution — contemporaneous record · strong. This Reuters dispatch reports Son's World Economic Forum remarks alongside Barry Diller, Michael Dell, and Michael Capellas. Their comments provide event context but are not Son's doctrine. paragraphs 74–96
• Argument and reasoning — investor-stated · moderate. An incumbent confronting e-commerce should isolate the Internet unit, appoint separate leadership, and authorize it to attack the legacy business. Son's “zero and infinity” framing assumes digital distribution compresses time and variable cost while expanding inventory, information, and reach; the dispatch supplies no unit economics or SoftBank case validating that abstraction. paragraphs 74–90
• Rule, operations, and limits — researcher inference · moderate. Make self-cannibalization a governed mandate rather than a side project, but specify ownership, customer conflict, transfer pricing, resource authority, and shutdown tests. The rule is a sharp organizational complement to the 2000 holding-company design; separation alone is not evidence of a viable challenger. paragraphs 74–79

7. Annual Report 2001 — Message from the President (2001)

• Context and argument — investor-stated · strong. After Internet valuations became volatile, Son separates market price from continuing adoption and technology progress. The new operating focus is broadband access plus transaction platforms: infrastructure has become the bottleneck preventing Internet services from flourishing. printed pp. 2–5; PDF pp. 63–66
• Reasoning and evidence — investor-stated · moderate. User growth, South Korea's broadband lead, Japanese deregulation, and a national connectivity goal support the view that low-priced always-on access can unlock creation and commerce. Yahoo! BB and procurement platforms are concrete actions, but projected adoption is not a return model. printed pp. 3–5; PDF pp. 64–66
• Rule, implication, and evolution — researcher inference · moderate. Ignore price volatility only when adoption and technical evidence improve; identify the binding bottleneck; exploit regulatory openings; and pair access with monetizable transactions. This materially shifts SoftBank from minority optionality into owned, capital-intensive infrastructure, so funding runway and stop conditions become indispensable. printed pp. 2–5; PDF pp. 63–66

8. Annual Report 2002 — To Our Shareholders (2002)

• Context and argument — investor-stated · strong. After an ¥88.755 billion consolidated net loss tied to broadband start-up costs, weak equity markets, valuation losses, and accounting change, Son concentrates rather than abandons the thesis. Leadership must span infrastructure, platforms, and content/services and eventually produce stable profit and cash flow. printed pp. 2–3; PDF pp. 4–5
• Capital evidence and assumptions — contemporaneous record · strong. Security-sale proceeds funded bond redemption and repurchases while Yahoo! BB built a nationwide low-price ADSL network. This shows offense alongside deleveraging, but not that the remaining concentration was economically optimal or sufficiently funded. printed pp. 2–3; PDF pp. 4–5
• Rules and implications — researcher inference · strong. Direct capital toward the bottleneck, seek scale across complementary layers, sell securities to protect solvency, and judge the program by cash flow rather than marks. A contemporary IC would require burn, liquidity, deleveraging, contribution-margin, and abandonment gates; continued conviction is not itself discipline. printed pp. 2–3; PDF pp. 4–5

9. Annual Report 2003 — To Our Shareholders (2003)

• Context and argument — contemporaneous record · strong. Released in July 2003 for the year ended March 31, 2003, the report records a ¥99.989 billion consolidated net loss, exceeding the prior year's loss. Son identifies Broadband Infrastructure start-up costs, acquisition expense, and equipment depreciation, while saying the rest of the group reached operating profitability; he still argues for accepting upfront losses to layer services onto a low-cost owned network. signed message, printed pp. 2–3; PDF pp. 4–5
• Evidence and assumptions — investor-stated · moderate. The report gives two correctly separated subscriber dates—about 2.36 million installed lines at March 31 and 2.82 million at June 2003—and a funding bridge from cash and asset sales to broadband investment and debt reduction. It provides no cohort margin, payback, cumulative-burn cap, or hurdle rate. printed p. 10; PDF pp. 12, 17
• Rules, operations, and evolution — researcher inference · strong. Separate acquisition cost and depreciation from group performance, hold measures to fixed dates, diversify funding, reduce debt concurrently, and demand eventual segment cash flow. Compared with 2002, the doctrine becomes more falsifiable because it names costs and funding, though still not the stop condition. signed message, printed p. 2; PDF p. 4

10. Annual Report 2004 — Message from the President (2004)

• Argument and context — investor-stated · strong. Broadband becomes the substrate joining television, telephony, and microprocessors; the ambition shifts from access to a “lifestyle company” combining infrastructure, distribution, and content. printed pp. 2–3; PDF pp. 4–5
• Evidence and counterevidence — contemporaneous record · strong. Broadband sales exceeded ¥100 billion and the segment loss narrowed, yet the group still reported a ¥107.0 billion net loss. An unauthorized acquisition of Yahoo! BB customer information led to 649 reported countermeasures—an early trust and governance failure that the original growth doctrine had not priced. printed pp. 2–3; PDF pp. 4–5
• Rules and implications — researcher inference · strong. Build infrastructure, use distribution to acquire customers, add content, and extend service to difficult geographies, but measure attach rate, margin, retention, security incidents, and remediation efficacy together. Segment improvement cannot validate the platform while group losses and trust failures persist. printed pp. 2–3; PDF pp. 4–5

11. Annual Report 2005 — Message from the President (2005)

• Argument and reasoning — investor-stated · strong. As broadband commoditizes access, competition moves to services and content. Son combines infrastructure cash flow, Yahoo Japan reach, and assets such as GungHo and the Hawks into a lifestyle platform; breadth is offered as synergy potential, not evidence that coordination exceeds its cost. printed pp. 4–5; PDF pp. 6–7
• Operating and financial evidence — contemporaneous record · strong. ADSL reached quarterly operating surplus while debt rose with Japan Telecom, Cable & Wireless IDC, and Hawks acquisitions. The evidence is a transition marker, not proof from a single quarter that the acquired platform earns its cost of capital. printed p. 5; PDF p. 7
• Rules, governance, and evolution — documented behavior · moderate. Use low-price infrastructure as a wedge, lift revenue through services, acquire customer bases and networks, and add group controls as complexity rises. Planned internal-auditor, disclosure, and security changes record intent rather than effectiveness; any modern adjacency acquisition needs attach economics, integration owners, leverage limits, and an organic-build counterfactual. printed pp. 4–5; PDF pp. 6–7

12. Annual Report 2006 — To Our Stakeholders (2006)

• Context and argument — contemporaneous record · strong. SoftBank returned to profit after broadband investment and then acquired Vodafone K.K. Son's “three-dimensional simultaneous growth” expands customers, services/content, and distribution beyond the existing network and Japan; mobile is presented as the next distribution layer, not a standalone carrier bet. printed pp. 6–9; PDF pp. 8–11
• Reasoning, examples, and assumptions — investor-stated · moderate. Yahoo! Streaming is the proposed video-era analogue to Yahoo Japan, while Vodafone contributes devices, software, content, and global distribution. These are ex-ante analogies before turnaround evidence and assume the portal/content advantage can transfer into a leveraged telecom acquisition. printed pp. 8–9; PDF pp. 10–11
• Rules, governance, and implications — researcher inference · strong. Harvest operating leverage after infrastructure scale, buy distribution when organic entry is too slow, and balance growth, profitability, and safety. A revised charter, conduct code, and security committee are described controls, not tested controls; build-versus-buy logic, covenant headroom, integration owners, milestones, and downside triggers remain the practical requirements. printed pp. 6–10; PDF pp. 8–12

13. Annual Report 2007 — CEO's Message (2007)

• Context and central argument — contemporaneous record · strong. One year after Vodafone K.K., the mobile segment drove growth, subscriber-addition share rose from 3.5% to 14.2%, and operating income reached ¥155.7 billion. Son's turnaround sequence is to fix inherited weaknesses, improve quantity and customer-valued quality together, and then deploy differentiated services and pricing. printed pp. 13–18; PDF pp. 15–20
• Evidence and operations — documented behavior · strong. The group expanded 3G base stations and handset variety, launched Yahoo! Keitai and streaming, separated handset payments, and introduced White Plan pricing. These are operational interventions and first-year outcomes, not a lifetime acquisition return. printed pp. 16–18; PDF pp. 18–20
• Capital rule and evolution — researcher inference · strong. Use operating free cash flow for acquisition debt, diversify finance, and add outside-director oversight while pursuing the turnaround. This is more falsifiable than the 2006 strategic analogy, but a modern board still needs cohorts, unlevered returns, milestones, and named owners before calling success. printed pp. 19–20; PDF pp. 21–22

14. Annual Report 2008 — CEO's Message (2008)

• Argument and evidence — investor-stated · strong. Mobile phones are becoming “Internet Machines,” shifting power toward portals, content, services, and usable devices as well as networks. Subscriber additions, White Plan, installment sales, Yahoo! Keitai, tailored handsets, and the iPhone 3G launch support operating momentum but not full investment return. printed pp. 13–18; PDF pp. 15–20
• Geography and portfolio reasoning — investor-stated · strong. Son proposes transferring Japanese mobile-Internet learning into Asia through Alibaba and other partnerships. Mature businesses should supply cash while expanding businesses reinvest; the report frames ROIC minus WACC as the value-creation test. printed pp. 17–19; PDF pp. 19–21
• Rule, assumptions, and evolution — researcher inference · moderate. Stage capital by business maturity, reuse operating learning geographically, and compare return on capital with its cost. The 2007 operating turnaround becomes a portfolio-allocation theory, but supplier power, country risk, partner economics, and ownership-level cash capture are missing; 2009 adds harder debt and free-cash-flow constraints. printed pp. 13–19; PDF pp. 15–21

15. Annual Report 2009 — To Shareholders and Investors (2009)

• Context and central argument — contemporaneous record · strong. After the global financial crisis, Son declares a turn from investment to harvest: positive ¥181.5 billion free cash flow and falling net debt support explicit targets to halve net debt by fiscal 2012 and reach zero by fiscal 2015. A maturing platform should earn market trust through predictable cash, debt reduction, and staged dividends. printed pp. 11, 17–18; PDF pp. 13, 19–20
• Evidence, attribution, and assumptions — investor-stated · moderate. Subscriber growth, data ARPU, churn, bad-debt, inventory, capex, and cross-segment benefits are selected operating evidence, not a full Vodafone acquisition return. Director Kazuhiko Kasai—not Son—is the source for adjacent funding-duration and liquidity detail. printed pp. 12–15, 19–20; PDF pp. 14–17, 21–22
• Rules and evolution — researcher inference · strong. Forecast operating income and FCF only after stabilization, apply cash to dated net-debt targets, stage shareholder payouts, and diversify funding under stress. The growth thesis survives, but near-term allocation becomes subordinate to solvency; modern use requires cash, refinancing, and payout gates that cannot be replaced by EBITDA or guidance. printed pp. 17–20; PDF pp. 19–22

16. Annual Report 2010 — To Shareholders and Investors (2010)

• Context and argument — contemporaneous record · strong. Fiscal 2010 operating income of ¥465.8 billion, FCF of ¥390.8 billion, and a ¥438.4 billion reduction in net interest-bearing debt support Son's claim that mobile growth and stabilization can coexist. The medium-term plan joins leadership in mobile Internet and Asian Internet with at least ¥1 trillion of aggregate FCF and the previously stated debt targets. printed pp. 15, 18; PDF pp. 17, 20
• Reasoning, evidence, and assumptions — investor-stated · strong. Subscriber additions, data ARPU, iPhone/PhotoVision, Alibaba Japan/Taobao, Ustream, and network-capacity investment illustrate a mechanism: create new mobile-data markets domestically, then scale services through partners and infrastructure. These are operating measures and strategic examples, not ownership-level returns from Asian investees. printed pp. 16–21; PDF pp. 18–23
• Rules and evolution — researcher inference · strong. Put growth initiatives inside explicit cash, debt, liquidity, and capacity envelopes; assign owners to partner and infrastructure trade-offs. This annual message supplies near-term controls distinct from the same year's 300-year vision, and prevents the long-range mission from being mistaken for a financing plan. printed pp. 15–23; PDF pp. 17–25

17. SoftBank Next 30-Year Vision (2010)

• Context, access, and central argument — investor-stated · strong. In a fully accessible 133-page official deck, with a related announcement treated as a second route rather than a second work, Son frames the next 30 years as the first step toward 300-year corporate DNA. The invariant purpose is human happiness through the Information Revolution; technologies and business models are replaceable means. deck PDF pp. 1, 5, 30–32, 125–127 announcement, paragraphs 7–9
• Reasoning and assumptions — investor-stated · provisional. Charts extrapolate computation, transistors, networks, storage, automation, and longevity toward machines beyond human-brain component counts. The forecasts are directional and method-light; they supply a search horizon, not a probability-weighted underwriting case. deck PDF pp. 40–75
• Organization, ownership, and succession — investor-stated · strong. The strategic-synergy group should use multiple brands, 20%–40% stakes, category-leading partners, decentralized autonomy, and a web of as many as 5,000 companies. SoftBank Academia is meant to train successors before Son's planned handoff in his sixties. deck PDF pp. 89–123
• Rule and contradiction — researcher inference · strong. Keep the mission fixed while switching technologies, use meaningful minority ownership rather than permanent operating control, and design succession early. The failure conditions are equally important: forecasts may not produce welfare or returns, minority stakes may lack governance, breadth may destroy accountability, and the announced succession timetable later slips. deck PDF pp. 89–123

18. Annual Report 2011 — Philosophy, Vision, and Strategy for Sustained Growth (2011)

• Context and argument — investor-stated · strong. Son defines a hierarchy—philosophy gives purpose, vision gives destination and time, and strategy follows—and says about 20,000 employees contributed responses to the 2010 vision. The human-benefit objective remains above any one information technology or business model. printed pp. 8–13; PDF pp. 10–15
• Organization and evidence — documented behavior · strong. The message repeats the 20%–40% strategic-synergy design and describes SoftBank Academia's candidate funnel, presentations, and evaluations. After the Great East Japan Earthquake, it also records network restoration and a planned ¥1 trillion investment, roughly 80% mobile, alongside debt reduction. printed pp. 14–19; PDF pp. 16–21
• Rules, implications, and evolution — researcher inference · moderate. Begin with human benefit, choose the era's technology, distribute authority among aligned category leaders, train successors through real decisions, and fund resilience when use and reliability require it. Unlike the vision deck, this is an implementation update; yet a shared philosophy does not substitute for minority protections or prove authority has transferred. printed pp. 8–19; PDF pp. 10–21

19. Masayoshi Son off on new quest to raise green profile — Asahi Q&A (2011)

• Context, attribution, and argument — investor-stated · strong. In an edited post-Fukushima Q&A preserved by the Japan Disasters Digital Archive, Son says the disaster widened his obligations beyond the information business. He advocates gradual replacement of nuclear and thermal generation with locally adapted renewables, while defining SoftBank as a temporary catalyst rather than a permanent energy operator. opening answer and “SON TO SERVE AS CATALYST”
• Reasoning, evidence, and access — investor-stated · moderate. Human error, involuntary local exposure, multigenerational waste, and accident compensation form the tail-risk case. Land availability, cost crossover, and 20%–30% growth are his forecasts rather than verified base rates. The archive stores 2011-09-01T00:00:00Z while its U.S. Eastern interface renders 2011-08-31; no original dateline or raw audio is separately visible, so speaker labels govern attribution. risk, policy, and archive-metadata fields
• Rule and operating implications — researcher inference · moderate. Intervene outside the core only after a severe shock exposes an unpriced societal risk; define a catalytic vehicle, make follower economics viable, ring-fence proceeds, and state the handoff condition. A modern IC would need the entity, capital cap, policy dependency, unit economics, and exit trigger, otherwise “catalyst” can conceal open-ended subsidy or mission drift. “SON TO SERVE AS CATALYST”
• Biography and contradiction — retrospective witness · provisional. Son separates a reported ¥10 billion personal donation from corporate restoration and network spending, and gives a contested retrospective account of the Aozora/Nippon Credit Bank exit. The work expands the 2010 mission while insisting the information revolution remains SoftBank's core; absent vehicle and outcome evidence, the boundary cannot be tested. opening setup and “SON'S ENTHUSIASM”

20. Goal-Oriented Management (2012)

• Context and argument — investor-stated · strong. Son attributes record fiscal 2011 results to anticipating mobile Internet and executing smartphone strategy early. “Goal-oriented management” means selecting an ambitious measurable end, reasoning backward to required actions, and executing rather than extrapolating current operations. printed p. 010; PDF p. 12
• Evidence and assumptions — contemporaneous record · strong. The target is ¥1 trillion of consolidated operating income in fiscal 2016, announced in February 2012 and described as certain. The one-page message gives no scenarios, base rates, downside buffer, or revision trigger, and assumes SoftBank's Internet heritage yields distinctive pattern recognition. printed p. 010; PDF p. 12
• Rule and limits — researcher inference · moderate. Backward planning is useful only when a secular shift connects to interim milestones and falsifiers. The work introduces a recurring strength—turning ambition into dated targets—and a recurring risk: confidence in a public target can substitute for probability, price discipline, or abandonment criteria. printed p. 010; PDF p. 12

21. Global No. 1 Is the Starting Point (2013)

• Argument and context — investor-stated · strong. Son ranks financial strength, global leadership, and even 300-year continuity as intermediate conditions beneath a social purpose: improving life through information technology and earning gratitude. The hierarchy is explicit; the terminal stakeholder outcome has no measurement system. printed p. 007; PDF p. 9
• Evidence and boundary — contemporaneous record · strong. The message was published in the month Sprint closed but does not state Sprint's price, financing, baseline, or turnaround case. It is contemporaneous ambition, not the acquisition memo. printed p. 007; PDF p. 9
• Rule and contradiction — researcher inference · strong. Scale belongs in a mission hierarchy only when financial and stakeholder outcomes are measurable. The hidden assumption—that corporate scale and social benefit align—has no counterexample here; modern misuse is using “global No. 1” or centuries-long intent to excuse value-destructive expansion. printed p. 007; PDF p. 9

22. We Can Change the World (2014)

• Context and argument — investor-stated · strong. Son divides SoftBank's history into a Japan-changing first 33 years and a global second stage beginning with Sprint. The transaction's stated mechanism is operating transfer: apply SoftBank's experience, technology, and commitment to build a leading U.S. mobile Internet service. printed p. 003; PDF p. 5
• Reasoning and assumptions — researcher inference · strong. The case extrapolates from Japan without disclosing the U.S. competitive baseline, spectrum, regulation, financing, timetable, capital need, or falsifier. Mission language from 2010 describes the hoped-for purpose, not economic or consumer proof. printed p. 003; PDF p. 5
• Rule and operational implication — researcher inference · strong. Transfer a home-market playbook only after explicitly testing jurisdiction differences, responsible operators, milestones, capital reserves, and stop rules. The work marks a shift from partnership-led internationalization to a control acquisition in a foreign regulated market. printed p. 003; PDF p. 5

23. Transformation into “SoftBank 2.0” (2015)

• Argument and diagnosis — investor-stated · strong. Son wants SoftBank to become a genuinely global, centuries-lasting company. Technology businesses, he argues, often have roughly 30-year lives because technologies, models, and founder dependence decay; a federation that repeatedly supports disruptive entrepreneurs is the proposed institutional answer. printed p. 005; PDF p. 7
• Evidence and assumptions — contemporaneous record · strong. Domestic telecom FCF, Alibaba's listing, Sprint stabilization, and growth at Snapdeal, Ola, and GrabTaxi are selected examples for funding a distributed Internet portfolio. Son argues local entrepreneurs outperform exported Japanese models, but leaves the autonomy/intervention balance undefined. printed pp. 006–007; PDF pp. 8–9
• People, succession, and implications — investor-stated · strong. Nikesh Arora is named the management partner who can help build the global company, based on operating experience and networks rather than disclosed decision-right tests. The work deepens local-partner sourcing and institutional succession, but the latter reverses in 2016; modern use requires authority, incentives, portfolio limits, and a key-person plan, not credentials alone. printed p. 007; PDF p. 9

24. Finding Opportunities in Adversity (2016)

• Argument and context — investor-stated · strong. Son uses Yahoo! BB's post-bust losses and later platform role to argue that adversity can offer the best moment to advance. He applies the analogy to Sprint while conceding that its turnaround is harder than expected. printed pp. 009–010; PDF pp. 11–12
• Reasoning, evidence, and attribution — investor-stated · strong. Direct work with Sprint engineers, network improvement, churn, procurement, finance, and transfer of Japanese know-how are the claimed mechanism. Comparing Sprint with Vodafone K.K. is a one-case analogy that does not control for spectrum, competition, leverage, regulation, or capital. printed pp. 009–010; PDF pp. 11–12
• Portfolio and succession evidence — contemporaneous record · strong. Growth holdings such as Snapdeal, Ola, Grab, Coupang, and SoFi sit beside sales of Alibaba, GungHo, and Supercell, but no security-level rule joins the lists. Son also cancels the Arora succession, citing another five-to-ten-year mission need rather than a disclosed board process. printed pp. 010–011; PDF pp. 12–13
• Rule and failure mode — researcher inference · strong. Adversity is investable only with a funded recovery plan, accountable operators, new evidence, milestones, and a refreshed downside. Without those, analogy and persistence can launder sunk-cost escalation into contrarian discipline. printed pp. 009–011; PDF pp. 11–13

25. Leading the Information Revolution to the Next Paradigm Shift (2016)

• Context and argument — investor-stated · strong. Signed the day after Arm's acquisition completed, Son presents Arm as a shared-mission platform: SoftBank contributes connectivity and Internet assets; Arm supplies pervasive, energy-efficient computing architecture. paragraphs 1–3
• Reasoning and operating plan — investor-stated · strong. Computing and connectivity converge, IoT becomes the next platform, Arm is a core layer, and SoftBank capital should accelerate products and infrastructure. Son promises to preserve senior leadership, customer neutrality, efficient design, partner economics, and engineering talent while expanding R&D. Execution and resulting operating outcomes remain attributable to Arm's leadership and engineers as well as SBG sponsorship; the statement provides no budget, milestone, price test, or return hurdle. paragraphs 2–7
• Rule, contradiction, and implications — researcher inference · strong. A foundational-IP acquisition should protect ecosystem neutrality and customer trust while accelerating R&D. The unresolved tension is whether ownership can increase speed without compromising neutrality; modern underwriting requires retention, roadmap, governance, financing, regulatory, and return gates rather than secular-platform conviction alone. paragraphs 5–7

26. On Our Journey to Make the World a Better Place (2017)

• Context and argument — investor-stated · strong. Achieving the ¥1 trillion operating-income target is offered as evidence that implausible goals can be met, then subordinated to an AI/IoT vision in which data, chips, and connectivity become the next Information Revolution stack. Arm and OneWeb are presented as complementary infrastructure. printed pp. 012–014; PDF pp. 14–16
• Arm reasoning and assumptions — investor-stated · strong. Son calls Arm a hidden gem and predicts it could become his most important investment, citing smartphone share and chip shipments. The thesis is contemporaneous and direct, but omits valuation, cash flows, rival architectures, and customer switching. printed p. 014; PDF p. 16
• Vehicle and team boundary — contemporaneous record · strong. More than $93 billion at SVF1 first close is committed capital, not deployment; SBG's $28 billion includes about $8.2 billion of Arm shares contributed in kind. Outside LPs are named, Rajeev Misra is assigned fund responsibility, and Simon Segars, Marcelo Claure, and Yasir Al-Rumayyan have distinct roles. printed pp. 014–015 and Note 48(2)(b), printed p. 179; PDF pp. 16–17, 181
• Rule and implications — researcher inference · strong. Map every strategic fund's mandate, GP/LP economics, capital-call timing, conflicts, governance, and liquidity before treating scale as access advantage. The work inaugurates the Vision Fund era but supplies no evidence that a larger pool preserves price discipline or converts early marks into cash. printed pp. 014–015; PDF pp. 16–17

27. Cluster of No.1 Strategy for 300 Years of Continuous Growth (2018)

• Argument and organization — investor-stated · strong. Son codifies a group of category leaders held typically at 20%–30%, often as largest shareholder, with founder autonomy and easier entry/exit than majority control. Biological evolution is the analogy for a self-propagating, self-renewing institution, not an empirical longevity base rate. “Creating a unique synergy group structure” and “Driving the evolution of the Group”
• Portfolio, people, and evidence — contemporaneous record · strong. The domestic telecom listing and Sprint–T-Mobile plan support a holding-company transition; Son credits Ken Miyauchi and Marcelo Claure for operating roles and redirects his own time toward Vision Fund investing. The fund is said to invest roughly ¥100 billion on average in established or near-leading firms, with selected growth examples but no hit-rate denominator. “Transforming from telecommunications business” and “A growing list”
• Valuation and contradiction — researcher inference · strong. The claimed sum-of-parts discount mixes public marks, acquisition cost, merger-premised value, private fair values, and manager economics of unequal observability. Minority influence is useful only when founder incentives, governance rights, price, and liquidity align; category rank and a 20%–30% stake are not substitutes for enforceable influence or unit economics. “Our share price of ¥8,070 is undervalued,” footnotes 4–5

28. Aiming to Orchestrate the AI Revolution (2019)

• Context and argument — investor-stated · strong. Before the WeWork IPO failure became public, Son compares AI with the Industrial Revolution and casts SoftBank as conductor while independent entrepreneurs remain performers. The metaphor promises value through coordination without operating every company. printed pp. 002–004; PDF pp. 4–6
• Evidence and assumptions — contemporaneous record · strong. Son calls the Vision Fund an unconventional venture investor making large checks into actual or likely unicorns and says he is extremely satisfied after roughly two years. The message does not separate cash realization, unrealized marks, or third-party interests, and analogy to historical industrialists does not validate the portfolio. printed pp. 003–004; PDF pp. 5–6
• Rules, succession, and limits — researcher inference · strong. Cross-portfolio coordination should have named owners, counterparties, consent, and measurable incremental value. The orchestra's changing conductor is a succession metaphor, not governance design; past target achievement and 16% annualized market-cap growth do not prove that public commitments cause results or calibrate the next target. printed p. 004; PDF p. 6

29. Leading the Digital Transformation While Increasing Shareholder Value (2020)

• Context and argument — investor-stated · strong. After the Vision Fund/WeWork break and during COVID-19, Son argues that crises accelerate infrastructure and behavior change. Usage at selected portfolio companies supports digital adoption, but not profitability or a portfolio-wide hit rate. printed pp. 002–003; PDF pp. 6–7
• Scorecard and capital response — contemporaneous record · strong. Son prioritizes shareholder value over operating income while acknowledging a ¥1.4 trillion operating loss and ¥961.6 billion net loss. The plan monetizes up to ¥4.5 trillion of assets, repurchases up to ¥2.0 trillion of shares, lowers debt, and builds cash; it does not quantify tax, foregone upside, or persistence of the discount. printed pp. 003–004; PDF pp. 7–8
• Sprint, governance, and attribution — investor-stated · moderate. Son concedes Sprint had dragged earnings, then calls the combined investment adequate after the T-Mobile merger without a complete cash-flow ledger. More outside directors and a nominating and compensation committee are recorded, but not tied to a named decision-process change after failure. printed pp. 004–005; PDF pp. 8–9
• Board-role entity control — contemporaneous record · strong. The same pages distinguish reciprocal board departures—Jack Ma left SBG’s board, while Son left Alibaba’s board on June 25, 2020—but later describe both as stepping down as Alibaba directors; that later phrase is internally ambiguous and must not erase the entity distinction. printed pp. 004–005; PDF pp. 8–9
• Rule and evolution — researcher inference · strong. A shock-accelerated adoption thesis needs durable retention, margin conversion, and liquidity preservation. The piece is the essential counterweight to 2019: ecosystem rhetoric survives, but asset sales and loss recognition reveal that marks, temporary engagement, and strategic adjacency are not realized returns. printed pp. 002–004; PDF pp. 6–8

30. Creating the Future with AI Entrepreneurs as a “Vision Capitalist” (2021)

• Context and argument — investor-stated · strong. After the 2020 Vision Fund/WeWork crisis and market recovery, Son rejects “investment company” as SBG's identity and describes a long-duration capital provider for the Information Revolution. An industrial-history analogy assigns invention to entrepreneurs and risk capital to SBG, but does not prove the analogy or that greater risk produces greater return. “SoftBank Group = A capital provider” and “The goal of a capital provider”
• Evidence, scorecard, and denominator — contemporaneous record · strong. The cited 264 investments span SVF1, SVF2, and the Latin America Fund and include 44 investment-committee-approved projects not yet closed. Reported 43% IRRs use management-defined scopes, combine SBG LP and manager economics, and include unrealized NAV; neither is an audited Son-personal cash return. “SoftBank Vision Funds have been driving NAV growth,” footnotes 1–3
• Rules and operations — investor-stated · strong. Select 10–30-year paradigm shifts, tolerate absent current profit when the eventual market can expand, use NAV rather than current income as SBG's principal scorecard, and weigh repurchases against financing and investment opportunities. The work supplies no valuation ceiling, loss limit, reserve rule, governance test, or mechanism to distinguish a durable shift from fashion. capital-provider section and share-buyback Q&A
• Biography, attribution, and contradiction — researcher inference · strong. Approvals belong to named fund investment committees and results to vehicles, teams, and LP structures, not Son alone. His intention to identify a successor in his sixties is later deferred, while 2022 adds explicit risk limits and 2023 returns to offense—evidence that long duration never makes financing conditions secondary. footnote 1 and succession Q&A

31. Investing in Technology Evolution While Adhering to Our LTV Commitment (2022)

• Context and central argument — investor-stated · strong. During the technology correction, Son separates thesis conviction from solvency: Vision Funds and Arm remain the growth engines, but normal-times LTV should remain below 25%, 35% is an emergency ceiling, and cash should cover at least two years of bond redemptions. “Investing in technology evolution” and “Proper LTV control”
• Reasoning and denominator — investor-stated · moderate. Son links SBG NAV and share price to Internet traffic and cites 475 portfolio companies, but the count mixes completed, exited, post-year-end, and approved-not-closed investments; Arm share figures are Arm estimates. Industry growth does not prove security selection, value capture, or the absence of leverage and dilution risk. “SBG's share price = NAV = Internet traffic,” footnotes 2–7
• Observed risk response — documented behavior · strong. As falling marks lifted LTV to 21.6% at December 2021, SBG tightened selection, slowed investment, sold assets, ended the year at 20.4% LTV, and held ¥2.9 trillion against ¥1.3 trillion of two-year bond redemptions. These are holding-company actions, not proof of improved security-level underwriting. “Proper LTV control,” footnotes 9–10
• Rule and evolution — researcher inference · strong. Precommit liquidity floors and deployment brakes before a downturn, but stress correlated marks, FX, asset-backed finance, capital calls, and illiquidity. The later offensive and concentrated AI plans test whether the 2022 limits govern position size and project finance as well as reported holding-company leverage. 2022 message

32. Going on the Offensive to Lead the Information Revolution (2023)

• Context and argument — investor-stated · strong. Son says generative AI and future superintelligence restore the reason to take offensive risk after fiscal 2022's defense. Arm and selected Vision Fund companies should sit around the platform shift; more than ¥5 trillion of cash and 11% LTV at March 31, 2023 are the corporate gate for changing posture. “An architect supporting the future of humanity” and “Empowering happiness,” footnotes 1–2
• Reasoning, evidence, and assumptions — investor-stated · moderate. Computing progresses from calculation to storage, search, inference, and creation; AI is predicted to exceed collective human intelligence within a decade, while 10,000-times superintelligence is placed “a few decades” away. Nearly 500 portfolio companies and a 52,000-proposal employee competition supply option breadth and experimentation, not validated products, returns, or forecast base rates. “Beyond the boundaries of wisdom,” “AI continues to evolve,” and “Arm has benefited”
• Biography and team boundary — retrospective witness · moderate. Son dates a confidence crisis to autumn 2022 and reports roughly 630 ideas in eight months, most expected to fail. Eleven employee finalists were reviewed with external director Yutaka Matsuo; this establishes Son's final-review role but not sole ideation or implementation credit. “An architect supporting the future of humanity” and employee-competition section
• Rules and limits — researcher inference · moderate. Build excess liquidity before offense, use high experiment throughput, combine infrastructure and application exposure, and delegate quarterly operations. The method requires technical milestones, governance, valuation, and downside cases; idea counts and predicted intelligence multiples are not investability evidence. 2023 CEO message

33. Realizing Artificial Super Intelligence for the Evolution of Humanity (2024)

• Argument and strategic evolution — investor-stated · strong. Son joins maximizing NAV through “evolution and multiplication” with an ASI mission. His six-stage history moves from software/Yahoo through broadband/mobile, Sprint, Alibaba, Vision Funds, and Arm-led AI; it is a useful management taxonomy but compresses overlaps, losses, financing, and team execution. “Maximizing shareholder value” and “Our mission”
• Arm evidence and return boundary — contemporaneous record · moderate. The company-reported ¥24.6 trillion return and 10x MOIC include IPO proceeds, retained shares at market, in-kind dividends, and the non-refundable NVIDIA advance, before tax and with expenses/intra-group effects excluded. Doubling engineers is linked to market expansion, but the resulting marked value cannot isolate management, engineering, market, customer, financing, or Son effects and is not a cash multiple. Figure 2 and “Maximizing shareholder value”
• Reasoning and operating implications — investor-stated · moderate. A network of AGI systems is expected to create ASI and move into the physical world; Arm, AI chips, data centers, power, and robotics become complementary layers. The architecture is more operational than 2023, but the causal and value-capture case remains forecast-led. “Our mission,” “Arm provides,” and “Driving innovation”
• Forecast acceleration — researcher inference · strong. The 10,000-times forecast accelerates from “a few decades” in 2023 to “about ten years” in 2024 without a disclosed forecasting bridge. 2023, “AI continues to evolve towards superintelligence” 2024, “Our mission: Evolution of humanity”
• Rule and contradiction — researcher inference · strong. Map each layer's bottleneck, owner, capital budget, dependency, milestone, and failure condition; decide which requires control and which can be partnered. High public-asset marks cannot both prove the broad forecast and automatically fund unrelated layers, and a normal-market LTV ceiling does not answer project-level concentration. Figure 2 and “Driving innovation”

34. Becoming the World's No. 1 ASI Platform Provider (2025)

• Argument and biography — investor-stated · strong. The ASI mission becomes a competitive objective: combine Arm and OpenAI with infrastructure and enterprise deployment to lead the platform. Son retrospectively traces this to a 1975 microchip image; that is first-person memory, not contemporaneous proof that ASI was a formulated 50-year investment thesis. “At the forefront of ASI,” “It all began with a photo,” and “The vision that shaped SoftBank”
• Economics and transaction boundary — contemporaneous record · strong. The up-to-$32.2 billion OpenAI exposure comprises SVF2's $2.2 billion initial investment and an effective follow-on of up to $30 billion subject to recapitalization conditions; it could instead be $10 billion. SBG's own 5%-of-GDP value forecast and near-50% margin suggestion have no probability tree, comparable base rate, capital bridge, or competitive analysis. “At the forefront of ASI,” footnotes 2–3
• Operating evidence and team — documented behavior · moderate. Arm chip shipments, completed Graphcore and expected Ampere acquisitions, and a plan to deploy Cristal across 2,500 internal systems illustrate vertical expansion and an internal design-partner loop. They do not yet show integration, productivity, security, retention, external revenue, or sole Son execution; OpenAI, Arm, acquired teams, partners, and SBG employees all matter. “At the forefront of ASI” and “A historical turning point”
• Rule and contradiction — researcher inference · strong. Seek a control point, combine compute with models, use the group as first customer, and treat conditional commitments as scenario trees. “Going all in” intensifies the conflict with 2022 discipline; winner-take-most expectations cannot justify any price or prove that partner access equals control, and the humane-ASI aspiration lacks evaluation, liability, and shutdown governance. 2025 CEO message

35. Leading ASI's Four Key Areas to Become the No. 1 Platform Provider (2026)

• Argument and evolution — investor-stated · strong. The 2025 two-engine story becomes four “corners”: frontier models, semiconductors, power/data-center infrastructure, and robots. OpenAI, Arm and a new chip business, infrastructure projects, and robotics assets form the intended system, while Son sets a ¥1 quadrillion NAV objective for 2042. “Securing the four corners” and “Aiming for NAV”
• Reasoning and project boundary — investor-stated · moderate. The Othello analogy argues that controlling all four layers creates positional advantage, but does not prove control, necessary ownership, or equal margins across software/IP and physical infrastructure. Ohio and France capacity plans and the announced ABB Robotics acquisition were future or unclosed at the message date, not operating or return evidence. “Securing the four corners,” footnotes 2–4
• Target and scorecard — investor-stated · moderate. Son compares approximately ¥3 trillion of 2010 NAV with ¥74 trillion on June 23, 2026 and extrapolates to a target about 13.5 times the latter. No annual cash-flow, dilution, leverage, or scenario bridge supports the next 16 years; the number is ambition, not forecast guidance. “Aiming for NAV of ¥1 quadrillion in 2042”
• Biography, operations, and contradiction — researcher inference · strong. Son explicitly extends his university-era life plan into his seventies, reversing the 2010/2021 succession timetable. The four-layer thesis requires a capital/dependency ledger by control, duration, technical milestone, demand contract, energy, unit economics, permit, counterparty, and stop condition; survivorship analogy to Internet winners cannot establish present entry-price returns or resolve the tension between NAV maximization and human-centered purpose. “Updating my 50-year life plan,” “The internet was called a bubble,” and “Aiming for NAV”

Cross-source synthesis

The durable architecture: mission, regime, bottleneck, vehicle

[researcher inference · strong] The corpus is not a stable list of sectors; it is a repeated four-step architecture. First, Son defines an invariant mission—human benefit through the Information Revolution. Second, he forecasts the next technological regime. Third, he identifies a bottleneck or aggregation point. Fourth, he changes the vehicle and degree of control to pursue it. The sequence runs from Internet traffic and portals in 1998–1999, through broadband infrastructure in 2001, mobile distribution in 2006, Arm's architecture in 2016, the Vision Fund's capital scale in 2017, and the model/compute/infrastructure/robotics stack in 2024–2026. FY1998, printed pp. 4–5 FY2001, printed pp. 2–5 2016 Arm statement, paragraphs 2–7 2026 CEO message, “Securing the four corners”

[researcher inference · strong] The enduring insight is technology flexibility inside mission continuity, not a prediction that every named technology wins. Son explicitly says SoftBank should not bind itself to one technology or business model, and the historical record documents large changes in operating form. Next 30-Year Vision, PDF pp. 89–99 The critical boundary is that a correct regime forecast does not establish value capture, entry-price return, governance, or survival: 2001 adoption coexisted with heavy broadband losses, 2019 AI optimism preceded the Vision Fund crisis, and current ASI layers have different capital intensity and transaction states. FY2003, printed pp. 2–3 FY2019 CEO message, printed pp. 003–004 2026 CEO message, footnotes 2–4

Foundational beliefs versus occasional positions

From rhetoric to documented behavior

• Bottleneck commitment — documented behavior · strong. SoftBank did not merely describe broadband; it funded a network through continuing losses, security sales, and debt reduction, then reported improving segment economics. That supports willingness to act on the thesis, not the optimality of price, financing, or ultimate return. FY2002–2004 messages FY2004, printed pp. 2–3
• Operating transfer — documented behavior · strong. The Vodafone turnaround record names base-station expansion, product breadth, installment sales, and pricing changes. Sprint's writings name engineering, network, procurement, and financing intervention but also concede greater difficulty; the comparison supports a repeatable intent, not a repeatable outcome. FY2007, printed pp. 16–20 FY2016, printed pp. 009–010
• Balance-sheet reflex — documented behavior · strong. Son repeatedly sells assets or slows investment when financing risk rises. The practice is real, but may force opportunity-costly sales and does not prove the original underwriting; it should be evaluated separately from the technology thesis. FY2009, printed pp. 17–20 FY2020, printed pp. 003–004 2022 CEO message
• Network synergy — researcher inference · provisional. Alliance density, strategic-synergy groups, clusters, and orchestras recur across decades, but the works seldom measure incremental revenue, cost savings, retention, or return caused by the network. Common investors, branding, and marks are not synergy. Forbes, alliance passages FY2019, printed pp. 003–004
• Failure learning — researcher inference · provisional. The corpus acknowledges dot-com losses, failed early STV positions, Sprint difficulty, WeWork/Vision Fund embarrassment, and portfolio companies that did not succeed. It discloses clear financing responses, but not a complete security-level before/after investment-committee rulebook. Forbes, STV failures 2021 CEO message, Vision Fund section 2022 CEO message

Actionable implications by investment function

• Sourcing — researcher inference · strong. Start with a regime map, then search for the aggregation point, technical bottleneck, or distribution owner. Use local operators and specialist teams rather than assuming the founder personally originated every opportunity. FY1999, printed pp. 2–5 FY2015, printed p. 007
• Picking — researcher inference · moderate. Category leadership, control points, increasing returns, and founder autonomy are recurring screens. The writings do not provide a stable founder test, valuation method, unit-economics threshold, or portfolio-company governance red flags; those must not be invented as Son rules. 2018 CEO message 2025 CEO message
• Pricing — researcher inference · insufficient public record. Aside from sum-of-parts arguments and management scorecards, the written corpus contains little repeatable entry-price discipline. A secular winner can still be a poor investment at the wrong price; this is one of the most important missing components. 2018 CEO message, valuation section 2024 CEO message, Figure 2
• Ownership and governance — researcher inference · strong. Choose deliberately among small test stakes, meaningful minority positions, largest-shareholder influence, and control acquisitions. Match rights to the intended intervention; autonomy cannot be assumed to coexist with influence, and control does not guarantee transfer. Forbes, Yahoo/E-Trade passages 2018 CEO message 2016 Arm statement
• Portfolio construction — researcher inference · strong. Early breadth followed by top-ups, a thousands-company network, large late-stage checks, and current full-stack concentration are materially different architectures. A usable policy must name the vehicle, denominator, reserves, correlation, position limit, and role of the dominant asset rather than call all of them “ecosystem.” Forbes, STV passages 2017 annual report, Note 48 2026 CEO message
• Exits and scorecards — researcher inference · strong. Asset sales, listings, M&A, and buybacks are recurring tools, but NAV and public marks mix realization states. Track gross and net cash, remaining value, tax, leverage, dilution, third-party interests, and foregone upside separately. FY2020, printed pp. 003–004 2021 CEO message, footnotes 2–3 2024 CEO message, Figure 2

Attribution boundary across the corpus

[researcher inference · strong] Son supplies mission, public targets, strategic analogies, and at times direct operating involvement. The record also names Gary Rieschel at STV; telecom leaders Ken Miyauchi and Marcelo Claure; Arm's Simon Segars and engineers; Vision Fund leader Rajeev Misra and fund investment committees; external director Yutaka Matsuo; sovereign and corporate LPs; and portfolio founders and teams. Forbes, STV people passages FY2017, printed pp. 014–015 2018 CEO message 2023 CEO message The writings do not expose complete sourcing, diligence, vote, board, follow-on, or exit attribution for most securities. Any thesis about Son's individual investment skill must therefore remain distinct from SBG corporate capital, managed-fund economics, LP capital, team execution, and partner-company performance.

Intellectual chronology, evolution, and contradictions

Dated evolution

Chronology alone does not prove influence. The 2010 vision predates the 2016 Arm thesis and resembles later ASI language, but it does not specify OpenAI, current counterparties, or the four-layer capital plan. Son's 2025 memory of a 1975 microchip image is evidence of how he narrates continuity, not contemporaneous proof that the current thesis existed then. [retrospective witness · moderate] 2010 deck, PDF pp. 40–75 2025 CEO message, “It all began with a photo”

Contradiction ledger

Decision playbook

Historical rules reconstructed from the corpus

1. Name the regime before the company — researcher inference · strong. Start with the structural shift—digital services, broadband, mobile Internet, AI—then specify the adoption evidence that could refute it. Ask: what changed in cost, capability, regulation, or behavior; which claim is forecast rather than observed; and what outcome would prove the regime is real but the investment wrong? FY1998 2023
2. Find the bottleneck or aggregation point — researcher inference · strong. Identify where information, distribution, compute, or infrastructure constrains the new system. Ask who controls it, whether control is durable, what alternatives exist, and whether value accrues there after capital cost. Failure mode: a technologically central layer with commoditized economics. FY1999 2016 Arm statement
3. Stage conviction against evidence — researcher inference · moderate. The early Forbes record supports a test-and-top-up pattern, but later giant checks make it unsafe to describe staging as universal Son doctrine. Define first-check size, milestone, reserve, valuation update, and kill criterion before investing. Failure mode: top-ups driven by price momentum or sunk cost. Forbes, Yahoo/STV passages 2018 CEO message
4. Choose ownership from the work required — researcher inference · strong. Use minority partnership when local founder autonomy and access matter; use control only when operating transfer or infrastructure ownership is necessary and fundable. Ask which rights, board seats, information, vetoes, or operating authority the thesis requires. Failure mode: influence without rights or control without transferable competence. 2010 deck FY2007
5. Write the operating-transfer plan before acquisition — researcher inference · strong. Name inherited weaknesses, responsible leaders, interventions, baselines, milestones, financing, and jurisdiction differences. Vodafone shows an observable intervention sequence; Sprint shows why one precedent is not a base rate. FY2007 FY2016
6. Protect the right to wait — documented behavior · strong. Put concentration inside liquidity, debt, refinancing, and deployment gates. Ask what a correlated mark decline, FX move, capital call, covenant shock, or closed exit window does to the plan. Failure mode: selling the best liquid assets to fund the most illiquid thesis. FY2009 2022
7. Separate operating, mark, and cash scorecards — researcher inference · strong. Report users and margins, public/private marks and NAV, and realized net cash as distinct ledgers with taxes, fees, leverage, dilution, and third-party interests. Failure mode: using a rising mark or commitment as proof of return. 2021, footnotes 1–3 2024, Figure 2
8. Make “synergy” earn a number — researcher inference · strong. Name the two counterparties, owner, customer consent, transfer, incremental revenue or savings, time, and counterfactual. Failure mode: counting common capital, introductions, or brand adjacency as value. Forbes, alliance passages FY2019
9. Precommit the failure response — researcher inference · strong. State warning signals, follow-on cap, governance escalation, asset-sale priority, and stop/exit rule before adversity. Failure mode: interpreting all drawdowns as opportunity and all persistence as long-termism. FY2004 FY2016 2020
10. Institutionalize credit and succession — researcher inference · strong. Record who sourced, diligenced, voted, operated, governed, funded, and exited; define authority transfer, not merely candidate training. Failure mode: a supposedly self-evolving institution whose strategy and public confidence remain founder-dependent. FY2017, printed pp. 014–015 2026, “Updating my 50-year life plan”

Modern VC translation (as of 2026-08-22)

Every row below is researcher inference; support grades describe the historical evidence, not a claim that Son stated the translation.

Best writings and unresolved record

The five best places to start

1. Master of the Internet (1999) — contemporaneous record · strong for learning. It is the rare early source that combines actual cases, check sizes, a mixed denominator, a named fund leader, team modeling, autonomy, follow-ons, failures, and an explicit skeptical alternative. Its edited-reporting boundary is clear and useful.
2. SoftBank Next 30-Year Vision (2010), deck PDF — investor-stated · strong. This is the operating constitution: stable mission, flexible technology, minority partnership, decentralized organization, extreme scale, values, and succession. It is also the cleanest place to see where vision outruns governance and forecast method.
3. Annual Report 2009 — To Shareholders and Investors — investor-stated · strong. It turns ambition into FCF, debt, liquidity, and dividend gates after a crisis. Those hard constraints are essential counterweight to the more famous visionary texts.
4. On Our Journey to Make the World a Better Place (2017) — contemporaneous record · strong. It freezes the Arm and SVF1 thesis near inception, exposes commitment versus deployment, names external LPs and senior operators, and makes the transition from operating company to scaled capital vehicle auditable.
5. Investing in Technology Evolution While Adhering to Our LTV Commitment (2022) — investor-stated · strong. It is Son's clearest public survival policy and the best test for every later “offense” or “all in” declaration: conviction is bounded by leverage, liquidity, and deployment speed.

Unresolved record

• 1992 HBR body — insufficient public record. A lawful full-text copy with page numbers could materially change the earliest intellectual baseline; title metadata cannot. HBR page
• Authorship apparatus — unresolved. Signed CEO messages support Son attribution for first-person doctrine, but the drafting roles behind charts, footnotes, translations, calculations, and selected examples are not public. Corporate report apparatus must remain SBG disclosure.
• Security-level cash-return ledger — unresolved and decision-critical. The works do not reconcile contributions, follow-ons, transfers, fees, leverage, tax, dilution, realizations, remaining marks, and outside LP interests across Son's career. Management IRR, NAV, public marks, and individual transaction snapshots cannot substitute. 2021 footnotes 2–3 2024 Figure 2
• Complete team and decision-right attribution — unresolved. Named executives and committees establish that Son did not act alone, but most deals lack public sourcing, diligence, dissent, vote, board, follow-on, and exit records. FY2017, printed pp. 014–015 2021 footnote 1
• Price discipline and kill criteria — insufficient public record. The corpus is rich on paradigm, category, scale, and financing after drawdowns, but thin on entry valuation, expected return, position ceilings, negative diligence, and precommitted stop rules.
• Failure-process change — provisional. Son says embarrassing investments produced lessons, and the company later slowed deployment and strengthened liquidity. The public writings do not identify the failed cases, revised investment-committee tests, or security-level evidence that new rules bound subsequent decisions. 2021 Vision Fund section 2022 risk section
• Renewable catalyst outcome — unresolved. The 2011 Q&A does not establish the legal vehicle, approvals, financing, capital cap, profit ring-fence, handoff, or performance, nor independently resolve the Aozora exit account. Asahi Q&A
• Succession and key-person continuity — unresolved. Training and candidate announcements did not produce the planned handoff; the 2026 extension makes authority after Son more important, not less. 2010 deck, PDF pp. 116–123 2026 life-plan section
• Current ASI rights and milestones — unresolved and potentially decisive. The ownership, governance, waterfalls, funding, permits, customer contracts, integration plans, security controls, and abandonment criteria across OpenAI, Arm, chip acquisitions, power/data-center projects, and robotics are not reconciled in Son's messages. Announced scale must not be treated as closed capacity or realized return. 2025 footnote 2 2026 footnotes 2–4
• Forecast calibration — unresolved. The writings repeatedly use exponential technology, intelligence multiples, GDP share, and NAV targets without disclosed probability distributions or consistent base rates. Future annual reports can test outcomes, but they cannot retroactively supply the ex-ante model that is absent here. 2010 deck, PDF pp. 40–75 2025 footnote 3 2026 NAV section

The useful way to read Son: follow the operating sequence, then audit the missing price

Masayoshi Son's public record is most useful as a map of how he turns a technological forecast into institutional action. He starts with a long secular curve, identifies a platform or category leader, recruits unusually ambitious people, mobilizes capital at a scale meant to alter the competitive field, and then seeks coordination across operating companies, minority holdings, infrastructure, and governments. The sequence is remarkably stable from mobile internet and Sprint through Arm, the Vision Funds, and the 2024–2026 artificial-superintelligence stack. What changes is the instrument: a leveraged operating acquisition becomes a fund ecosystem, then a concentrated group-wide stack of chips, models, data centers, agents, and robots. Sprint SEC transcript 2026 presentation, pp. 27–38 [researcher inference · strong]

That consistency is not evidence of accuracy or effectiveness. The recurring weakness is the missing bridge from technological importance to security-level return: entry price, dilution, financing, governance rights, downside probability, portfolio concentration, and exit rules are often absent. Son's best public admissions—overpaying for WeWork, trusting its entrepreneur too much, and later saying the No. 1-company process could still overpay—show that identifying a real wave or a surviving leader does not close the underwriting case. Forbes, node 10 FY2022 Q1 transcript, A9 The reader should therefore use these appearances for hypothesis generation and decision chronology, then test every economic claim against transaction documents, vehicle-level cash flows, and contemporaneous forecasts. [researcher inference · strong]

Coverage, denominator, and evidence rules

• Underlying works: 46 processed appearances. Priority is 14 ★★★ / 18 ★★ / 14 ★. Access is 11 full / 35 partial; source reliability is 43 A / 3 B. An appearance remains one work when a publisher page, embedded player, deck, transcript, recap, or mirror provides another route to it. The frozen work-level priority mapping is encoded on each dated H3 and is an editorial research priority independent of route-level stars; route priorities measure retrieval importance and do not aggregate to work priority. Mapping context: the fully processed August 2022 composite is ★★★, while the separate partial 2010 TALKS digest stays ★★ despite higher-priority WRITINGS deck/announcement routes.
• Checked source-map routes: sources.md contains 144 checked / 0 open TALKS route rows, split 34 early-AGM + 31 later-AGM + 39 SoftBank World + 40 official, dialogue, interview, and current-era routes. Route priority is 39 ★★★ / 78 ★★ / 27 ★. This is the checked source-map route denominator and does not change the 46-work denominator.
• Note-header URL declarations: the frozen notes contain 75 URL-bearing route declarations in their metadata/header blocks. This is an inventory of URLs surfaced in note headers, not the checked source-map route denominator.
• Broader evidence universe: the corpus contains 152 distinct cited URL strings once timestamp/query variants, contextual controls, comparison-year documents, and supporting releases are included. That figure is deliberately not a route denominator and must not be collapsed into the 75 note-header declarations or the 144 checked source-map rows.
• Speech boundary: deck text, press-release prose, and event descriptions are corporate written evidence, not Son's verbatim speech. Unverified automatic captions support cautious paraphrase only. No local or non-OpenAI audio transcription informed this corpus.
• Attribution boundary: SBG corporate capital, SBIA-managed funds, SVF1, SVF2, regional funds, operating subsidiaries, founders, boards, lenders, LPs, and Son personally are not interchangeable. Commitment, cost, mark, NAV, proceeds, distribution, and realized return are also different measures.
• Interpretive boundary: repeated themes establish persistence of belief. They do not prove forecast truth, causal contribution, or investment effectiveness.

Dated source-by-source guide

1. 2008-06-25 — 28th AGM: governance flexibility meets shareholder constraint — ★

Source note · Access: partial. The official archive preserves the meeting route but no speech, deck, Q&A, or timestamps; the accessible evidence is the linked corporate record. The key action was a proposed preferred non-voting share class intended to widen financing and acquisition options without ordinary-vote dilution, followed by withdrawal after shareholder feedback. That sequence shows a control-preserving capital-design impulse constrained by owners, but it is a board/company action rather than a Son-only investment rule. [contemporaneous record · strong] AGM archive, “VOD, Materials” table May 8 release, “Purpose” and “Reasons” May 20 release, “Reason for cancellation”

Operationally, the appearance begins the record with a useful negative lesson: strategic flexibility is not authorization, and a proposed financing instrument is not deployed capital. No valuation, target, leverage limit, or expected return survives in the accessible meeting record, so later acquisition logic cannot be backfilled into 2008. [researcher inference · strong] AGM archive, “VOD, Materials” table May 8 release, “Reasons” May 20 release, “Reason for cancellation”

2. 2009-06-24 — 29th AGM: infrastructure plus content as a digital flywheel — ★

Source note · Access: partial. The webcast is closed, while the notice, business report, and resolutions survive. SoftBank described itself as a comprehensive digital-information company, joining communications infrastructure with content and services; iPhone 3G momentum supplied the operating example even as FY2008 sales and net income had declined and charges remained material. [contemporaneous record · strong] Notice/business report, PDF pp. 2–3 AGM archive, “VOD, Materials” table

The investable inference is a distribution flywheel: own the access layer, use compelling devices/content to improve demand, and reinvest network economics. Yet the documents do not disclose churn, acquisition cost, handset subsidy payback, or incremental returns. The thesis may explain why a platform matters without establishing whether the owner paid an attractive price. [researcher inference · moderate] Notice/business report, PDF pp. 2–3

3. 2010-06-25 — 30th AGM: an explicit, dated deleveraging regime — ★

Source note · Access: partial. The closed webcast prevents speech attribution; the official notice and governance documents control. After mobile-led improvement, SoftBank set unusually concrete corporate targets: halve net interest-bearing debt in three years, reduce it to zero in six, and generate at least ¥1 trillion of free cash flow over six years. The perimeter and later policy changes matter, so this is a dated balance-sheet program rather than a permanent Son doctrine. [contemporaneous record · strong] Notice/business report, PDF pp. 4–5 AGM archive, “VOD, Materials” table

This is one of the clearest early risk rules in the corpus: translate strategic ambition into a cash-generation and debt path. It later contrasts with Sprint leverage and fund-scale expansion, showing that Son's capital regime changes with opportunity and perceived capacity. [researcher inference · strong] Notice/business report, PDF pp. 4–5 Sprint SEC transcript, lines 1083–1087 and 1408–1416 2017 presentation, pp. 36–40

4. 2010-06-25 — Next 30-Year Vision digest: mission and horizon, not a transcript — ★★

Source note · Access: partial. SoftBank supplies a six-minute digest of a two-hour presentation, an event notice, an announcement, and a 133-page deck; the player has no caption track. Only the verified publisher description and written announcement are used. They define a long-run mission—contribute to human happiness through the IT revolution and become a group needed around the world—without proving which detailed slide claims Son spoke in the edit. [contemporaneous record · strong] Official digest, description Same-day announcement

The decision value is the horizon itself: build an institution able to rotate across technologies rather than optimize a single product cycle. The access limit is substantive; brain-computer, robot, IoT, 300-year, and valuation targets in written routes are not reconstructed as speech. [researcher inference · strong] Official digest, description Same-day announcement

5. 2011-06-24 — 31st AGM: a self-renewing group and a conditional energy initiative — ★

Source note · Access: partial. A 108-page official deck survives, but the webcast is closed. The deck operationalizes the long-horizon vision as a 5,000-company synergy group and an immense market-cap aspiration; these are organizational ambitions, not an achieved portfolio or probability-weighted forecast. After the earthquake, it also advances renewable energy subject to feed-in tariffs, grid access, and governance approval—a useful example of a public mission becoming investable only when institutional conditions align. [contemporaneous record · moderate] Strategy deck, PDF pp. 54–75 AGM archive, “VOD, Materials” table

The implied rule is conditional rather than purely visionary: pair a social need with a policy and distribution mechanism before committing. Missing are project returns, capital requirements, failure thresholds, and evidence that the proposed 5,000-company architecture improved selection or governance. [researcher inference · moderate] Strategy deck, PDF pp. 54–75

6. 2012-06-22 — 32nd AGM: debt capacity is reallocated from survival to growth — ★

Source note · Access: partial. The 137-page deck is available; speech and Q&A are not. After reporting a fall in net interest-bearing debt from roughly ¥2.4 trillion to ¥0.5 trillion, SoftBank shifted from the earlier debt-zero objective toward balancing growth investment, shareholder returns, and debt reduction; the annual dividend rose from ¥5 to ¥40 and an operating-income target of ¥1 trillion was presented. [contemporaneous record · strong] Strategy deck, PDF pp. 31–56 and 115–127

This is an explicit regime change: deleveraging created option value, then management chose to spend some of that capacity. The claimed 960 group companies and 1.4 billion aggregate users are not a deduplicated exposure measure, while hurdle rates, concentration limits, and veto rules remain missing. [researcher inference · strong] Strategy deck, PDF pp. 31–56 and 115–127

7. 2012-10-15 — Sprint announcement: capital plus operating strategy, before the outcome — ★★★

Source note · Access: full. The SEC-filed corrected transcripts preserve both the investor call and Tokyo press conference with line and timestamp locators. Son's ex-ante case was that the U.S. mobile market was large, the shift from voice to mobile internet created a platform transition, and Sprint needed both capital and network/procurement strategy. He acknowledged higher transaction risk but argued that refusing the challenge could create greater strategic risk. [investor-stated · strong] SEC transcript, Son remarks and press conference, lines 701–787 and 1065–1264

The operational playbook is speed after conviction: improve the network rapidly and transfer Japan experience. The missing underwriting bridge is material—entry value, spectrum economics, integration cost, downside scenarios, and exit conditions are not turned into a repeatable return model. This was a leveraged corporate-control transaction, not a Vision Fund bet. [researcher inference · strong] SEC transcript, Son remarks and press conference, lines 701–787 and 1065–1264

8. 2013-06-21 — 33rd AGM: Sprint synergy forecasts before closing — ★

Source note · Access: partial. The meeting occurred after DISH abandoned its competing bid and before Sprint closed; the official deck survives, not the speech. Son's corporate case combined U.S. market scale, common LTE/smartphone technologies, and transfer of the Japan turnaround playbook. Crucially, the presentation gave falsifiable operating targets: average annual operating-expense synergies above $2 billion in FY2014–FY2017, above $3 billion in FY2017, and capital-expenditure reductions of 32%–36%. [contemporaneous record · strong] Strategy deck, PDF pp. 57–86 AGM archive, “VOD, Materials” table

Those targets make the source more valuable than a generic vision speech: later analysis can test timetable and scope. They also suggest a kill-or-resize discipline if technology transfer, local execution, or regulatory constraints block the expected path, although the deck does not state such a rule explicitly. [researcher inference · moderate] Strategy deck, PDF pp. 79–86

9. 2013-07-23 — SoftBank World 2013: data closes the operating loop — ★★

Source note · Access: partial. The official 52-minute VOD has no public transcript or captions; the 64-page deck and official/contemporaneous recaps control the analysis. Twelve days after Sprint closed, the keynote tied global scale to a longer technology curve and described using tens of billions of network records to optimize service. [contemporaneous record · moderate] Official deck, PDF pp. 4–9 and 33–42 Official recap, opening

The framework is data-to-decision: obtain distribution, collect proprietary operational data, turn it into optimization, and embed the result in customer workflows. That mechanism is more actionable than the global slogan, but the accessible evidence still omits model lift, causal attribution, cost, data rights, and return. Deck language is not treated as verified Son speech. [researcher inference · strong] Official deck, PDF pp. 33–42 and 47–58

10. 2014-03-27 — CCA Global Expo: coalition building after Sprint — ★★★

Source note · Access: full. An official eight-page transcript is controlling; YouTube automatic captions are only timestamp backup. Son argued that AT&T and Verizon's combined subscriber share had risen from 56% to 73% in five years and presented Sprint plus smaller carriers as a structural coalition against incumbent scale. The thesis required assets, capital, technology, and reciprocal commercial arrangements, not entrepreneurial passion alone. [investor-stated · strong] Official transcript, pp. 1–6

This is post-close but pre-outcome evidence. It reveals the board-level intervention logic—use Sprint chairmanship to recruit partners and improve challenger economics—while leaving price, leverage ceilings, contract terms, and failure triggers unresolved. The equal-opportunity mission explains urgency but is not a required-return threshold. [researcher inference · strong] Official transcript, pp. 1–6

11. 2014-06-20 — 34th AGM: investor-role boundary and Pepper's platform hypothesis — ★

Source note · Access: partial. The official strategy deck and governance documents survive; the webcast is closed. The presentation combines Sprint's scale opportunity with U.S. LTE, pricing, and satisfaction deficiencies, while explicitly speaking as investor rather than Sprint operator. It also advances Pepper as a cloud-connected, emotion-oriented robot whose learning could compound across devices. [contemporaneous record · moderate] Strategy deck, PDF pp. 25–36 and 49–61 AGM archive, “VOD, Materials” table

The two cases share an operating-platform logic: acquire or seed a distribution layer, transfer know-how, collect data, and improve the system. The Sprint actor boundary is unusually valuable; the Pepper claims are much weaker because customer value, learning validation, safety, cost, and unit economics are missing. Alibaba and partner/director material also reinforces that outcomes belong to founders, companies, and boards, not Son alone. [researcher inference · strong] Strategy deck, PDF pp. 25–36 and 49–61

12. 2014-07-15 — SoftBank World 2014: technology as an answer to demographic constraint — ★★

Source note · Access: partial. The 64-minute VOD is uncaptioned; the 66-page official deck and recap provide the stable record. The keynote starts from Japan's shrinking labor force and GDP pressure, then presents smartphones, tablets, cloud systems, connected objects, and robots as ways to raise productivity per worker. Company before-and-after examples are issuer-selected operating metrics, not audited customer returns. [contemporaneous record · moderate] Official deck, PDF pp. 3–27 and 39–58

The reusable reasoning chain is: identify the binding social or operational constraint, map a technology to it, distribute through an adoption platform, and learn from live feedback. Pepper as a labor substitute depends on task quality, deployment cost, reliability, and complementary process change; none is quantified. The deck therefore establishes a hypothesis and market-selection method, not verified speech or economics. [researcher inference · strong] Official deck, PDF pp. 39–58

13. 2015-06-19 — 35th AGM: SoftBank 2.0 and the permanent-capital selection claim — ★

Source note · Access: partial. The webcast is closed; the official strategy deck survives. SoftBank's stated transition is from a Japan telecom operator to a global technology group. The selection frame joins permanent capital with the right market, idea, and team, while a company-reported 43% aggregate IRR is used to support the historical case. That IRR spans ongoing holdings and incomplete cash-flow disclosure, so it is neither a fully realized series nor Son's personal return. [contemporaneous record · strong] Strategy deck, PDF p. 7 and pp. 19–27 AGM archive, “VOD, Materials” table

The source is an early articulation of a platform-investor identity: recycle operating cash and appreciated assets into the next information-revolution layer. It also exposes the recurring omission—entry price, ownership, realized distributions, losses, and exit policy are not supplied alongside the winner narrative. [researcher inference · strong] Strategy deck, PDF p. 7 and pp. 19–27

14. 2015-07-30 — SoftBank World 2015: structural adoption can outlive a market crash — ★★

Source note · Access: partial. The official VOD is uncaptioned; deck, recap, and a speaker-labeled third-party transcript provide partial text. Son distinguishes the dot-com bubble from the underlying IT regime, using SoftBank's roughly 99% share-price collapse and longer-run recovery as evidence that adoption can continue through financial destruction. He then links connected devices, data, analysis, and new business models, with Pepper for Biz as a product-platform example. [investor-stated · moderate] Official deck, PDF pp. 2–10 and 43–44 Logmi transcript, “インターネット産業の20年”

The sound rule is to track users, engagement, and workflow value separately from market price. The dangerous extension is to treat secular correctness as protection against overvaluation or insolvency. AI labor forecasts and robot economics remain weakly specified, so staged pilots and cohort evidence should precede large commitments. [researcher inference · strong] Official deck, PDF pp. 2–10 and 43–44 Logmi transcript, “インターネット産業の20年”

15. 2016-06-22 — 36th AGM: technology plus finance becomes the corporate identity — ★

Source note · Access: partial. The closed webcast leaves the official deck and governance record. SoftBank separates operating assets from investment assets and presents an ambition to combine leading free cash flow with leading investment IRR. A company-selected history reports roughly $3.7 billion invested and $92 billion of value, or 44% IRR and 25 times cost; the scope mixes realized and unrealized holdings and lacks full cash-flow reconstruction. [contemporaneous record · strong] Strategy deck, PDF pp. 3–5 and 34–40 AGM archive, “VOD, Materials” table

The organizational rule is to use operating cash and financing capacity to buy exposure to the next technology paradigm, then recycle appreciation. Sprint and the domestic telecom operations supplied distributed operating evidence; singularity forecasts supplied the next theme. The deck does not establish a stable leverage perimeter, independent return series, or successor-ready decision system. [researcher inference · strong] Strategy deck, PDF pp. 7–17, 34–40, and 41–58

16. 2016-07-21 — SoftBank World 2016: wait for the mapped asset, then move very fast — ★★

Source note · Access: full. Official video, English upload, deck, and recap cover the appearance, with automatic captions treated cautiously. Three days after announcing Arm, Son said he had followed the possibility for roughly a decade, met Arm's leadership over lunch about two weeks earlier, and completed diligence and funding work in that compressed period. [investor-stated · moderate] Official English-upload video, 02:03–05:38

Arm's rationale was architectural: a licensing ecosystem with broad mobile share and roughly 14.8 billion annual chip shipments could become the control point for IoT. The actionable sourcing rule is to maintain a long-term target map so that speed follows preparation, not discovery from scratch. But the talk supplies little entry-price analysis, no rejected-case record, and no explicit walk-away threshold; speed therefore remains both capability and risk. [researcher inference · strong] Official English-upload video, 02:03–05:38 Official deck, PDF pp. 19–25 and 30–39

17. 2017-06-21 — 37th AGM: the Vision Fund adds external capital to the control-point strategy — ★

Source note · Access: partial. The official deck is complete; the Brightcove webcast has no captions. After moving from FY2003 losses to FY2016 operating income above ¥1 trillion, the presentation connects AI and a trillion-device IoT forecast to Arm, OneWeb, robotics, and Guardant Health. The ¥10 trillion Vision Fund figure is committed vehicle capital, not deployed SBG cost, Son's money, NAV, distribution, or return. [contemporaneous record · strong] Presentation PDF, pp. 4–13, 14–35, and 36–37

The portfolio architecture becomes explicit: control selected infrastructure, own minority positions in category leaders, and use outside LP capital to broaden reach. Founders, directors, SBIA, SBG, and portfolio boards remain separate actors. Missing are entry-price rules, conflict management, measured synergy, reserves, and a vehicle-level loss budget. [researcher inference · strong] Presentation PDF, pp. 21–43

18. 2017-07-20 — SoftBank World 2017: a capital-and-technology industrial revolution — ★★

Source note · Access: full. The official 132-minute keynote, English upload, deck, and recap are available; automatic captions and guest voices remain separated. Son frames Arm, OneWeb, IoT, data, and deep learning as an integrated stack and gives ¥10 trillion of Vision Fund commitments as scale. A selected historical display reports about $11 billion invested and $175 billion accumulated value, 15 times and 44% gross IRR; it is issuer-selected, gross, estimated, and heavily unrealized. [investor-stated · moderate] Official deck, PDF pp. 10–45 Official English-upload video, 22:20–28:03

The important counterweight is concentration: the top five names account for 94% of the reported cumulative return. Portfolio-company examples voiced by guests belong to those executives, not Son. The talk shows why Son values a network of complementary leaders, but not how SBIA prices coordination, governs conflicts, or protects LPs when the same ecosystem contains competitors. [researcher inference · strong] Official deck, PDF pp. 31–33

19. 2017-09-19 — Bloomberg/Rubenstein: biography, founder selection, leverage, and succession in one record — ★★★

Source note · Access: full. Bloomberg's publisher transcript controls the recorded September 19 interview; Omny audio is a duplicate route. Son ties the Vision Fund to a 30-year singularity thesis and retells Fujita, Berkeley invention, Sharp, Yahoo, Alibaba, the dot-com crash, Vodafone Japan, and Arm. These stories show his self-described process—pursue future industries, judge a founder's intensity, and mobilize large capital—but most deal amounts and early biography are retrospective testimony, sometimes prompted by Rubenstein. [retrospective witness · provisional] Bloomberg transcript, singularity paragraphs beginning “So one vision which is singularity” and “That is the point” Bloomberg transcript, Fujita excerpt through “That's the one you should focus” Bloomberg transcript, deal-story paragraph groups beginning “But when you were there,” “One of the investments,” “Before Yahoo was so famous,” “Now let's talk about one big mistake,” “At that time I said,” and “Why did you spend”

The most decision-useful contrasts are within the interview. Alibaba is a leader-first story despite no revenue or correct plan; Vodafone Japan is a concentrated acquisition financed with roughly $2 billion of equity and $18 billion of debt; Arm is a platform forecast. At sixty, Son also planned to choose a successor in his sixties and become a coach—a dated intention that the 2026 life-plan extension later contradicts. [retrospective witness · moderate] Bloomberg transcript, Alibaba paragraphs Bloomberg transcript, Vodafone Japan financing paragraphs Bloomberg transcript, succession paragraphs 2026 presentation, pp. 6–8 and 62

20. 2018-06-20 — 38th AGM: the minority-owned “Cluster of No. 1” becomes the formal design — ★

Source note · Access: partial. The full deck and governance page survive; the webcast has no captions. SBG explicitly moves from telecom operator to strategic holding company and proposes a 300-year, self-renewing cluster: hold roughly 20%–40% of category leaders, recycle mature assets, and add new leaders rather than preserve one controlled conglomerate. [contemporaneous record · strong] 2018 presentation, pp. 7–33

The deck's examples mix SBG subsidiaries, SB Delta Fund, affiliates, and Vision Fund exposure, so portfolio labels do not define one vehicle. The shareholder-value response uses a company-assumption SOTP of ¥14,199 per share versus an ¥8,070 price, not executable liquidation value. A purportedly self-evolving organization also remains centered on Son, without a succession mechanism or consistent definition of category leadership. [researcher inference · strong] 2018 presentation, pp. 44–76

21. 2018-07-19 — SoftBank World 2018: invest before the precursor becomes obvious — ★★

Source note · Access: partial. The official 167-minute VOD lacks captions and no official deck was found; speaker-labeled Logmi text and contemporaneous recaps provide partial access. Son's framework is to identify precursors of a future industry and act early, using an Arm-at-the-edge, GPU/cloud/network stack and predictive applications as examples. Portfolio-company guests supply their own operating claims and must remain separately attributed. [investor-stated · moderate] Logmi transcript, “未来のことはわからない” through “AIの進化は二次曲線” AINOW recap, “それぞれの産業” and “これからはAI群戦略”

The stated selection ambition is to become a leading shareholder in the company most likely to redefine each industry through AI, then encourage family-company cooperation. That architecture may create option value, but the talk offers no entry-price test, governance protocol, conflict rule, measured synergy, loss limit, or exit discipline. Acting early is not enough when timing and valuation determine survival. [researcher inference · strong] AINOW recap, “これからはAI群戦略” Business Network recap, opening

22. 2019-03-25 — Milken Japan: the clearest full statement of the late-stage AI screen — ★★

Source note · Access: full. The 43-minute official video has timestamped English automatic captions, with speaker roles established by Milken. Son describes AI as an economy-wide platform shift and states a narrow Vision Fund screen: AI exposure, late-stage or pre-IPO unicorns, and the No. 1 company in each segment. In geographically segmented ride sharing, he instead sought multiple regional leaders. [investor-stated · strong] Official video, 24:42–29:27

The talk also contains the essential counterexample inside the origin story: Son says a $30 million gap cost him a large private Amazon position, motivating a much larger pool of capital. Removing the funding constraint can prevent missed opportunities, but it can also weaken price and sizing discipline. Product adoption through a market crash is a useful operating signal; OYO, autonomy, and AI/fintech forecasts remain untested here. [researcher inference · strong] Official video, 08:48–12:45 and 29:28–34:16

23. 2019-06-19 — 39th AGM: the AI orchestra and the fragility of headline IRR — ★

Source note · Access: partial. The 103-page deck is full; the webcast is uncaptioned. SBG casts itself as conductor of an AI orchestra—a holding company and funds coordinating founders—and reports $97.2 billion of Vision Fund commitments. March 2019 performance varies sharply by definition: 45% LP net equity IRR, 29% LP net blended IRR, and 62% from an SBG-LP-plus-manager-fee perspective. All include important NAV, leverage, coupon, fee, transfer, and unrealized-value effects. [contemporaneous record · strong] 2019 presentation, pp. 40–44 and 102

Vision Fund 2 and expansion from 415 to 1,000 SBIA employees were still plans, not outcomes. The ¥200 trillion 2040 path assumes no dividends and sustained 19% gross IRR for twenty years; it is explicitly non-assured. The orchestra metaphor clarifies intended coordination while leaving investment-committee dissent, capacity, pricing, and succession unresolved. [researcher inference · strong] 2019 presentation, pp. 45–80

24. 2019-09-19 — Sozo Summit with Larry Fink: a documented appearance with no recoverable claims — ★★★

Source note · Access: partial. SoftBank's official page establishes a 47:53 conversation between Son and BlackRock CEO Larry Fink and supplies the Brightcove player. It provides no transcript, captions, recap, moderator, or content timestamps; the permitted OpenAI transcription route was unavailable. Therefore no investment claim, decision rule, deal account, quotation, or forecast is inferred from speaker identity or event label. [contemporaneous record · strong] Official page and speaker paragraph

This is a substantive negative result: media availability is not textual access. The September 2019 event date controls over Brightcove's later ingestion metadata, and Son, Fink, SBG, and BlackRock remain separate people and institutions. This appearance is best treated as a high-value retrieval target, not filled with plausible themes from surrounding events. [researcher inference · strong] Official page and speaker paragraph

25. 2019-11-06 — FY2019 Q2 and WeWork: economic exposure without voting control — ★★★

Source note · Access: partial. The webcast has no captions; the official investor deck controls. At September 30, SoftBank presented $7.8 billion of total WeWork equity exposure split between an SBG wholly owned subsidiary and SVF1, across multiple securities and affiliates. Economic ownership could rise to 80%, yet WeWork remained an associate because SoftBank lacked majority shareholder and board voting rights. [contemporaneous record · strong] Official investor deck, Accounting slides 4–5

The valuation bridge fell roughly 74% from an earlier $29.6–$33.1 billion range to $7.8 billion post-transaction. Management proposed stopping new office openings, cutting costs, changing oversight, and supplying more capital while repeating that portfolio companies should be self-financing and not rescued. The enduring lesson is to underwrite governance rights, security terms, runway, and cross-vehicle exposure together; remedial process after impairment does not prove prospective control. [researcher inference · strong] Official investor deck, pp. 48–49 and 61, 77–84

26. 2019-12-06 — Tokyo Forum with Jack Ma: mutual recognition, fast conviction, and its base-rate problem — ★★★

Source note · Access: full. The official hour-long video and University of Tokyo recap provide timestamped, speaker-controlled evidence. Son recalls sampling about twenty Chinese startups in 2000 and choosing Ma after a short meeting because Ma expressed a future for small businesses and could attract followers; Ma recalls Son moving toward funding within minutes. Their recollections agree on fast mutual recognition but are not the 2000 term sheet or an audited return record. [retrospective witness · provisional] Official video, 03:14–08:22 and 13:36–17:11

Son's current rule was to back ambitious AI founders and connect them in a renewable 300-year cluster, while treating entrepreneurs as company captains. Ma supplies the strongest countervoice: too much money can generate mistakes and capital is a responsibility. The dialogue spotlights an extreme winner but omits the comparison set, failures, false positives, formal board rights, and exit rules—exactly the evidence needed to test charisma-based selection. [researcher inference · strong] Official video, 22:16–29:18 and 54:06–59:17

27. 2020-02-12 — FY2019 Q3: a smaller loss is not a turnaround — ★★

Source note · Access: partial. The official webcast has no captions; the earnings deck controls. SoftBank called the SVF segment's movement from a ¥970.3 billion Q2 loss to a ¥225.1 billion Q3 loss the beginning of a turnaround. Public-mark improvement and a smaller quarterly loss are directional changes, not recovery of cumulative value or a realized LP return. [contemporaneous record · strong] Official earnings deck, slides 4–8

WeWork's $5.3 billion liquidity plan assumed signed-building openings, a pause elsewhere, cost reduction, and 2022 break-even. Installing Sandeep Mathrani added real-estate operating experience, but biography and funding do not prove the forecast. The repeated no-rescue rule remains unresolved against continuing support, and the pre-COVID date prevents later office-demand damage from being treated as known. [researcher inference · strong] Official earnings deck, slides 26–30 and 58–59

28. 2020-04-06 — Forbes under pressure: Son admits price and founder errors — ★★★

Source note · Access: full. Forbes supplies an edited article with direct, publisher-vetted Son quotations from a private interview and observed investor summit, not the full raw exchange. Son admits two WeWork mistakes: paying too high a valuation and believing too much in the entrepreneur. He nonetheless separates tactical regret from strategic continuity and forecasts that new management can produce a decent return. [investor-stated · strong] Forbes, node 10 and node 23

His portfolio defense is power-law: tolerate roughly fifteen failures if a similar number of breakouts can carry the fund. The reported $150 billion sufficiency calculation comes from insiders and depends on principal, a 7% preferred distribution, and portfolio outcomes; it is not Son's guarantee. Forbes' counter-thesis is equally important: huge checks can inflate burn, late-stage valuations are vulnerable, and cash conservation requires a different managerial skill set. Parent asset sales are not SVF1 returns. [researcher inference · strong] Forbes, nodes 38 and 44–52

29. 2020-05-18 — FY2019 results: survival liquidity alongside long-duration conviction — ★★

Source note · Access: partial. The webcast is uncaptioned; the release, 61-page deck, and financial report control. SoftBank reported negative ¥1.365 trillion consolidated EBIT and negative ¥962 billion attributable net income. From inception through March 2020, 26 SVF1 markups contributed $13.4 billion while 47 markdowns produced $14.2 billion of losses, including $0.2 billion realized loss; the net $0.8 billion loss is an SBG-consolidated mark bridge, not investor IRR. [contemporaneous record · strong] Official earnings deck, slides 11–27

The crisis response raised $11.5 billion through Alibaba-linked forward, floor, and collar contracts while presenting 14% LTV, then described a pandemic valley in which digital services might lead recovery. Arm's 166 billion cumulative shipments and falling adjusted EBITDA are company-reported operating inputs, not an investment return. The playbook is a barbell: monetize liquid holdings to survive while retaining the technology thesis; mission continuity does not cancel financing risk. [researcher inference · strong] Official earnings deck, slides 34–54

30. 2020-06-25 — 40th AGM: NAV replaces EBIT as the central scoreboard — ★★

Source note · Access: partial. The official deck is full; the webcast has no captions. SBG treats COVID-era online adoption as acceleration and elevates shareholder value—company-calculated at $218 billion on June 24—over EBIT as its key KPI. Holdings less net debt is a mark-based measure, not cash, distributions, or Son's return. [contemporaneous record · strong] 2020 presentation, pp. 4–32 and 58–62

Sprint's stated 25% equity IRR includes T-Mobile shares, sale proceeds, options, and a contingent right while excluding a potential indemnity; the asset-monetization program likewise mixes executed transfers, financing contracts, and expected proceeds. The useful rule is to keep marks, cash, forwards, and contingent value separate. The counter-risk is that NAV can foreground capital allocation while obscuring portfolio-company unit economics and private-mark realizability. [researcher inference · strong] 2020 presentation, pp. 34–43

31. 2021-06-23 — 41st AGM: “Vision Capital” scales breadth faster than governance evidence — ★

Source note · Access: partial. The deck is full; no captions or transcript exist. SBG reframes itself as capital provider to the information revolution and reports ¥26 trillion NAV. Its 264-company ecosystem mixes SVF1, SVF2, LatAm, JVs, exited names, 44 approved-but-unclosed firms, and duplicate companies; it is not 264 unique, live, completed investments. [contemporaneous record · strong] 2021 presentation, pp. 17–32

Company-selected Internet and combined-fund IRRs of 43% contain unrealized NAV, financing, manager-fee, and transferred-asset effects. The ambition to expand toward 500 companies may improve sourcing breadth but raises diligence, board-capacity, and accountability questions. Purpose language and founder support do not remove funding constraints; the 2022 LTV/cash reset is later counterevidence. [researcher inference · strong] 2021 presentation, pp. 33–56

32. 2022-02-08 — Arm/NVIDIA termination: change the exit route, not the platform thesis — ★★

Source note · Access: partial. The official interview player has no captions; same-day releases provide direct written attribution and transaction controls. SoftBank and NVIDIA terminated the sale because regulatory challenges prevented completion, and Son immediately pivoted to preparing an Arm IPO. The $1.25 billion non-refundable deposit was termination consideration recognized as profit by an SBG subsidiary, not Arm operating return or full sale proceeds. [contemporaneous record · strong] Joint release, opening and Son paragraphs Transaction notice, paragraphs 1–3

The strategic claim—Arm as a platform across mobile, cloud, automotive, IoT, metaverse, and AI—survived the failed counterparty route. SBGC and SVF1 held different portions, including 24.99% attributable to SVF1, so later marks and proceeds require vehicle allocation. The adaptive rule is valuable; the source gives no minimum IPO valuation, timing probability, or net whole-position return. [researcher inference · strong] Transaction notice, paragraph 1 and footnote 1

33. 2022-05-12 — FY2021 results interview package: a stated 30% hurdle for stressed follow-ons — ★★

Source note · Access: partial. The post-results interview itself is uncaptioned; quotations and claims come only from the distinct same-day, speaker-labeled global-call transcript. Son argued that private marks could lag public repricing by one or two years and said a stressed down- or flat-round investment should proceed only if SoftBank expected a 30% IRR, not merely because the company needed rescue. [investor-stated · strong] Official global-call transcript, A3 and A15, pp. 1–5

Defense had operating form: reduce deployment by one-half to three-quarters over the next year, mark quarterly with specialist teams, monitor runway, and retain cash and realization optionality. Son described about thirty geography/sector teams, SVF1 as concentrated, SVF2 as more distributed, and Arm as a core position he wanted largely to retain. The 30% figure lacks inputs, duration, probability, and later cash-flow verification; it is a stated hurdle, not performance evidence. [researcher inference · strong] Official transcript, A18, A20–A21, A26 and A32

34. 2022-06-24 — 42nd AGM: theme persistence inside an LTV and cash envelope — ★★

Source note · Access: partial. The year page and deck are full, while the webcast and edited highlight have no captions. The 475-company ecosystem mixes SVF1, SVF2, LatAm, approved-but-unclosed firms, JVs, exits, token exposure, and duplicates, so it cannot be used as a clean diversification denominator. Arm remains the enabling-platform anchor. [contemporaneous record · strong] 2022 presentation, pp. 15–24

The new portfolio rule is explicit: lower LTV, strictly select new investment, monetize appropriately, and cover maturities. FY2021 year-end LTV was 20.4% under company definitions, versus an illustrated 25% normal threshold and 35% emergency boundary; reported standalone cash included an undrawn line and must not be read as cash on hand. Relative to 2021 expansion, the theme stays constant while pace becomes subordinate to financing survival. [researcher inference · strong] 2022 presentation, pp. 33–43

35. 2022-08-08 — FY2022 Q1: the strongest public defense-mode postmortem — ★★★

Source note · Access: full. The paired official global-call transcript gives full Son Q&A; the separate video interview remains uncaptioned and is not treated as a duplicate transcript. Son directed continuous defense—accumulate cash, cut SBG expenses, and avoid excess risk because the downturn could last months or years. He also identified the process error precisely: selecting the No. 1 company did not prevent overpayment. [investor-stated · strong] Official transcript, A1, A4, A6, A9, A16 and A23

Son said roughly 80%–90% of portfolio companies had 24 months of cash, but memory-qualified the estimate and did not define the vehicle denominator. Rajeev Misra remained responsible for SVF1 while Son took deeper leadership of SVF2; this is responsibility evidence, not sole attribution. The operating rule is category strength plus an independent price test, with willingness to sell non-core assets at a reasonable price. Neither “reasonable” nor the trigger for renewed offense is quantified. [researcher inference · strong] Official transcript, A5, A8, A22 and A32–A34

36. 2023-06-21 — 43rd AGM: defense flips to offense without a disclosed risk budget — ★

Source note · Access: partial. The official deck is full; the webcast has no captions. SBG declares a move from defense to offense and uses GPT-4, an internal contest that generated 52,000 proposals, and a broad application map to frame generative AI as a shift from calculation/search to inference/creation. The contest proves experimentation, not commercial value, and future capability claims are illustrative or non-guaranteed. [contemporaneous record · strong] 2023 presentation, pp. 3 and 7–23

Arm-supplied metrics support the infrastructure anchor, while Rene Haas's SBG board election shows alignment and related-party complexity. The implied playbook is to anchor in Arm, test applications internally, and resume thematic exposure. The bridge does not state whether prices, liquidity, fundamentals, or conviction justified the pivot, nor does it replace the prior LTV ceiling with a new exposure cap or stop rule. [researcher inference · strong] 2023 presentation, pp. 25–33

37. 2024-06-21 — 44th AGM: Arm becomes both proof point and concentration risk — ★★★

Source note · Access: partial. The official deck is complete; the webcast is uncaptioned. SBG reports ¥34 trillion NAV on June 20, up ¥20 trillion in a year, with Arm at 55% of holding equity value. It also presents a company-reported ¥24.6 trillion Arm investment result, 10 times gross multiple, and 34% equity IRR combining realized IPO proceeds, marked retained shares, in-kind distributions, and an NVIDIA deposit, before tax and excluding related costs and intra-group effects. It is not fully realized or Son's personal return. [contemporaneous record · strong] 2024 presentation, pp. 5–15

The thesis accelerates from generative AI to AGI/ASI and a stack of chips, data centers, and robots. Arm is both strategic control point and financing capacity; the same 55% weight makes NAV and borrowing capacity sensitive to one listed asset. The deck's claim that SoftBank was founded for ASI is a 2024 reinterpretation, not a contemporaneous 1981 fact. [researcher inference · strong] 2024 presentation, pp. 17–40

38. 2024-10-03/04 — SoftBank World 2024: the forecast accelerates and agents become the mechanism — ★★

Source note · Access: partial. The full official video has only Japanese automatic captions; claims are timestamped paraphrases and no quotation is admitted. Son compresses an earlier within-ten-years AGI forecast to roughly two or three years and places ASI at roughly 10,000 times human intelligence within ten years. That is a dated acceleration, not evidence that either milestone occurred. [investor-stated · moderate] Official video, 00:00:56–00:03:22 and 00:05:53–00:07:15

The mechanism is an organization of agents: many systems perform trial and error against rewards, select stronger outputs, and coordinate into swarms. Son reports thousands of agents and 188 invention candidates submitted for patents, while acknowledging that grants were uncertain. Submission is not defensible IP, product, revenue, or return; the personal trading-agent example is illustrative, not a live account. The missing controls are reward design, human approval, cost, false-positive rate, permissions, and shutdown conditions. [researcher inference · strong] Official video, 00:27:02–00:41:03

39. 2024-10 — FII 8 with Richard Quest: forecast the bottleneck, conserve capital, but disclose no deal — ★★

Source note · Access: partial. Only FII's edited executive-report page is lawfully accessible; there is no verified full-video route. The organizer attributes to Son an ASI forecast around 2035, at 10,000 times human intelligence, a $9 trillion generative-AI capital-expenditure estimate, at least 5% additional global GDP, 400 GW of AI-data-center power, and 200 million chips. None comes with period, model, benchmark, probability, or security-level valuation inputs. [investor-stated · moderate] FII 8 Executive Report, p. 61

Son also said he was conserving funds for an unidentified large move. The investable chain is visible—forecast superintelligence, infer compute and energy scarcity, preserve liquidity for concentration—but the vehicle, target, approval, size, and downside rules are not. FII's page says ASI, not merely AGI, and its editor's transitions must remain separate from Son's short direct attributions. [researcher inference · strong] FII 8 Executive Report, p. 61

40. 2024-11-14 — NVIDIA AI Summit Japan with Jensen Huang: stack orchestration and outcome-aware regret — ★★★

Source note · Access: full. NVIDIA's creator-provided English captions cover the complete dialogue; the On-Demand session is a duplicate route. Son describes founder recognition through shared vision, passion, and instinct, while Huang supplies the celebratory claim that Son chose the winner in every computing generation. Son's response that he was lucky and the omitted denominator prevent Huang's praise from becoming an audited selection record. [retrospective witness · strong] Official video, 00:57:42–01:00:21

The operating plan is an AI grid for Japan: data centers train, AI-RAN distributes, the telecom network and consumer services seed applications, and subsidized compute encourages local experimentation. Their NVIDIA history—privatization discussion, market purchase, and attempted Arm combination—is recalled after enormous appreciation, so regret is outcome-aware and says little about financing feasibility or hold discipline. Speaker and entity boundaries matter: Huang, Son, Miyakawa, SBG, SoftBank Corp., Arm, and NVIDIA contribute different claims and actions. [researcher inference · strong] Official video, 01:03:49–01:21:38

41. 2025-01-21 — Stargate White House remarks: announcement scale is not deployed capital — ★★

Source note · Access: partial. The official White House page and embedded YouTube are duplicate routes with automatic English captions; the two-minute speech is paraphrased, not quoted. President Trump, not Son, opens with arithmetic combining earlier U.S. commitments. Son separately recounts a $50 billion commitment becoming $70 billion, a later $200 billion commitment, and a $500 billion SoftBank–OpenAI–Oracle–MGX Stargate ambition. The clip supplies no deployment schedule, asset list, vehicle reconciliation, or return. [contemporaneous record · moderate] Official White House video, 00:00–01:27

The causal thesis is that compute infrastructure enables AI across industries and that U.S. innovation leadership should anchor the buildout. Son's institutional role moves from selecting companies toward convening and financing a consortium, but announcement-stage leadership is not closed financing, commissioned capacity, or cash realization. Partner obligations, governance, power contracts, demand, unit economics, and stop rules remain absent. [researcher inference · strong] Official White House video, 00:00–01:27

42. 2025-06-27 — 45th AGM: the minority cluster becomes a vertically linked ASI platform — ★★★

Source note · Access: partial. The full deck survives; the webcast is uncaptioned. SBG reports ¥31 trillion NAV with Arm at 51% and Vision Funds at 27%, then defines a ten-year ambition across chip design, models, agents, and robots. Graphcore was completed; Ampere remained pending. An “up to $32 billion” OpenAI figure combines $2.2 billion in SVF2 with a conditional SBG follow-on up to $30 billion that could fall to $10 billion depending on recapitalization timing. [contemporaneous record · strong] 2025 presentation, pp. 8–23

The architecture is changing: Arm IP, adjacent chip acquisitions, a frontier-model partner, agents, and physical systems create coordination potential but also common-factor exposure. Cristal's goal across employees and customers is non-guaranteed and lacks adoption or ROI evidence. The deck's Intel-8080-at-seventeen origin story and 1981 ASI purpose are retrospective leads, not contemporaneous founding records; governance dissent on stock-option revisions is a relevant counterweight. [researcher inference · strong] 2025 presentation, pp. 4–7 and 24–38

43. 2025-07-16 — Son and Sam Altman at SoftBank World: strategic association, content inaccessible — ★★★

Source note · Access: partial. SoftBank's official page establishes a distinct 18:23 joint appearance by Son and OpenAI CEO Sam Altman, held July 16 and published September 1. The Brightcove player exposes no transcript or caption track, and the adjacent second-half keynote is not a transcript of this session. No claim, forecast, commitment, quotation, or division of labor is reconstructed. [contemporaneous record · strong] Official webcast, title and description

The recoverable evidence is relationship and timing: the two leaders appeared publicly amid SoftBank's agent, OpenAI, and Stargate initiatives. Named roles do not make Son, Altman, SBG, SoftBank Corp., OpenAI, SB OpenAI Japan, Stargate, or any fund interchangeable. This is the corpus's highest-priority modern transcript gap because the relationship is central and metadata alone cannot resolve governance or commitments. [researcher inference · strong] Official webcast, title and description

44. 2025-07-16 — SoftBank World 2025 keynote: from agent experiment to a one-billion-agent operating target — ★★

Source note · Access: partial. The official webcast and YouTube are duplicate routes; only Japanese automatic captions are available, so the note uses paraphrase. Son turns the 2024 swarm idea into the Cristal program and a one-billion-agent target, with telecom-network monitoring as an early architecture and a vision of employees managing hundreds or thousands of agents across functions. Target count is not production deployment. [investor-stated · moderate] Official caption route, 00:00:42–00:04:00

The platform claim is an agent operating system that lets systems chain work, create descendants, and optimize against goals and rewards. A stated ¥40 monthly cost per agent omits utilization, model/compute, integration, evaluation, supervision, security, and incident response, so it is not all-in unit economics. The reader should demand agent inventories, authorization scopes, rollback controls, error rates, and realized productivity before treating abundance as operating leverage. [researcher inference · strong] Official caption route, 00:17:03–00:37:40

45. 2026-02-24 — TIME: current allocation thesis under a reporter's chronology check — ★★★

Source note · Access: full through lawful search-index route. Direct canonical and AMP requests returned 406, but the full canonical article, byline, timestamp, and section structure were exposed through a lawful search index. Son says AI will transform essentially every industry, forecasts 10,000-times-human ASI within a decade, and gives 10%–30% global-GDP scenarios for ASI plus physical AI. Those claims lack a model, probability, and falsification rule. [investor-stated · strong] TIME, opening and paragraph beginning “Not that Son agrees”

The profile's critical control is voice separation. Son says he sold NVIDIA because of COVID-era financial constraints; Charlie Campbell corrects that SoftBank sold in January 2019, a year before the pandemic. Son says new investment in China is zero and U.S. frontier research is the focus; Campbell, Lionel Barber, Alok Sama, and Forrest Mozer supply separate reporting, interpretation, witness accounts, and dispute. NVIDIA regret reveals capital scarcity across admired assets, while WeWork and Barber's founder-syndrome critique show why visionary affinity needs governance and price controls. [researcher inference · strong] TIME, paragraphs beginning “Then there were the misses,” “When it comes to selling,” and “Son's break-up with China”

46. 2026-06-24 — 46th AGM: physical ASI, a ¥1 quadrillion target, and an extended key-person horizon — ★★★

Source note · Access: partial. The official 66-page deck is full; the webcast has no captions. Son's life plan changes from an earlier handoff in his sixties to realizing ASI in his seventies. The deck presents cognitive agents and physical-ASI robots across a four-part stack of models, semiconductors, infrastructure, and robots, while Ohio and France remain announced projects and ABB Robotics a pending acquisition. [contemporaneous record · strong] 2026 presentation, pp. 6–38

SBG reports ¥74 trillion NAV on June 23 and targets ¥1 quadrillion in 2042, about fourteen times in sixteen years; the target is expressly non-guaranteed and the NAV is a company-defined financing construct, not cash. The succession change heightens key-person risk just as execution spreads across OpenAI, Arm, financiers, subsidiaries, fund managers, and robot companies. Announced capacity, pending acquisition, operating asset, committed capital, mark, debt-adjusted NAV, and realized proceeds must remain distinct. [researcher inference · strong] 2026 presentation, pp. 40–66

Recurring stories and the decision frameworks behind them

1. Start with the future state, then work backward to control points. Son rarely begins with a current-market screen. He forecasts the next dominant interface or intelligence layer, identifies bottlenecks, and then seeks assets at those layers: mobile distribution, spectrum and networks, Arm's instruction-set ecosystem, AI models, compute, data centers, agents, and robots. [investor-stated · strong] Sprint transcript, lines 701–787 2026 presentation, pp. 27–38

The strength of this method is coherence: it helps an allocator ignore incumbent category labels and map enabling infrastructure. The weakness is forecast stacking. Each layer can be technologically important while the security is overpriced, the owner lacks control, demand arrives late, or a cheaper architecture wins. The 2024–2026 ASI stack needs independent probabilities for model progress, power, chips, project financing, utilization, regulation, and physical deployment rather than one shared conviction multiplier. [researcher inference · strong] Sprint transcript, lines 701–787 2026 presentation, pp. 27–38

2. Treat founder quality as a force multiplier—but not a substitute for governance. The Alibaba, Fujita, and pioneer stories emphasize persistence, conviction, recruiting power, and a felt recognition of unusual people. Son's own best counterevidence is WeWork: he says he paid too much and believed too much in the entrepreneur. The later No. 1-company postmortem adds that even a surviving leader can be a poor investment at the wrong price. [investor-stated · strong] Tokyo Forum, 03:14–16:18 Forbes, node 10 FY2022 Q1 transcript, A9

The modern translation is a two-key approval: exceptional founder evidence plus an independent security-and-governance case. Required controls include unit economics, runway, board and voting rights, related-party risk, check-size limits, follow-on obligations, reference calls beyond admirers, and a pre-written disconfirming case. Charisma should increase diligence, not reduce it. [researcher inference · strong] Tokyo Forum, 03:14–16:18 Forbes, node 10 FY2022 Q1 transcript, A9

3. Use capital scale and speed as competitive products. Sprint, Arm, the Vision Fund, Stargate, and OpenAI all express the belief that unusually fast, large capital can change what a company or market can attempt. The 2016 Arm appearance suggests the best version: track an asset for years, then compress execution once it becomes actionable. The Amazon regret story shows the motivation to remove funding constraints. [retrospective witness · moderate] SoftBank World 2016 route Milken, 08:48–10:37

Scale creates its own failure mode. A larger check can bid up price, subsidize uneconomic growth, concentrate the portfolio, weaken syndicate price discovery, and create implicit rescue expectations. The proper gate is not whether capital can accelerate a company, but whether the incremental capital raises probability-weighted value per share after governance, dilution, burn, and financing risk. [researcher inference · strong] Forbes, nodes 46–47

4. Build ecosystems around scarce infrastructure, but measure rather than narrate synergy. Son moves repeatedly from a single asset to a network: telecom distribution plus content; Sprint plus smaller carriers; Arm plus IoT; Vision Fund founders; the Japan AI grid; or models, chips, data centers, agents, and robots. The hypothesis is that common infrastructure, data, and relationships lower costs and accelerate adoption. [investor-stated · moderate] CCA transcript, pp. 2–6 NVIDIA AI Summit, 01:03:49–01:19:11

The missing evidence is consistently the same: contracted cross-selling, shared-customer conversion, incremental margin, data-rights economics, integration cost, governance when portfolio companies compete, and attribution of benefits by vehicle. Without those measures, an ecosystem can be a collection of correlated exposures rather than a flywheel. [researcher inference · strong] CCA transcript, pp. 2–6 NVIDIA AI Summit, 01:03:49–01:19:11

5. Separate market-price collapse from operating adoption, but never from financing survival. Son's dot-com lesson is that users and engagement can grow while shares collapse. The 2020 valley model and 2022 defense posture add the missing half: an investor must retain enough liquidity and reduce burn to survive until the thesis can mature. [investor-stated · strong] Milken, 10:41–12:45 FY2022 Q1 transcript, A16 and A20–A23

This yields a usable cycle framework: monitor adoption separately from valuation; model portfolio-company runway; impose holding-company LTV, liquidity, and maturity limits; slow deployment until private marks reset; and define the conditions for offense before the crisis. The public record supplies many elements but leaves the offense trigger discretionary. [researcher inference · strong] Milken, 10:41–12:45 FY2022 Q1 transcript, A16 and A20–A23

6. Treat data and experiments as a bridge from vision to operations. Network telemetry in 2013, the 52,000 internal AI proposals in 2023, 188 patent candidates in 2024, and the 2025 Cristal target show a repeated instinct to turn a thesis into internal trials. That is stronger evidence than a forecast alone. [documented behavior · moderate] 2023 presentation, p. 8 SoftBank World 2024, 00:34:20–00:41:03

The evaluation discipline is underdeveloped in public. Proposal count, patent submission, agent count, and low nominal unit cost are activity measures. Decision-quality evidence would include control groups, total cost, false positives, grants, adoption, error and incident rates, human-review load, counterfactual productivity, and realized cash benefit. [researcher inference · strong] 2023 presentation, p. 8 SoftBank World 2024, 00:34:20–00:41:03

7. Keep every return numerator and denominator attached. Across the talks, Son and SBG use gross and net IRR, equity and blended IRR, SBG-perspective manager economics, SOTP, NAV, markups, accumulated value, and transaction-specific return. Those measures answer different questions. Many include unrealized holdings, financing, transfers, coupons, fees, or contingent value. [contemporaneous record · strong] 2019 presentation, pp. 43–44 and 102 2024 presentation, p. 15

The analyst's rule is simple: record entity, vehicle, cash-flow perspective, cost, dates, realized proceeds, remaining mark, debt, fees/carry, tax, currency, contingent components, and benchmark before comparing any two returns. Repetition of a high IRR under shifting definitions is not corroboration. [researcher inference · strong] 2019 presentation, pp. 43–44 and 102 2024 presentation, p. 15

How the framework changed, 2008–2026

2008–2012 — operator discipline and balance-sheet option value. The early record is corporate rather than fund-centric: combine telecom infrastructure with content, improve mobile economics, pay down debt under explicit cash-flow targets, and then reallocate capacity toward growth and dividends. The proposed preferred-share instrument and its withdrawal show that governance can constrain control-preserving finance. [contemporaneous record · strong] 2008 archive 2010 archive 2012 archive

2012–2016 — control transactions and platform transfer. Sprint is a leveraged operating-control bet justified ex ante by market scale, mobile-internet transition, and transfer of capital and network strategy. Arm is a different control point: an IP-licensing platform that can span devices and industries. In both, Son treats speed and board-level influence as part of the value proposition, while public price discipline remains thin. [researcher inference · strong] Sprint SEC transcript SoftBank World 2016

2017–2019 — external capital and the minority-company cluster. The Vision Fund scales the playbook from selected corporate assets to scores of late-stage AI/category leaders financed partly by LPs. SBG becomes the “conductor,” and the 20%–40% cluster is supposed to renew itself over centuries. The gain is reach; the new problems are manager/LP alignment, private marks, governance capacity, cross-vehicle exposure, and winner concentration. [contemporaneous record · strong] 2018 presentation, pp. 7–33 2019 presentation, pp. 27–50

2019–2022 — intervention, liquidity, and explicit process regret. WeWork forces the public record to confront valuation, founder trust, control without votes, follow-on financing, and cross-vehicle concentration. COVID then makes asset monetization and holding-company liquidity central. By August 2022 Son has added the missing price clause to category leadership and adopted cash, cost, LTV, runway, and slower deployment as survival rules. [researcher inference · strong] WeWork deck FY2022 Q1 transcript

2023–2026 — renewed offense and vertical AI concentration. The organization returns to offense around Arm, then broadens into chip companies, OpenAI, Stargate, agents, and robotics. This is not merely a return to the 2018 minority cluster: the stack is more vertically coordinated and more exposed to a common AI factor. Forecasts accelerate from a long singularity horizon to AGI in two or three years and ASI within about a decade, while announced infrastructure and acquisition states expand faster than disclosed project economics. [researcher inference · strong] 2023 presentation, pp. 3–32 2025 presentation, pp. 15–28

Biographical turning points and what they can—and cannot—prove

• Identity and discrimination: Son says Korean ancestry and discrimination made him work harder; later profiles connect dual identity and mortality to urgency. These are useful motivation accounts, not causal proof of investment performance. [retrospective witness · moderate] Bloomberg interview, discrimination paragraphs TIME, biography sections
• Den Fujita and future-industry orientation: Son recalls persistently securing a teenage meeting and receiving advice to study future industries and computers. The story plausibly explains top-down theme selection, but no contemporary Fujita record survives in the corpus. [retrospective witness · provisional] Bloomberg excerpt
• Berkeley invention routine and early capital: daily invention, the Sharp translator, and arcade-machine accounts support an experimentation-and-self-financing identity. Amounts and contributor economics vary across later sources; the Mozer dispute and conflicting arcade figures remain unresolved. [retrospective witness · provisional] TIME, final biographical section
• Dot-com collapse: Son uses a roughly 99% SoftBank share-price fall and Alibaba endurance as proof that adoption can outlive a crash. It supports psychological tolerance and the market-versus-product distinction, but does not show that all positions survived, that valuation was irrelevant, or that later crises share the same economics. [investor-stated · moderate] Milken, 10:41–12:45
• Illness and succession: later profiles tie a hepatitis diagnosis and mortality awareness to urgency. In 2017 Son planned a handoff in his sixties; in 2026 he extended the personal mission into his seventies. That dated change is evidence of key-person persistence, not a completed succession process. [contemporaneous record · strong] Bloomberg succession paragraphs 2026 presentation, pp. 6–8 and 62

Deal rationales reconstructed from the public record

Alibaba — founder and market before metrics. Son's later accounts say he entered China as the next internet market, sampled founders quickly, and chose Jack Ma for conviction, purpose, and recruiting force despite little revenue and an inadequate plan. This is the purest expression of the people heuristic and the strongest positive outcome associated with it. It is also retrospective and winner-selected: the corpus lacks the original memo, security, exact cost, ownership path, rejected-founder denominator, and attribution of Alibaba's execution among Ma, staff, directors, shareholders, and markets. [retrospective witness · provisional] Tokyo Forum, 03:14–17:11

Vodafone Japan and Sprint — buy control when distribution blocks the thesis. The Vodafone Japan recollection links a mobile-internet thesis to the need for spectrum and a heavily leveraged carrier acquisition. Sprint's SEC record supplies the stronger ex-ante logic: large market, mobile-data transition, weak challenger network, procurement and operating transfer, and rapid capital deployment. Later T-Mobile consideration can be evaluated only against corporate cost, financing, options, contingent rights, indemnities, and ownership period; it cannot be converted into a Vision Fund result. [investor-stated · strong] Sprint SEC transcript 2020 presentation, pp. 34–35

Arm — own the instruction-set tollbooth across successive computing waves. In 2016 Son presented Arm as a decade-watched licensing ecosystem positioned for IoT. In 2022, regulatory failure ended the NVIDIA sale and shifted the exit route to an IPO without changing the platform thesis. In 2024, SBG's company-reported result was dominated by marked retained equity and must remain pre-tax and partly unrealized. By 2025–2026 Arm had become both the core ASI asset and a source of correlated NAV and financing risk. [researcher inference · strong] Arm termination release 2024 presentation, p. 15

Vision Fund — remove capital scarcity and cover the late-stage AI winner set. The stated screen is late-stage/pre-IPO AI category leaders, supplemented with multiple geographic leaders where network effects remain local. External LP capital widened the search and check-size frontier. The economic case is power-law: a minority of extraordinary outcomes must cover failures, preferred distributions, fees, and costs. The public record does not show that bigger checks improved price discovery, and the 2017/2019 metrics were concentrated, largely unrealized, and definition-sensitive. [investor-stated · strong] Milken, 24:42–29:27 2019 presentation, pp. 43–44 and 102

WeWork — a technology-platform interpretation overwhelmed price and control. Son continued to view real estate through a technology-and-community lens, but the contemporaneous record shows a severe valuation reset, large exposure across SBG and SVF1, and economic ownership without majority voting or board control. His 2020 admission closes the two most important errors—valuation and entrepreneur trust—while the rescue/no-rescue question remains unresolved because the expected return on follow-on financing was not disclosed. [investor-stated · strong] WeWork deck, Accounting slides 4–5 and SVF slide 6 Forbes, node 10

NVIDIA — correct platform recognition with an unproven hold counterfactual. The 2024 Huang dialogue supports that Son valued NVIDIA and explored several routes, while the 2026 TIME interview records deep regret and current capital prioritization toward OpenAI, data centers, and Stargate. Campbell's January 2019 sale-date correction defeats the COVID chronology Son offered. Even if conviction was real, the record does not establish feasible privatization terms, authorization, vehicle constraints, or the ability to hold through later volatility. [researcher inference · strong] NVIDIA AI Summit, 01:19:13–01:21:38 TIME, final section

OpenAI, Stargate, and the ASI stack — integrate the bottlenecks rather than pick one application. The 2025 AGM distinguishes SVF2's funded amount from SBG's conditional maximum; Stargate is a consortium ambition, not a single funded security. The rationale is vertical: model partnership creates demand for chips, data centers, distribution, agents, and physical systems. This may capture more of the value chain, but also couples capital needs and correlated failure across the stack. Partner economics, exclusivity, governance, project finance, power, utilization, and stop rules are not public in these appearances. [contemporaneous record · strong] 2025 presentation, p. 23 White House remarks, 01:00–01:46

Recollection versus contemporaneous record

• Alibaba: the 2017 and 2019 interviews corroborate what Son and Ma said years later, not what was documented in 2000. Dollar figures are partly interviewer/moderator framing, and neither route supplies the original term sheet, ownership, dilution, or realized distributions. Record the people heuristic as retrospective testimony, then audit the transaction separately. [retrospective witness · provisional] Bloomberg Alibaba paragraphs Tokyo Forum, 03:14–17:11
• Sprint: the 2012 SEC-filed transcript is the controlling ex-ante rationale. The 2013 synergy deck is a dated forecast, CCA is a post-close strategy appeal, and the 2020 25% IRR is a later company calculation containing marked and contingent components. Later recollection must not overwrite those states. [contemporaneous record · strong] SEC transcript 2020 presentation, pp. 34–35
• WeWork: the November 2019 deck documents exposure, control, valuation, and the proposed intervention contemporaneously. The April 2020 Forbes interview records Son's admission and forward-looking recovery belief. Neither later remorse nor later outcome changes the pre-existing governance and price facts. [contemporaneous record · strong] WeWork deck Forbes
• NVIDIA: the 2024 Son–Huang dialogue documents shared recollection of attempted routes; the 2026 profile documents Son's COVID explanation and Campbell's contradiction. The January 2019 transaction date has precedence over a memory-based pandemic cause. [contemporaneous record · strong] TIME, paragraph beginning “When it comes to selling”
• Arm: 2016 supplies post-announcement rationale; 2022 supplies failed-sale and IPO-pivot evidence; 2024 supplies SBG's result definition. Each answers a different question. Combining them into an inevitable success story would erase regulatory failure, ownership across SBG/SVF1, deposit treatment, IPO proceeds, retained marks, tax, and intra-group effects. [researcher inference · strong] Transaction notice 2024 presentation, p. 15
• Founding purpose and ASI: 2010's contemporaneous announcement speaks of happiness through the IT revolution. The 2024–2025 claim that SoftBank was founded to realize ASI is later reinterpretation. It may express current identity, but it is not an original 1981 artifact. [contemporaneous record · strong] 2010 announcement 2024 presentation, pp. 20–22
• Forecast timing: the record moves from a roughly thirty-year singularity horizon in 2017–2019 to, in 2024, AGI within two or three years and ASI within ten, followed by 2025–2026 ten-year platform ambitions. This is forecast acceleration and reframing, not cumulative confirmation. [researcher inference · strong] Bloomberg singularity paragraphs SoftBank World 2024, 00:00:56–00:03:22

Contradictions and counterevidence that should travel with the playbook

• Category leadership versus valuation: the 2019 selection screen favors the No. 1 company; the 2022 postmortem says SoftBank may have overpaid while applying that screen. Category quality and investment quality are separate. [investor-stated · strong] Milken, 24:42–27:27 FY2022 Q1 transcript, A9
• No rescue versus WeWork follow-on capital: 2019–2020 materials repeat self-financing/no-rescue while SoftBank supplies substantial financing and intervention. The record never discloses the expected-return threshold that would distinguish rescue from new value-seeking investment. [researcher inference · strong] Q3 deck, slides 26–30 and 58–59
• Debt elimination versus strategic leverage: 2010 sets a path toward zero net debt; later Sprint, fund, asset-backed, margin-loan, and infrastructure strategies use financing as a strategic tool. The rational reading is regime dependence, not permanent adherence to a debt-zero principle. [contemporaneous record · strong] 2010 archive 2026 presentation, p. 66
• Self-renewing institution versus Son-centered succession: the 300-year cluster is supposed to transcend any product or leader, yet the 2019 conductor metaphor and 2026 life-plan extension keep Son central. Distributed execution does not resolve allocator succession. [researcher inference · strong] 2019 presentation, pp. 47–50 2026 presentation, pp. 6–8 and 62
• Minority diversification versus vertical concentration: the 2018 cluster holds many category leaders at minority stakes; the 2025–2026 stack concentrates capital and coordination around Arm, OpenAI, infrastructure, and robotics. Company count can rise while common-factor diversification falls. [researcher inference · strong] 2018 presentation, pp. 23–33 2025 presentation, pp. 15–28
• Defense versus offense: 2022 emphasizes cash, lower LTV, slower deployment, and strict selection; 2023 declares offense without publishing the new numerical risk budget. A discretionary confidence change may be legitimate, but it is not a reproducible capital-allocation rule. [researcher inference · strong] 2022 presentation, pp. 33–41 2023 presentation, pp. 3 and 32
• Forecast certainty versus changing horizon: a long singularity horizon compresses into near-term AGI and ten-year ASI. None of the public appearances defines the benchmark, baseline human, confidence interval, or failure date needed to distinguish learning from moving the goalposts. [researcher inference · strong] Bloomberg singularity paragraphs SoftBank World 2024, 00:00:56–00:03:22
• Cash rhetoric versus economic reality: Son's 2022 claim that cash alone does not change value is rhetorical. Currency, inflation, counterparty risk, and opportunity cost remain, and the answer itself acknowledges currency movement. [researcher inference · strong] FY2022 Q1 transcript, A16
• Regulatory caution versus maximum-speed deployment: Son supports safeguards and innovation acceleration, while agent and infrastructure plans scale rapidly without public detail on testing, liability, permissions, project downside, or kill switches. The policy position is directional, not an operational safety rule. [researcher inference · strong] TIME, Stargate section

Listen to these 5 first

1. 2012 Sprint announcement call and press conference. This is the best ex-ante decision record: full SEC-filed text, clear transaction state, Son's own market/operating logic, explicit risk-of-action versus risk-of-inaction reasoning, and enough specificity to test later outcomes. Source note SEC transcript [researcher inference · strong]
2. 2017 Bloomberg interview with David Rubenstein. This is the highest-density biographical and decision-style source, spanning founder selection, access-seeking, invention, Alibaba/Yahoo, leverage, Arm, Vision Fund scale, drawdown psychology, and a dated succession plan. Read every early fact as retrospective testimony. Source note Bloomberg interview [researcher inference · strong]
3. 2019 Milken Japan conversation. This gives the clearest single statement of the late-stage AI/No. 1 selection screen, the geographic-leader exception, the operating-adoption-versus-market-price distinction, and the Amazon regret that motivated greater capital scale. Source note Official video [researcher inference · strong]
4. 2022 FY2022 Q1 global call. This is the best skeptical companion to the expansion-era talks: a full, speaker-labeled postmortem with defense rules, overpayment admission, fund-leadership boundaries, runway caveats, asset-sale flexibility, and continued AI conviction. Source note Official transcript [researcher inference · strong]
5. 2024 NVIDIA AI Summit Japan with Jensen Huang. This is the best current operating-stack source and has creator-provided English captions. It shows founder recognition, infrastructure orchestration, sovereign AI, subsidized ecosystem formation, and NVIDIA regret with both speakers' voices preserved. Source note Official video [researcher inference · strong]

For the fastest corrective reading after those five, read the April 2020 Forbes interview: it contains the cleanest Son admission on WeWork price and founder trust, plus a contemporaneous counter-thesis to the power-law defense. [researcher inference · strong] Forbes, node 10 and nodes 44–52

Access and retrieval limits; highest-value next routes

The largest systematic gap is AGM speech and Q&A. Nineteen annual meetings from 2008 through 2026 are in the corpus, but their archive quality ranges from governance-only pages to full decks plus uncaptioned webcasts. Decks are valuable contemporaneous corporate records, yet they cannot show Son's exact language, improvisation, questions, dissent, or how management defended assumptions. Official speaker-labeled transcripts or caption files for the Brightcove AGM archive would have the highest cross-period value. [researcher inference · strong] 2008 AGM archive 2026 AGM archive

Two modern relationship appearances remain content-dark. The 2019 Son–Fink Sozo conversation and 2025 Son–Altman session have playable official media but no lawful text. Because BlackRock and OpenAI are strategically important counterparties, official transcripts or OpenAI-API transcription from the first-party media—with speaker diarization and manual verification—would be high-value. No local or non-OpenAI transcription was used here. [researcher inference · strong] Sozo official page Son–Altman official webcast

FII 8 and the 2010 vision need fuller first-party text. FII's single report page preserves only edited fragments of large forecasts; the Next 30-Year Vision digest omits most of a two-hour speech. Obtain the full FII Quest interview and official captions/transcript for Brightcove 6314097694112 before assigning omitted reasoning or treating decks as speech. [researcher inference · strong] FII 8 Executive Report, p. 61 Next 30-Year Vision digest

Automatic-caption keynotes need primary transcript replacement. SoftBank World 2024, Stargate, and SoftBank World 2025 are deeply analyzed only to the confidence permitted by platform-generated captions. The corpus therefore paraphrases, avoids direct quotation, and keeps forecast/agent claims at moderate confidence. Creator-provided subtitles would upgrade voice and wording without changing the work denominator. [researcher inference · strong] SoftBank World 2024, 00:00:56–00:03:22 Stargate official video SoftBank World 2025, 00:00:42–00:04:00

Transaction and performance closure requires non-talk evidence. The highest-value missing records are original Alibaba terms and cash flows; Sprint financing, synergy realization, and T-Mobile consideration; Arm acquisition/IPO/retained-share cash bridge by vehicle; complete Vision Fund vintage cash flows and fee/coupon treatment; WeWork follow-on underwriting; NVIDIA sale authorization and liquidity context; and Stargate/OpenAI binding obligations, project finance, utilization, and returns. These are not failures of speech retrieval—they are the evidence required to test the speeches. [researcher inference · strong] Sprint SEC transcript Arm transaction notice WeWork deck TIME profile

The retrieval standard remains conservative: alternate publisher pages and mirrors improve access but do not create new appearances; an indexed canonical article can be full access even when direct HTTP returns 406 if the lawful route exposes the entire text and voice boundaries; and unavailable content remains unavailable rather than reconstructed from adjacent decks, event titles, or plausible themes. [researcher inference · strong] TIME canonical article Sozo official page Next 30-Year Vision digest

Keep the mission fixed; cannibalize the method

• Rule. Define the enduring human or customer outcome, then treat every technology, product, business model, and organizational form as replaceable; fund the challenger before the incumbent economics force the change. [researcher inference · strong] Next 30-Year Vision, PDF pp. 30–32 and 89–99
• Evidence status. The mission/means distinction is investor-stated and strongly repeated; the governed self-cannibalization procedure is a researcher inference with strong behavioral support but no published security-level checklist. [investor-stated · strong] FY2011 report, printed pp. 8–13 / PDF pp. 10–15 [researcher inference · strong] Wired/Reuters, paragraphs 74–79
• Provenance and origin. The earliest accessible corpus evidence is the 1998–2000 pivot from PC software distribution and publishing toward Internet services; in 2000 Son explicitly urged cannibalization. The 2010 vision later formalized a fixed philosophy above changing technologies and business models. [investor-stated · strong] FY1998 report, printed pp. 2–5 / PDF pp. 4–7 Wired/Reuters, paragraphs 74–79 Next 30-Year Vision, PDF pp. 89–99
• Mechanism and reasoning. A stable mission lowers identity costs when the profit pool moves, while an explicit challenger prevents installed assets and current managers from vetoing the next architecture. It works only if the new unit has authority, talent, capital, customer access, and a measurable reason to exist; separation without those inputs merely duplicates cost. [researcher inference · strong] FY2000 report, printed pp. 5–6 / PDF pp. 38–39
• Historical and vehicle context. SoftBank could move among operating subsidiaries, corporate minority stakes, control acquisitions, and later managed funds. That flexibility is not available to every closed-end VC fund, and SBG's public shareholders, debt, LPs, regulators, and operating-company boards constrain which method can be replaced. [researcher inference · strong] investments vehicle map FY2022 report, printed pp. 91–93
• When to use. Use it when a durable customer need survives a technology discontinuity and the incumbent has a credible organizational conflict. Do not use it to rename unrelated diversification as mission continuity, or when the proposed challenger has no independent adoption, economics, or right to compete. [researcher inference · strong] writings synthesis Wired/Reuters, paragraphs 74–79
• How to apply it. (1) State the invariant outcome in observable terms; (2) map the incumbent profit pool and the new substitute; (3) appoint a separate accountable owner; (4) preallocate a bounded budget and customer/data rights; (5) set adoption, unit-economics, and shutdown gates; (6) decide when to migrate, sell, or close the incumbent. Ask which stakeholder loses, which board can approve the conflict, and what evidence would prove the new method is merely fashionable. [researcher inference · strong] Wired/Reuters, paragraphs 74–79
• Examples.
  • Documented application — Internet pivot · [documented behavior · strong]. SoftBank sold or de-emphasized older publishing/distribution assets while redirecting capital toward Internet gateways and services; this is a change of method, not proof every Internet investment worked. FY1998 report, printed pp. 2–5
  • Documented application — Yahoo! BB · [documented behavior · strong]. SBG built subsidized broadband despite losses because it treated access as the next delivery layer; the operating build is excluded from the external-investment denominator. FY2003 report, printed pp. 2–3 and 10 / PDF pp. 4–5 and 12 case lab
  • Researcher-applied illustration — Arm and the AI stack · [researcher inference · strong]. The move from mobile distribution to processor IP, models, infrastructure, agents, and robots fits method replacement, but the 2026 stack remains a mixture of holdings, commitments, partnerships, and pending assets rather than proven value capture. 2026 CEO message, “Securing the four corners”
• Evidence. Independent support comes from contemporaneous annual reports, the Reuters-reported 2000 remarks, the full 2010 vision, and later operating/deal records. Together they establish repeated adaptation and one explicit cannibalization statement; they do not establish a common return series or Son-only causation. [researcher inference · strong] writings chronology investments evidence boundary FY1998 report, printed pp. 2–5 Wired/Reuters, paragraphs 74–79 Next 30-Year Vision, PDF pp. 89–99 2026 CEO message, “Securing the four corners”
• Limits, failure modes, and exception ledger. A broad mission can become unfalsifiable, moral language can hide poor price or governance, and simultaneous bets can create correlated exposure rather than renewal. The 2010 claim of a self-renewing 300-year institution is also weakened by repeated succession reversals and current key-person dependence. [researcher inference · strong] 2010 vision, PDF pp. 105–123 FY2026 CEO message, “Updating my 50-year life plan”
• Evolution and contradictions. The method evolved from portfolio-company autonomy to operating control, then to a fund-led cluster and a concentrated AI stack. The unresolved contradiction is whether “technology flexibility” permits abandoning a thesis, or merely allows every new concentration to be narrated as the same mission. [researcher inference · strong] 2018 CEO message, “Driving the evolution” 2025 CEO message, “Going all in on ASI”
• Modern VC translation (as of 2026-08-22). What transfers is a thesis-level mission with explicit permission to obsolete portfolio support tools and sector labels. What fails is pretending a ten-year fund can imitate a listed conglomerate's duration or cross-subsidies. Signals are a durable customer problem, falling cost curve, incumbent conflict, and an independently improving challenger; misuse is mission drift. A grounded current example is SBG's four-corner AI map: use it as a dependency inventory, but require product adoption, security, unit economics, financing, and realized value at each layer. [researcher inference · strong] 2026 AGM deck, PDF pp. 27–38

Backcast from a future state, but make forecasts falsifiable

• Rule. Describe the future state, work backward to dated operating and financing milestones, and attach probabilities, base/downside cases, falsifiers, and an abandonment review before committing capital. [researcher inference · strong] FY2012 report, printed p. 010 / PDF p. 12
• Evidence status. Large goals and backward planning are investor-stated with strong recurrence. [investor-stated · strong] 2010 vision, PDF pp. 40–75 2026 CEO message, “Aiming for NAV” Probability trees and kill gates are a researcher-applied correction because the public record repeatedly omits them. [researcher inference · strong] writings contradiction ledger FII 8 report, p. 61
• Provenance and origin. The 2010 Next 30-Year Vision supplies the first accessible full architecture: philosophy, future destination, technology curves, organizational design, and a thirty-year horizon. The 2012 message then turns the method into a dated ¥1 trillion operating-income target, asserted without a disclosed scenario bridge. [investor-stated · strong] 2010 vision, PDF pp. 1, 30–32 and 125–127 FY2012 report, printed p. 010
• Mechanism and reasoning. A vivid destination can coordinate talent and capital across long projects, while backward milestones reveal dependencies early. Falsifiers prevent the target from becoming its own evidence: each technological, demand, financing, regulatory, and execution layer must carry an independent probability and observable stop condition. [researcher inference · strong] talks framework 2010 vision, PDF pp. 40–75
• Historical and vehicle context. Public target-setting spanned an operating telecom group, a strategic holding company, managed funds, and current infrastructure consortia. The same headline ambition has different capital-call, governance, liquidity, and fiduciary consequences in each vehicle, so the backcast must begin with the legal payer and decision owner. [researcher inference · strong] investments vehicle map FY2022 report, pp. 91–93
• When to use. Use it for long-cycle infrastructure, platform transitions, market creation, and organizational renewal where current screens miss future bottlenecks. Do not use it for a one-way commitment when intermediate evidence cannot arrive before most capital is spent, or where the destination is too vague to falsify. [researcher inference · strong] 2025 CEO message, OpenAI conditional commitment footnote 2
• How to apply it. (1) Define the end-state customer behavior; (2) decompose technology, adoption, unit economics, capital, regulation, and exit; (3) assign base/upside/downside probabilities without a shared conviction multiplier; (4) date the next observable; (5) prewrite capital releases and kill criteria; (6) refresh the tree rather than moving the deadline. Ask which assumption can fail while the secular forecast remains right. [researcher inference · strong] 2026 AGM deck, PDF pp. 27–38
• Examples.
  • Documented application — fiscal 2016 operating-income target · [investor-stated · moderate]. Son publicly backcast to ¥1 trillion, but the one-page record provides no base rate, downside buffer, or revision trigger; it shows mobilization more clearly than forecast method. FY2012 report, printed p. 010 / PDF p. 12
  • Researcher-applied illustration — Sprint · [researcher inference · strong]. The acquisition materials specified network and operating synergies, yet lacked a public probability-weighted valuation, capex downside, and exit tree; later difficulty shows why a prior Vodafone outcome was not a base rate. SEC-filed Sprint transcript, lines 701–814 and 1794–1928 case lab
  • Boundary case — ASI and ¥1 quadrillion NAV · [investor-stated · moderate]. The 2026 targets orient the organization but lack a disclosed annual cash-flow, dilution, leverage, or probability bridge; they are ambitions, not investment guidance. 2026 CEO message, “Aiming for NAV”
• Evidence. The model is supported by independent 2010, 2012, Sprint, 2024, and 2026 records plus the cross-talk forecast audit. Those sources establish repeated target use and repeated missing scenario controls; they do not show that the same internal backcast governed every decision. [researcher inference · strong] writings contradiction ledger Next 30-Year Vision, PDF pp. 40–75 FY2012 report, printed p. 010 Sprint SEC transcript, lines 701–814 FII 8 report, p. 61 2026 CEO message, “Aiming for NAV”
• Limits, failure modes, and exception ledger. Targets can mobilize without predicting, deadlines can move, correlated assumptions can be multiplied by rhetoric rather than arithmetic, and goal certainty can suppress dissent. A correct long-run direction can still yield a poor security because price, dilution, governance, financing, or duration is wrong. [researcher inference · strong] FY2022 Q1 transcript, A9
• Evolution and contradictions. Early targets emphasized operating income and debt; later targets emphasize NAV, ASI capability, GDP impact, compute, and infrastructure. The contradiction is a move from nearer operating milestones toward increasingly interdependent forecasts without a more visible probability discipline. [researcher inference · strong] FY2009 report, printed pp. 17–20 TIME 2026, paragraph beginning “Not that Son agrees”
• Modern VC translation (as of 2026-08-22). What transfers is future-back roadmapping; what fails is using a charismatic forecast as valuation evidence. Today's AI models, power projects, and regulation change faster than telecom buildouts, so signals must include falling cost per useful task, contracted demand, utilization, interconnection/power milestones, security, and financing closes. Misuse is stacking optimistic layer probabilities. SBG's current OpenAI/Stargate plan is the grounded test: treat each funded tranche, partner obligation, site, power contract, and deployment milestone as a separate branch. [researcher inference · strong] OpenAI follow-on release, sections 1–5

Name the regime before the company

• Rule. Begin with the structural shift in capability, cost, behavior, or regulation; define how to observe adoption and then search for companies, rather than projecting a favored company into a fashionable theme. [researcher inference · strong] FY1998 report, printed pp. 2–5
• Evidence status. Son repeatedly states long-horizon information-revolution and AI regime theses. [investor-stated · strong] FY1999 report, printed pp. 2–5 / PDF pp. 3–5 2021 CEO message, “SoftBank Group = A capital provider” 2026 CEO message, “Securing the four corners” Converting that habit into a company-selection sequence is a researcher inference with strong cross-era support. [researcher inference · strong] portfolio forensics FY2025 SVF1 table, p. 78
• Provenance and origin. The earliest full written corpus frames Internet services as the next economic layer before listing gateway, commerce, finance, and distribution assets. Later materials apply the same order to broadband, mobile Internet, IoT, AI, and ASI. [investor-stated · strong] FY1999 report, printed pp. 2–5 / PDF pp. 3–5 FY2001 report, printed pp. 2–5 / PDF pp. 63–66 2018 CEO message, “Driving the evolution” 2026 CEO message, “Securing the four corners”
• Mechanism and reasoning. Regime-first work expands the opportunity set beyond current sector labels and focuses diligence on secular adoption. Its causal weakness is that technological importance says nothing by itself about industry structure, value capture, price, ownership, survival, or the winning company. [researcher inference · strong] talks opening synthesis FY1999 report, printed pp. 2–5
• Historical and vehicle context. The method emerged when public Internet adoption was visible but monetization was immature, then migrated into capital-intensive telecom and giant managed funds. Later-stage checks and current infrastructure commitments are less reversible than early minority optionality, so regime confidence cannot justify the same position size across vehicles. [researcher inference · strong] FY2022 report, printed pp. 22–25
• When to use. Use it in genuine technology or regulatory discontinuities where independent adoption evidence exists and the investable landscape is still forming. Do not use it when the “regime” is defined only by capital inflows, valuation, management forecasts, or a single company's promotional metric. [researcher inference · strong] Paytm lab FY2025 SVF1 table, p. 78
• How to apply it. Write a regime memo before the company memo: changed primitive, user behavior, cost curve, adoption measure, incumbent response, regulatory constraint, value-chain map, alternative architectures, and disconfirming evidence. Only then score companies on durable value capture, price, rights, financing, and team. [researcher inference · strong] writings decision playbook FY1999 report, printed pp. 2–5
• Examples.
  • Documented application — Internet gateways · [documented behavior · strong]. SoftBank mapped traffic, portals, commerce, and financial services before building the Yahoo and Yahoo Japan positions; the record also contains failed Internet bets, so the regime did not select every winner. FY1999 report, printed pp. 2–5 Forbes, paragraphs beginning “All told”
  • Documented application — mobile Internet · [investor-stated · strong]. Sprint materials start from rising mobile-data demand and then propose network, procurement, device, and operating transfer. Sprint SEC transcript, lines 701–787 Cross-period outcome boundary · [researcher inference · strong]. Son's later admission that the turnaround was harder than expected separates regime accuracy from deal quality. FY2016 report, printed pp. 9–10
  • Boundary case — late-stage category leaders · [researcher inference · strong]. Uber, Grab, DiDi, and Paytm occupied real adoption categories yet produced reported gross results of 1.2x, 0.8x, 0.3x, and 0.7x at the stated 2026 perimeter. A valid regime did not rescue entry price or regulation. FY2025 results, SVF1 table, PDF p. 78
• Evidence. Independent support spans signed annual reports, SEC-filed Sprint remarks, fund-return tables, and the frozen writings/talks syntheses. The evidence establishes a repeated search orientation and its counterexamples, not a proprietary regime-prediction accuracy rate. [researcher inference · strong] writings synthesis portfolio forensics FY1999 report, printed pp. 2–5 Sprint SEC transcript, lines 701–787 FY2025 SVF1 table, p. 78
• Limits, failure modes, and exception ledger. Themes can be too broad to falsify, adoption can coexist with commodity economics, adjacent companies can share hidden factor exposure, and a regime label can suppress price discipline. WeWork's workspace/community narrative and the mobility basket show how a plausible category can sit atop weak governance, subsidies, or regulatory risk. [researcher inference · strong] WeWork S-1, pp. 49–51
• Evolution and contradictions. The sequence moved from many Internet categories to mobile control, then hundreds of fund companies, and now a concentrated AI stack. Breadth and concentration are both called regime exposure; without vehicle-level position and correlation limits, they are contradictory portfolio architectures. [researcher inference · strong] FY2019 report, portfolio map, printed pp. 87–89 2026 CEO message
• Modern VC translation (as of 2026-08-22). What transfers is separating a technical/economic map from company enthusiasm. Changed conditions include faster AI model cycles, hyperscale compute/power needs, data rights, export controls, and uncertain value capture. Signals are independent retention, falling cost per useful task, contracted demand, durable margins, and switching costs; misuse is calling every AI company a regime winner. The current four-corner plan is useful as a research map, not as proof that every layer or security is investable. [researcher inference · strong] 2026 AGM deck, PDF pp. 27–38

Own or influence the bottleneck where value can accrue

• Rule. Identify the scarce gateway, distribution channel, infrastructure layer, standard, or coordination point in a new system; invest only if durable control or influence can convert that scarcity into cash economics after capital and regulation. [researcher inference · strong] FY1999 report, printed pp. 2–5
• Evidence status. Bottleneck search is a researcher-reconstructed rule with strong repeated behavior across portals, broadband, mobile, processor IP, and AI infrastructure; no source publishes this exact unified checklist. [researcher inference · strong] FY1999 report, printed pp. 2–5 FY2001 report, printed pp. 2–5 2006 Vodafone acquisition release, sections 1–5 2016 Arm statement, paragraphs 2–7 2026 CEO message, “Securing the four corners”
• Provenance and origin. Yahoo appears in the late-1990s record as an Internet gateway and aggregation point. [contemporaneous record · strong] FY1999 report, printed pp. 2–5 / PDF pp. 3–5 The 2001 message then names broadband infrastructure as the constraint on Internet services. [investor-stated · strong] FY2001 report, printed pp. 2–5 / PDF pp. 63–66 Those precede the later Vodafone customer-access and Arm processor-IP control rationales. [researcher inference · strong] 2006 Vodafone acquisition release, sections 1–5 2016 Arm strategic rationale, paragraphs 2–7
• Mechanism and reasoning. A scarce layer can tax or accelerate many downstream applications, collect data, create switching costs, and preserve option value as products change. The mechanism fails when the layer commoditizes, customers multi-home, standards remain neutral, capital cost absorbs rents, or ownership damages ecosystem trust. [researcher inference · strong] Arm 2023 prospectus, business and risk sections
• Historical and vehicle context. Bottlenecks were pursued through minority corporate stakes, organic infrastructure, leveraged control acquisitions, in-kind fund transfers, and current preferred equity/project commitments. The value-capture test must therefore use the relevant entity's rights, financing, duration, and cash-flow perimeter. [researcher inference · strong] investments comprehensive ledger FY2022 report, pp. 91–103
• When to use. Use it when demand depends on a genuinely scarce interface or asset and the investor can specify why rents persist. Do not use it when “foundational” means merely technically important, when open standards shift value elsewhere, or when the investor owns an adjacent asset but no contractual right to the bottleneck. [researcher inference · strong] Arm lab 2016 Arm acquisition release
• How to apply it. Map the value chain; name the constrained resource; test substitutes, multi-homing, customer bargaining, regulation, capacity expansion, and capital intensity; specify required ownership/votes/contracts; model unit economics and reinvestment; set milestones for durable rent rather than gross traffic. [researcher inference · strong] writings decision playbook CCA transcript, pp. 2–6
• Examples.
  • Documented application — Yahoo and Yahoo Japan · [documented behavior · strong]. Gateway traffic plus local rights created distribution optionality, but the full cash ledger and original security terms remain incomplete. Yahoo 1999 10-K, Notes 6–7 Yahoo Japan lab
  • Documented application — broadband and mobile access · [documented behavior · strong]. Yahoo! BB and Vodafone Japan placed SBG at customer-access layers; operating milestones are documented, but acquisition/build returns are not fully reconstructed. FY2007 report, printed pp. 13–20
  • Boundary case — Arm · [researcher inference · strong]. Arm's instruction-set ecosystem is strategically central, yet ownership must preserve licensee neutrality, and SBG's reported ¥24.6T/10x perimeter depends heavily on retained/unrealized value. Arm prospectus 2024 CEO message, Arm return discussion
• Evidence. Independent corporate reports, SEC filings, transaction releases, and case labs support repeated bottleneck ownership and its economic boundaries. The record does not establish a denominator of attempted bottleneck bets or that bottleneck centrality caused the reported outcomes. [researcher inference · strong] denominator boundary FY1999 report, printed pp. 2–5 Sprint SEC transcript, lines 701–814 2016 Arm statement, paragraphs 2–7 FY2022 report, pp. 91–103
• Limits, failure modes, and exception ledger. A central layer can have weak margins, require perpetual capex, invite regulation, or lose to an open substitute. Sprint and OneWeb show that infrastructure relevance does not solve operating or financing dependence; the current AI stack could couple power, compute, model, and project-finance risk rather than diversify it. [researcher inference · strong] OneWeb outcome, financial report printed p. 14 / PDF p. 18
• Evolution and contradictions. The rule moved from influencing gateways through minority stakes to owning networks and Arm, then to partnering around OpenAI/Stargate. “Control point” now includes assets SBG does not control, so contractual access, governance, exclusivity, and partner economics matter more than narrative adjacency. [researcher inference · strong] FY1999 report, printed pp. 2–5 2016 Arm acquisition release 2026 report, printed pp. 23–25
• Modern VC translation (as of 2026-08-22). What transfers is dependency mapping; what fails is assuming the most expensive AI layer captures the most value. Signals are constrained supply, contractual demand, switching costs, pricing power, neutral ecosystem growth, and returns above reinvestment needs. Misuse is paying any price for technical importance. The current Arm/OpenAI/Stargate stack should be evaluated layer by layer: IP royalties, model access, compute utilization, power/offtake, project finance, and governance cannot share one valuation shortcut. [researcher inference · strong] 2026 AGM deck, PDF pp. 27–38 and 66

Use the time machine through local operators, not copy-paste

• Rule. Transfer a proven model into a less-mature market only through operators who understand local demand, regulation, distribution, and culture; preserve their autonomy while contracting for the information and influence the investment case requires. [researcher inference · strong] FY1999 report, printed pp. 3–5 / PDF pp. 4–5
• Evidence status. The “time machine” and local-partner logic are investor-stated and well documented in the early Internet record. [investor-stated · strong] Forbes, Yahoo Japan and joint-venture passages The full rights-and-localization checklist is researcher inference because original agreements and governance maps are incomplete. [researcher inference · strong] Yahoo 1999 10-K, Note 6
• Provenance and origin. The earliest accessible use is the 1995–1999 Yahoo/Yahoo Japan sequence: take exposure to a leading U.S. gateway, then form a 60/40 Japanese joint venture rather than export the service unchanged. [contemporaneous record · strong] Yahoo 1999 10-K, Note 6 and Yahoo Japan discussion
• Mechanism and reasoning. A model already validated elsewhere reduces product/category uncertainty, while a local operator supplies distribution, regulation, language, and adaptation. It fails when market maturity, payments, data rules, labor, consumer behavior, or competition differ enough that the foreign analogy hides the real underwriting problem. [researcher inference · strong] FY2015 report, printed pp. 6–7 / PDF pp. 8–9
• Historical and vehicle context. Early localization used corporate capital and joint ventures; the Vision Fund later held large minority stakes through Cayman and other entities across regions, sometimes without matched votes. Cross-border data, listing, national-security, and regulated-affiliate risk became more binding than in the 1990s portal cases. [researcher inference · strong] DiDi 2021 20-F, principal-shareholder table Paytm RHP, printed pp. 223 and 261
• When to use. Use it when the core behavior is independently proven, local adoption lags for identifiable reasons, and a named operator has incentives and authority to adapt it. Do not use it where regulation blocks the ownership/data model, subsidies create apparent adoption, or “local leader” is only a fundraising rank. [researcher inference · strong] failure/ordinary labs Alibaba prospectus, “Our History”
• How to apply it. Compare origin and target markets across infrastructure, income, payments, regulation, data, labor, distribution, unit economics, and exit markets; identify the local decision maker; define territory/IP/data rights, board/information rights, pro-rata capital, milestones, and a country-specific kill case. [researcher inference · strong] Yahoo Japan IC worksheet Alibaba prospectus, “Our History”
• Examples.
  • Documented application — Yahoo Japan · [documented behavior · strong]. The localized JV joined Yahoo's brand/technology to SoftBank's Japanese operating capacity and later became a separately managed public platform. Yahoo 1999 10-K
  • Documented application — E*Trade Japan and other early JVs · [contemporaneous record · moderate]. Forbes reports a repeated alliance/localization program, but edited reporting and incomplete security terms prevent treating every venture as a proven application or success. Forbes, E*Trade and alliance passages
  • Counterexample — Sprint · [researcher inference · strong]. SBG attempted to transfer Japanese network, procurement, and operating experience into the United States; the harder-than-expected turnaround shows that a home-market playbook does not travel without a jurisdiction-specific base rate and operator plan. FY2016 report, printed pp. 9–10 / PDF pp. 11–12
• Evidence. Independent SEC filings, early corporate reports, Forbes's contemporaneous profile, and later deal records establish repeated localization and its failures. They do not reveal a complete country-decision denominator, uniform security terms, or the share of outcomes created by local teams rather than SoftBank. [researcher inference · strong] people map Forbes 1999, paragraphs beginning “Masayoshi Son” Yahoo 1999 10-K, Note 6 Alibaba prospectus, “Our History” Sprint SEC transcript, lines 701–814
• Limits, failure modes, and exception ledger. Copy-paste can import the wrong maturity curve; autonomy without rights leaves the investor unable to intervene; control can displace the local knowledge the strategy needs. DiDi and Paytm add regulator/entity risks, while Sprint adds financing and operating-transfer risk. [researcher inference · strong] RBI Paytm action, items 1–5
• Evolution and contradictions. Early evidence favored small or meaningful minority partnerships and local autonomy; Vodafone and Sprint used control, and the current AI program again mixes foreign minority equity, subsidiaries, and consortia. The record offers no universal choice rule, so ownership must be derived from the work required rather than from the “time machine” label. [researcher inference · strong] writings contradiction ledger Forbes 1999, Yahoo Japan passages 2006 Vodafone acquisition release Sprint SEC transcript, lines 701–814 2026 report, printed pp. 23–25
• Modern VC translation (as of 2026-08-22). What transfers is market-by-market localization; what fails is assuming software distribution erases national boundaries. Signals are local retention, contribution after incentives, regulatory clearance, data portability, enforceable rights, and an accountable local leader. Misuse is treating global brand or a 20%–30% stake as control. A current AI example would require country-specific model, data, safety, compute, and distribution rights before funding a local frontier-model distributor; this is an analyst-created application, not a documented Son deal rule. [researcher inference · moderate] TIME 2026, paragraph beginning “Son's break-up with China”

Back ambitious category leaders with a separate price-and-governance key

• Rule. Exceptional founder ambition and category leadership may open the diligence gate, but approval requires an independent case for valuation, security, votes, board rights, related-party controls, runway, and disconfirming evidence. [researcher inference · strong] FY2022 Q1 transcript, A9
• Evidence status. Founder intensity and No. 1 selection are investor-stated. [investor-stated · strong] Tokyo Forum, 03:14–16:18 The two-key approval is a researcher reconstruction strongly supported by WeWork and ordinary outcomes; the record does not disclose a stable Son-authored founder scorecard. [researcher inference · strong] Forbes pressure interview, node 10 FY2025 SVF1 table, p. 78
• Provenance and origin. The 1999 profile documents Son's Yahoo and U.S. venture selections, while Alibaba's prospectus documents SoftBank's 2000 investment and 2005 transaction but not an original founder scorecard. [contemporaneous record · strong] Forbes 1999, Yahoo/STV passages Alibaba prospectus, “Our History” and “SoftBank” Son's and Jack Ma's founder-selection account is retrospective, not decision-near proof. [retrospective witness · moderate] Tokyo Forum, 03:14–16:18 The 2018 holding-company message later formalizes a cluster of autonomous No. 1 companies. [investor-stated · strong] 2018 CEO message, “Creating a unique synergy group structure”
• Mechanism and reasoning. Ambitious founders can recruit, learn, and compound faster in increasing-return markets. The same charisma can weaken price discipline and oversight, so founder evidence must increase—not reduce—reference depth, downside work, and governance protection. [researcher inference · strong] talks framework Forbes 1999, paragraphs beginning “All told”
• Historical and vehicle context. Early corporate stakes, control acquisitions, and SVF minorities assigned different rights. Founder supervotes, fund-manager authority, regulated subsidiaries, and outside LP capital mean economic ownership is not governance and Son's relationship is not the legal security. [researcher inference · strong] Coupang 424B4, principal shareholders, printed pp. 178–179
• When to use. Use it in markets with increasing returns where leadership is independently evidenced and founder capability is relevant to the next stage. Do not use it when rank depends on subsidy, an unaudited forecast, temporary valuation, or when governance risk cannot be priced or controlled. [researcher inference · strong] Uber/Grab/Paytm labs FY2019 Q2 presentation, pp. 6–19
• How to apply it. Run two separate memos: founder/category evidence and security/governance. Require customer and former-colleague references beyond admirers, cohort/unit-economics proof, cap table and votes, board/audit rights, related-party map, valuation cases, reserve cap, and written kill criteria. A pass on one memo cannot offset failure on the other. [researcher inference · strong] WeWork S-1, pp. 49–51
• Examples.
  • Researcher-applied illustration — Alibaba · [researcher inference · strong]. The result supports early founder sponsorship and long duration, but the public record lacks the original founder scorecard and complete security cash flows; Alibaba's team created the business. Alibaba 2014 prospectus, history and principal shareholders
  • Documented boundary — Coupang · [contemporaneous record · strong]. A large reported SVF1 gain coexisted with Bom Kim's supervoting control; operating success and the investor's governance power were separate. Coupang 424B4, pp. 178–179 FY2025 results, p. 78
  • Counterexample — WeWork · [investor-stated · strong]. Son admitted paying too high a valuation and believing too much in the entrepreneur; the decision-near S-1 had already disclosed capital dependence, leases, controls, and related parties. Forbes, node 10 WeWork S-1, pp. 49–51
• Evidence. Direct Son remarks, company filings, corporate reports, and reported fund outcomes provide independent support across wins, ordinary rows, and failure. They show that leadership/founder belief recurred and that price/governance mattered; they do not yield a founder-heuristic hit rate. [researcher inference · strong] decision-learning ranking Tokyo Forum, 03:14–16:18 WeWork S-1, pp. 49–51 Forbes pressure interview, node 10 FY2025 SVF1 table, p. 78
• Limits, failure modes, and exception ledger. False positives, founder syndrome, supervotes, related parties, subsidy-led rank, and survivor selection are the central risks. Uber, Grab, DiDi, and Paytm show that a category leader can be a middling or losing security; no evidence supports treating charisma as proprietary diligence. [researcher inference · strong] FY2025 SVF1 table, p. 78
• Evolution and contradictions. The rhetoric stayed founder-friendly while the public lessons added price after WeWork and defense in 2022. Yet current OpenAI concentration again combines founder relationship, category leadership, and large commitments, testing whether the separate governance/price key is institutionalized. [researcher inference · strong] Forbes pressure interview, node 10 FY2022 Q1 transcript, A16 and A20–A23 2026 OpenAI follow-on release, sections 1–5
• Modern VC translation (as of 2026-08-22). What transfers is testing founder learning velocity and market leadership; what fails is equating media rank or frontier-AI access with investability. Signals are retained users, normalized unit economics, reference consistency, governance maturity, transparent conflicts, and price-supported return. Misuse is lowering standards for a famous founder. OpenAI is the grounded current test: preferred terms, ownership, governance, syndication, tranches, and concentration must stand independently of model leadership. [researcher inference · strong] FY2026 report, printed pp. 23–25

Stage conviction; re-underwrite every follow-on and precommit the stop

• Rule. Buy optionality with a bounded first commitment, release reserves only when prewritten evidence arrives, and treat every follow-on, rescue, or refusal as a fresh IC with a maximum total exposure and stop rule. [researcher inference · strong] Forbes, Yahoo/STV passages
• Evidence status. Early test-and-top-up behavior is contemporaneously reported. [contemporaneous record · strong] Forbes 1999, paragraphs beginning “Since setting up” The universal re-IC/stop discipline is a researcher inference strongly supported by later failures and the Coupang exception; staging was not universal across Son's record. [researcher inference · strong] portfolio forensics, follow-ons WeWork S-1/A, pp. 118–119 FY2025 Coupang row, p. 78
• Provenance and origin. The 1999 Forbes profile describes small initial Internet positions and subsequent top-ups, while also reporting a $230M STV pool spread across 55 companies. [contemporaneous record · moderate] Forbes, paragraphs beginning “In 1995 Son discovered” and “In a frenetic first year” Later very large checks and support packages make the early pattern a useful origin, not a timeless doctrine. [researcher inference · strong] FY2019 Q2 WeWork presentation, pp. 6–29 OneWeb loss disclosure, printed p. 14 / PDF p. 18
• Mechanism and reasoning. Staging buys information and limits initial loss; a fresh IC prevents sunk cost, reputation, and a rising prior mark from laundering escalation into conviction. It works only if milestones arrive before capital is exhausted and the investor can actually refuse the next tranche. [researcher inference · strong] OneWeb financial report, Note 6(2), printed p. 14 / PDF p. 18
• Historical and vehicle context. Corporate follow-ons, fund reserves, warrants, convertibles, loans, and rescue finance create different seniority and approval paths. Outside LPs, SBG balance-sheet support, and cross-vehicle exposure must be aggregated before a manager calls a follow-on “small.” [researcher inference · strong] WeWork S-1/A, printed pp. 118–119
• When to use. Use it whenever uncertainty can resolve in stages and the investor retains contractual discretion. Do not pretend a huge first check is staged because accounting closes occur in tranches, or apply “never double down” when the company is meeting evidence and the security remains attractive. [researcher inference · strong] Coupang lab FY2025 SVF1 table, p. 78
• How to apply it. Before entry, write first-check purpose, next milestone, reserve, valuation reset, outside-financing proof, governance escalation, maximum look-through exposure, and kill conditions. At every follow-on, rebuild the cap table, security stack, runway, cohort economics, conflicts, and counterfactual use of capital; record dissent and the reason to invest, hold, sell, or stop. [researcher inference · strong] operational rules FY2019 Q2 presentation, pp. 6–19
• Examples.
  • Documented application — early Yahoo/STV · [contemporaneous record · moderate]. The early profile supports test-and-top-up behavior across a broad pool, but team attribution and complete losses prevent calling it a Son-only staged system. Forbes, STV and Yahoo passages
  • Counterexample — WeWork · [contemporaneous record · strong]. Convertibles, warrants, preferreds, JV instruments, and support capital accumulated while governance and lease/cash risks were visible; the $6.71B later loss perimeter is fund-accounting, not total support cash. WeWork S-1/A, pp. 118–119 FY2025 report, printed p. 91 / PDF p. 93
  • Counterexample and exception pair — OneWeb/Coupang · [researcher inference · strong]. OneWeb escalated across equity, loans, warrants, and shareholder loans before a financing failure; Coupang shows why automatic abandonment would also be wrong because retained exposure preserved a large reported gross gain. OneWeb report, printed p. 14 / PDF p. 18 FY2025 Coupang row, p. 78
• Evidence. Independent contemporary journalism, issuer filings, corporate loss disclosures, and standardized fund rows support both early staging and later escalation. The case set supports conditional re-underwriting, not a claim that published post-2019 controls caused better outcomes. [researcher inference · strong] failure packet Forbes 1999, STV and Yahoo passages WeWork S-1/A, pp. 118–119 OneWeb loss disclosure, printed p. 14 / PDF p. 18 FY2025 SVF1 table, p. 78
• Limits, failure modes, and exception ledger. Milestones can be gamed, insiders may suppress bad news, reserves create implicit rescue expectations, and pro rata rights can become psychological obligations. A small first check is not disciplined if the fund has already promised unlimited support; a refusal can destroy option value if the evidence still supports the company. [researcher inference · strong] FY2019 Q2 deck, slides 20–29
• Evolution and contradictions. Early optionality gave way to front-loaded scale; WeWork and the 2022 drawdown produced monitoring, runway, slower-deployment, and liquidity language. The renewed AI offense tests whether stop and concentration limits govern giant conditional commitments as well as ordinary portfolio follow-ons. [researcher inference · strong] FY2021 results transcript, A18, A20–A21 and A26 OpenAI follow-on release
• Modern VC translation (as of 2026-08-22). What transfers is evidence-gated reserves; what fails is cosmetic tranche language when commitments are economically irreversible. Signals are audited milestone attainment, third-party funding, improving unit economics, governance compliance, and sufficient remaining runway. Misuse is rescuing to protect a mark or abandoning every down round. For current frontier-model deals, release capital against model capability, paid retention, safety, compute efficiency, governance, and syndication—not a calendar alone. [researcher inference · strong] 2026 OpenAI release, sections 1–5

Match ownership and control to the work required

• Rule. Choose minority, largest-shareholder, joint-venture, or control ownership from the exact information, veto, board, operating, and financing work the thesis requires; never infer governance from economic percentage alone. [researcher inference · strong] FY2019 Q2 WeWork deck, accounting slides 4–5
• Evidence status. The autonomy and strategic-synergy architecture is investor-stated. [investor-stated · strong] FY1999 report, printed pp. 3–5 2018 CEO message The rights-to-work decision rule is a researcher inference with strong evidence across minority and control cases. [researcher inference · strong] 2006 Vodafone acquisition release 2016 Arm acquisition release
• Provenance and origin. The 1999 record combines meaningful minority stakes, local JVs, operating autonomy, and centralized capital allocation. [contemporaneous record · strong] FY1999 report, printed pp. 3–5 Later Vodafone, Sprint, and Arm acquisitions document the distinct financing, operating, regulatory, and governance obligations created by control. [documented behavior · strong] 2006 Vodafone acquisition release, sections 1–5 2012 Sprint SEC-filed remarks, lines 701–814 2016 Arm acquisition release
• Mechanism and reasoning. Rights should match the intervention: local founders need room to operate, while turnarounds may require management, budget, and integration authority. Too few rights produce influence theater; too much control adds debt, regulation, fiduciary, integration, and execution risk. [researcher inference · strong] Vodafone release, sections 1–5
• Historical and vehicle context. SBG corporate control, opco governance, SVF manager rights, founder supervotes, regulated affiliates, and LP capital are not interchangeable. The legal holder and manager—not Son's public relationship—determine most enforceable rights. [researcher inference · strong] vehicles and people map FY2022 report, pp. 91–103
• When to use. Use it before every investment and again after dilution, IPO, merger, affiliate formation, or rescue. Do not seek control merely for accounting consolidation, or accept a large economic stake when the thesis requires rights that the security does not provide. [researcher inference · strong] Paytm RHP, pp. 223 and 261
• How to apply it. List the work—information, hiring, capital allocation, related-party approval, audit, regulation, product integration, follow-on, or exit—then map ownership, votes, board seats, reserved matters, manager authority, affiliate boundaries, duration, and enforcement. Price any missing right and name the walk-away gap. [researcher inference · strong] Coupang 424B4, pp. 178–179
• Examples.
  • Documented application — Yahoo Japan JV · [contemporaneous record · strong]. A 60/40 JV joined local operation and partner technology; the public record still omits the original reserved-matters map. Yahoo 1999 10-K
  • Documented application — Vodafone Japan control · [documented behavior · strong]. The 97.7% acquisition enabled operating intervention but added layered financing, lender, preferred-holder, and board obligations. 2006 transaction release
  • Counterexample — WeWork, DiDi, and Coupang · [researcher inference · strong]. Large economics did not equal control: WeWork's proposed roughly 80% economics lacked majority votes/accounting control; DiDi's 20.0% economics carried 10.7% votes; Coupang's founder retained supervotes. WeWork deck, slides 4–5 DiDi 20-F Coupang 424B4
• Evidence. Corporate filings, prospectuses, transaction releases, and the people/vehicle maps independently support the ownership-rights distinction. They do not reveal every side letter, IC vote, or informal influence channel, so unknown remains the correct default. [researcher inference · strong] case-specific decision-role map Alibaba prospectus, “Our History” and “SoftBank” 2006 Vodafone acquisition release Sprint SEC transcript, lines 701–814 2016 Arm acquisition release
• Limits, failure modes, and exception ledger. Rights may be legally weak, expire at IPO or dilution thresholds, conflict with regulation, or be unusable without board support. Control can still fail operationally; minority autonomy can still create excellent outcomes. This model chooses a governance fit, not a guarantee. [researcher inference · strong] Sprint/T-Mobile lab Sprint SEC transcript, lines 701–814
• Evolution and contradictions. Son's early minority-alliance rhetoric coexists with direct control and intervention, then returns to large minority/fund positions and consortium leadership. The tension is resolvable only case by case; the public writings never state a stable ownership algorithm. [researcher inference · moderate] writings contradiction ledger FY1999 report, printed pp. 3–5 2006 Vodafone acquisition release FY2017 report, Vision Fund governance, printed pp. 70–73 2026 report, printed pp. 23–25
• Modern VC translation (as of 2026-08-22). What transfers is designing rights from intended work; what fails is relying on reputation or headline ownership. Signals are enforceable information, independent board/audit rights, founder alignment, regulatory separation, and rights that survive financing. Misuse is demanding control without operating competence. In current frontier AI, a minority investor must distinguish model access, equity rights, compute/offtake agreements, safety governance, and project-company control rather than treat partnership as ownership. [researcher inference · strong] FY2026 report, pp. 23–25 and 80–83

Write the operating-transfer plan before a control acquisition

• Rule. Before buying control, name the inherited weaknesses, transferable capabilities, accountable operators, jurisdiction differences, financing envelope, milestones, and conditions for abandoning or selling the turnaround. [researcher inference · strong] FY2007 report, printed pp. 13–20
• Evidence status. Direct operating intervention is documented behavior. [documented behavior · strong] FY2003 Yahoo! BB report, printed pp. 2–3 and 10 FY2007 Vodafone report, printed pp. 13–20 FY2016 report, printed pp. 42–47 The pre-acquisition transfer-plan checklist is researcher inference with strong comparative support from Vodafone, Sprint, Yahoo! BB, and Japan Telecom. [researcher inference · strong] control-acquisition labs Sprint SEC transcript, lines 1065–1264
• Provenance and origin. Yahoo! BB first exposed the operating-build pattern, but Vodafone Japan provides the earliest accessible control-acquisition sequence with financing layers and later network, product, pricing, and distribution interventions. [documented behavior · strong] 2006 acquisition release, sections 1–5 FY2007 report, pp. 13–20
• Mechanism and reasoning. Control creates authority to change capital allocation, management, product, pricing, and infrastructure, but value arises only if the buyer owns a relevant capability and can transfer it across context. The plan converts “synergy” into owners and baselines and exposes integration, leverage, and regulator dependencies before closing. [researcher inference · strong] Sprint SEC transcript, lines 1065–1264
• Historical and vehicle context. Vodafone, Sprint, Japan Telecom, and Arm were corporate-control transactions, not ordinary SVF minority checks. Their debt, capex, employees, boards, regulators, and consolidated operating results belong to SBG/opcos; they cannot be evaluated with a venture mark alone. [researcher inference · strong] vehicles map FY2016 report, printed pp. 42–47
• When to use. Use it when the thesis requires authority over operations or scarce infrastructure and the acquirer has proven, transferable competence. Do not use control to compensate for a weak thesis, when the buyer lacks operators, or when leverage leaves no time to learn. [researcher inference · strong] Japan Telecom lab FY2007 report, printed pp. 13–20
• How to apply it. Freeze the pre-close baseline; assign leaders for network/product/pricing/talent/finance; compare the prior playbook with local regulation and customers; budget capex and integration; set churn, revenue, margin, FCF, leverage, and covenant gates; define management escalation and exit. Ask which capability is truly portable and what failure would prove it is not. [researcher inference · strong] Vodafone IC worksheet FY2007 report, printed pp. 13–20
• Examples.
  • Documented application — Vodafone Japan · [documented behavior · strong]. Base-station expansion, handset breadth, installments, Yahoo! Keitai, and White Plan were named interventions; one-year subscriber and operating-income improvements are milestones, not a complete acquisition return. FY2007 report, pp. 13–20
  • Counterexample — Sprint · [documented behavior · strong]. SBG proposed transferring network, device, procurement, and financing experience, but Son later admitted the rebuild was harder than expected and credited Marcelo Claure's execution. FY2016 report, printed pp. 9–10
  • Boundary case — Arm · [researcher inference · strong]. Arm required protecting neutral ecosystem trust while accelerating R&D, not importing a telecom turnaround; control did not simplify Arm China governance or make the later reported return cash. 2016 Arm statement, paragraphs 5–7 Arm prospectus
• Evidence. Transaction releases, annual reports, SEC materials, and the ex-ante labs independently establish operating plans, outcomes, and attribution limits. They do not provide a common acquisition-to-date cash IRR or prove Son personally executed the turnarounds. [researcher inference · strong] major-case packet 2006 Vodafone acquisition release Sprint SEC transcript, lines 1065–1264 2016 Arm strategic rationale FY2016 report, printed pp. 42–47
• Limits, failure modes, and exception ledger. Prior success can create false confidence; jurisdiction, spectrum, labor, technology, and customer economics may differ; leverage can force decisions before learning; control can damage talent or ecosystem neutrality. Japan Telecom's impairment and Sprint's difficulty prevent a Vodafone-only lesson. [researcher inference · strong] 2006 Japan Telecom impairment release
• Evolution and contradictions. The operating-transfer idea moved from domestic broadband/mobile to U.S. telecom and then foundational semiconductor IP. Current Ampere and robotics plans revive control acquisition, but the public record does not yet provide equivalent integration owners, baselines, or completed outcomes. [researcher inference · strong] FY2007 Vodafone report, printed pp. 13–20 Sprint SEC transcript, lines 1065–1264 2016 Arm strategic rationale 2026 AGM deck, PDF pp. 27–38 and 66
• Modern VC translation (as of 2026-08-22). What transfers is writing the value-creation plan before price and closing; what fails is importing telecom milestones into semiconductors, robotics, or AI infrastructure. Signals are named operators, customer retention, technical roadmap delivery, unlevered FCF, integration velocity, and covenant headroom. Misuse is calling a generic network an operating playbook. Ampere and the announced robotics build are current tests that require foundry/customer, talent, roadmap, integration, and financing evidence before claiming SBG synergy. [researcher inference · strong] FY2026 report, pp. 23–25

Treat capital scale and speed as a product inside vehicle and price limits

• Rule. Use unusually large, fast, or patient capital only when its incremental amount changes a valuable feasible action and increases probability-weighted value per share after price, dilution, burn, governance, concentration, and financing risk. [researcher inference · strong] Forbes pressure interview, nodes 46–47
• Evidence status. Capital scale and speed are investor-stated strategic advantages with strong recurrence. [investor-stated · strong] FY2017 report, printed pp. 14–15 Investment effectiveness is only moderately supported because the same mechanism can bid up prices, subsidize loss, and create rescue expectations. [researcher inference · moderate] Forbes 1999, Yahoo follow-on passages Forbes pressure interview, WeWork admission at node 10
• Provenance and origin. Early Yahoo follow-ons, Son's 2016 account of tracking Arm before committing, and the 2017 Vision Fund first close support a cross-record synthesis: mobilize unusually large capital after a strategic asset becomes actionable. This is a researcher reconstruction, not a Son-published rule. [researcher inference · moderate] Forbes 1999, Yahoo follow-on passages SoftBank World 2016 official English video, 02:03–05:38 SVF1 first close
• Mechanism and reasoning. Capital can fund infrastructure, outlast competitors, compress negotiation, recruit talent, and make SBG a preferred counterparty. Beyond the useful amount, it can inflate valuation, reduce external price discovery, weaken operating constraints, correlate the portfolio, and make the investor lender of last resort. [researcher inference · strong] Tokyo Forum, 54:06–59:17
• Historical and vehicle context. SBG corporate cash/debt, SVF1 outside-LP commitments, SVF2 preferred/co-invest economics, and nonrecourse project finance have different durations, costs, approvals, and loss bearers. “SoftBank capital” is not one balance sheet and Son's public commitment is not necessarily funded cash. [researcher inference · strong] FY2022 report, pp. 91–93
• When to use. Use it when a defined capital bottleneck blocks a positive-unit-economics plan, speed changes strategic access, and the security retains return at the proposed price. Do not use it to manufacture demand, win a valuation auction, or conceal a company that cannot attract outside capital. [researcher inference · strong] OneWeb lab FY2019 financial report, printed p. 14
• How to apply it. Calculate the minimum capital that unlocks the milestone; compare staged, syndicated, debt, project-finance, and control alternatives; model per-share return and total look-through exposure; set outside-financing proof, runway, governance, reserve, and stop gates; record who funds and who receives the economics. [researcher inference · strong] vehicles map FY2022 report, pp. 91–103
• Examples.
  • Documented application — Arm · [documented behavior · strong]. SBG executed a £24B cash acquisition; the transaction amount and structure are documented contemporaneously. 2016 acquisition release Preparedness account · [investor-stated · moderate]. Son later said he had tracked Arm for roughly a decade and committed within two weeks; that recollection supports speed, while price and return still require separate ledgers. SoftBank World 2016 official English video, 02:03–05:38
  • Documented application — Vision Fund and OpenAI · [contemporaneous record · strong]. SVF1 pooled $98.6B of committed capital at FY2022; the 2026 OpenAI plan separates $34.6B invested from a decided further $30B through SVF2. Scale is documented, realized return is not. FY2022 report, pp. 91–92 OpenAI follow-on, sections 1–5
  • Counterexample — WeWork · [researcher inference · strong]. Son's admission that he overpaid and believed too much in the entrepreneur supports the researcher-applied conclusion that capital scale can amplify a selection/governance error rather than create advantage. Forbes, node 10
• Evidence. Independent fund disclosures, transaction releases, direct talks, and failure records establish scale, speed, and their adverse mechanisms. They do not isolate the causal contribution of capital from founders, employees, markets, other investors, or later marks. [researcher inference · strong] people/capital map 2016 Arm acquisition release FY2017 report, printed pp. 70–73 Forbes pressure interview, WeWork admission at node 10 OneWeb loss disclosure, printed p. 14 / PDF p. 18
• Limits, failure modes, and exception ledger. More capital can lower discipline, fund negative unit economics, create governance mismatch, crowd out syndicate signals, and force follow-ons. [researcher inference · strong] Forbes, WeWork admission at node 10 OneWeb loss disclosure, printed p. 14 / PDF p. 18 Jack Ma's warning that too much money produces mistakes and Son's Amazon regret are retrospective witness accounts; neither proves that every larger check was unjustified or justified. [retrospective witness · moderate] Tokyo Forum, 54:06–59:17 Milken Japan, 08:48–10:37
• Evolution and contradictions. Scale moved from corporate follow-ons and leveraged acquisitions to an LP-backed fund, then defense and back to concentrated AI equity/infrastructure. The 2022 liquidity discipline and 2025–2026 “all in” posture remain in tension until position, project-finance, and stop limits are disclosed. [researcher inference · strong] FY2017 report, printed pp. 14–15 2022 CEO message, “Proper LTV control” 2025 CEO message
• Modern VC translation (as of 2026-08-22). What transfers is treating capital certainty and speed as part of the founder offer; what fails is copying a conglomerate's balance sheet inside a finite fund. Signals are capital-constrained positive economics, milestone leverage, outside validation, disciplined price, and explicit reserves. Misuse is “blitzscaling” without normalized contribution. Stargate is the current test: announced capacity and partner names do not substitute for site equity, nonrecourse closes, power/offtake, utilization, or customer contracts. [researcher inference · strong] Stargate announcement

Build ecosystems around scarce infrastructure, but make synergy earn a number

• Rule. Assemble complementary assets only when named counterparties can produce measurable incremental revenue, savings, retention, learning, or risk reduction after consent, integration cost, conflicts, and vehicle allocation. [researcher inference · strong] CCA transcript, pp. 2–6
• Evidence status. Cluster, alliance, and ecosystem creation are investor-stated and recurrent. [investor-stated · strong] Forbes 1999, alliance passages 2010 vision, PDF pp. 105–115 FY2019 CEO message, printed pp. 3–4 / PDF pp. 5–6 Realized synergy is only provisionally supported because the corpus rarely quantifies counterfactual economics; the measurement rule is researcher-applied. [researcher inference · provisional] FY2024 report, Note 46(3)(a), p. 138
• Provenance and origin. The late-1990s Internet alliance network is the earliest accessible version; later texts rename it strategic-synergy group, cluster, and orchestra while retaining autonomous category leaders around common capital and relationships. [investor-stated · strong] Forbes, alliance passages 2018 CEO message
• Mechanism and reasoning. Shared infrastructure can lower fixed cost, common distribution can accelerate adoption, and portfolio relationships can improve information or customer access. The mechanism fails when companies compete, consent/data rights block transfer, benefits are captive or subsidized, integration costs exceed gains, or common capital merely creates correlated exposure. [researcher inference · strong] NVIDIA AI Summit, 01:03:49–01:19:11
• Historical and vehicle context. Synergy crossed corporate subsidiaries, JVs, minority stakes, SVF companies, and current infrastructure partners. Benefits and costs accrue to different entities and LPs, so SBG NAV, fund gross marks, opco revenue, and partner economics cannot be merged into one “ecosystem return.” [researcher inference · strong] vehicles map FY2022 report, pp. 91–103
• When to use. Use it when complementarity is contractual or experimentally testable and participants have aligned incentives. Do not use it when the only commonality is theme, investor, brand, or valuation, or where cross-portfolio data/customer transfer violates governance or competition rules. [researcher inference · strong] portfolio factor analysis FY2024 financial report, Note 46(3)(a), p. 138
• How to apply it. Name both counterparties, asset/data/customer transferred, consent, integration owner, baseline, incremental metric, time window, cost, competing interests, and benefit allocation. Require a counterfactual and stop the program when measured incremental value does not cover cost and concentration. [researcher inference · strong] writings rule CCA transcript, pp. 2–6
• Examples.
  • Researcher-applied illustration — Yahoo/telecom distribution · [researcher inference · strong]. Yahoo content, broadband, and later mobile distribution formed a plausible access stack, but customer and segment improvements are not a complete cross-asset return attribution. FY2007 report, pp. 13–20
  • Documented thesis — Sprint/CCA and the Vision Fund cluster · [investor-stated · moderate]. Son proposed network and portfolio coordination, yet the sources omit incremental revenue, savings, and counterfactuals attributable to the network itself. CCA transcript, pp. 2–6 FY2019 message
  • Counterexample — Greensill/Katerra and correlated factors · [researcher inference · strong]. Portfolio links can transmit financing/counterparty risk rather than synergy; allegations remain disputed, and the public record does not yield security-level net losses for the pair. FY2024 report, Note 46(3)(a), p. 138
• Evidence. Independent writings, official talks, operating reports, and failure records establish the ecosystem thesis and repeated measurement gap. Because the model lacks quantified causal outcomes across three cases, its efficacy remains provisional despite strong evidence that Son advocated it. [researcher inference · provisional] talks framework Forbes 1999, alliance passages CCA transcript, pp. 2–6 FY2019 CEO message, printed pp. 3–4 FY2024 report, Note 46(3)(a), p. 138
• Limits, failure modes, and exception ledger. Narrative synergy can hide conflicts, data-rights problems, transfer pricing, cross-subsidy, double counting, and thematic concentration. A large portfolio increases possible connections but also coordination load; introductions, shared investors, and common branding are not measured economic benefit. [researcher inference · strong] FY2019 portfolio map, pp. 87–89
• Evolution and contradictions. The network moved from loosely allied Internet companies to telecom integration, a 5,000-company vision, an SVF orchestra, and now a more vertically linked AI stack. Increasing integration may improve coordination while reducing autonomy and increasing common-factor risk. [researcher inference · strong] 2010 vision, PDF pp. 105–115 2026 AGM deck, pp. 27–38
• Modern VC translation (as of 2026-08-22). What transfers is using portfolio companies as opt-in design partners and distribution tests; what fails is forcing captive cross-selling or treating correlation as a moat. Signals are signed pilots, independent conversion, incremental margin, lower acquisition cost, authorized data use, and benefit attribution. Misuse is counting meetings or shared branding. The current model/chip/data-center/agent/robot stack should publish customer contracts, utilization, integration cost, and entity-level value before claiming synergy. [researcher inference · provisional] 2026 CEO message, “Securing the four corners”

Separate adoption from market price, but never from financing survival

• Rule. Track user adoption and operating proof separately from market valuation, while maintaining enough portfolio-company runway and vehicle liquidity to survive until the thesis can mature or be disproved. [researcher inference · strong] Milken Japan, 10:41–12:45
• Evidence status. The adoption/price distinction and defense posture are investor-stated. [investor-stated · strong] FY2002 report, printed pp. 2–3 FY2022 Q1 transcript, A16 and A20–A23 The integrated cycle checklist is researcher inference with strong behavior across the dot-com bust, GFC, 2020, and 2022. [researcher inference · strong] FY2002 report, printed pp. 2–3 FY2009 report, printed pp. 17–20 FY2020 report, printed pp. 3–4 FY2022 Q1 transcript, A16 and A20–A23
• Provenance and origin. The early-2000s record shows Internet/broadband use growing while SoftBank's share price and finances deteriorated. Son made the adoption-versus-market distinction explicit again in the 2015 presentation and the 2019 Milken interview; annual reports document asset sales and debt reduction needed to continue. [investor-stated · strong] FY2002 report, printed pp. 2–3 SoftBank World 2015 deck, pp. 2–10 Milken Japan, 10:41–12:45
• Mechanism and reasoning. Markets can reprice long before operating adoption changes; the investor earns the horizon only if companies and the capital vehicle remain solvent. Separate scorecards prevent a falling mark from automatically disproving demand and prevent rising usage from excusing leverage, burn, or an unaffordable valuation. [researcher inference · strong] 2022 CEO message, “Proper LTV control”
• Historical and vehicle context. SBG could sell listed assets and manage parent LTV, while private portfolio companies depended on outside funding and fund support. A closed-end fund, opco, project company, and listed holding company require different cash, maturity, covenant, capital-call, and reserve tests. [researcher inference · strong] FY2022 report, pp. 91–93
• When to use. Use it during rapid repricing, financing shocks, or long adoption curves. Do not use it to deny evidence of product failure, defend stale private marks, or postpone a kill decision when retention, margin, regulation, or technical progress has broken. [researcher inference · strong] failure labs FY2019 financial report, printed p. 14
• How to apply it. Maintain separate adoption, operating, valuation, and liquidity dashboards; stress revenue, burn, FX, marks, collateral, capital calls, maturities, and closed exits; precommit minimum runway, parent/fund liquidity, and deployment brakes; define the evidence that reopens offense. [researcher inference · strong] 2022 CEO message
• Examples.
  • Documented application — dot-com/broadband · [documented behavior · strong]. Broadband adoption rose while SoftBank endured major losses, sold assets, and reduced debt; persistence was possible because financing actions accompanied the operating thesis. FY2003 report, printed pp. 2–3 and 10
  • Documented application — 2020–2022 defense · [documented behavior · strong]. SBG monetized assets, accumulated cash, slowed deployment, monitored runway, and stated normal/emergency LTV and two-year redemption policies. These are parent/fund controls, not proof security selection improved. FY2020 report, printed pp. 3–4 2022 CEO message
  • Counterexample — OneWeb · [contemporaneous record · strong]. Strategic adoption potential did not cure financing dependence; failed funding preceded Chapter 11 and a $1.593B reported FY2019 loss on $2.023B cumulative exposure. FY2019 financial report, printed p. 14 / PDF p. 18
• Evidence. Direct talks, annual reports, and loss disclosures independently support the distinction and the financing reflex. They do not provide a public, consistently applied offense trigger or prove that the later AI concentration remains inside the same risk limits. [researcher inference · strong] talks change chronology FY2002 report, printed pp. 2–3 Milken Japan, 10:41–12:45 OneWeb loss disclosure, printed p. 14 / PDF p. 18 FY2022 Q1 transcript, A16 and A20–A23
• Limits, failure modes, and exception ledger. “The thesis is intact” can become a sunk-cost defense; runway can be bought at destructive dilution; selling liquid winners can preserve weak illiquid holdings; asset-backed finance can correlate collateral and mark risk. Adoption is necessary evidence, not sufficient return. [researcher inference · strong] NVIDIA lab FY2019 report, NVIDIA note, printed p. 66
• Evolution and contradictions. The balance-sheet reflex recurs after crises, but public offense repeatedly returns. The unresolved question is whether 2022's LTV/cash rules constrain current project-level commitments, position concentration, and conditional capital—or only the reported parent perimeter. [researcher inference · strong] FY2020 report, asset-sale program, printed pp. 3–4 2022 CEO message, “Proper LTV control” 2026 CFO message, financing policy
• Modern VC translation (as of 2026-08-22). What transfers is dual underwriting of adoption and survival; what fails is copying SBG's 25% LTV threshold into a venture fund with different assets and liabilities. Signals are retention, contribution, runway, financing diversity, covenant headroom, and realistic exit timing. Misuse is marking to narrative or holding solely because price fell. For AI infrastructure, contracted demand, utilization, power, construction draws, model efficiency, and nonrecourse financing must be tracked alongside capability progress. [researcher inference · strong] FY2026 report, pp. 80–83

Attach every return numerator and denominator

• Rule. Before quoting or comparing performance, state the entity, vehicle, cash-flow perspective, cost, dates, ownership/dilution, realized proceeds, remaining mark, debt/hedges, fees/carry, tax, currency, contingencies, and benchmark. [researcher inference · strong] 2019 AGM presentation, pp. 43–44 and 102
• Evidence status. The underlying figures are contemporaneous corporate records. [contemporaneous record · strong] 2019 AGM presentation, pp. 43–44 and 102 FY2022 report, pp. 91–103 2024 AGM presentation, p. 15 The common denominator discipline is researcher inference because SBG uses materially different IRR, MOIC, NAV, SOTP, operating, and transaction perimeters. [researcher inference · strong] denominator field controls FY2025 SVF1 table, p. 78
• Provenance and origin. Early reports already separated public marks, sales, and operating results, but the need becomes acute in the SVF era when SBG LP economics, manager performance fees, third-party interests, in-kind transfers, and unrealized NAV appear beside corporate returns. [contemporaneous record · strong] FY1998 report, financial review, printed pp. 26–31 FY2017 report, Note 48 FY2022 report, fund structure and economics, pp. 91–103
• Mechanism and reasoning. A return ratio is meaningful only if cash flows and ownership share the same perimeter and time axis. The ledger prevents marks, commitments, intra-group transfers, operating revenue, and contingent securities from being misread as distributable cash or Son-personal performance. [researcher inference · strong] denominator field controls FY2022 report, pp. 91–103
• Historical and vehicle context. SBG corporate capital, SBGC, opcos, SVF1/SBIA, SVF2/SBGA, outside LPs, LatAm funds, SB Northstar, and Son-controlled co-investment have different economics. A reported gross company row is not LP net, consolidated IFRS income, parent cash, or personal return. [researcher inference · strong] vehicles map FY2022 report, pp. 91–103
• When to use. Use it for every return, ownership, valuation, commitment, exit, impairment, or “value created” claim. Do not compare figures when dates, currencies, realization states, security stacks, or legal entities cannot be reconciled; preserve not computable. [researcher inference · strong] uncertainty ledger FY2025 SVF1 table, p. 78
• How to apply it. Build security-level dated cash flows; reconcile shares and votes through dilution; separate primary, secondary, in-kind, debt, hedge, and contingent flows; allocate fees, tax, financing, third-party and manager economics; report operating, mark/NAV, gross-return, and realized-net-cash ledgers separately; show exclusions. [researcher inference · strong] full denominator packet FY2022 report, pp. 91–103
• Examples.
  • Documented measure — Alibaba · [contemporaneous record · strong]. SBG reported ¥9.7T total return and 57% IRR through 2023-05-11, net of monetization costs, pre-tax, and excluding the unmonetized stake; the stated ¥7.4B cost does not reconcile every public historical security flow. 2023 CFO message
  • Documented boundary — Arm · [contemporaneous record · strong]. SBG's reported ¥24.6T/10x includes retained market value and internal/in-kind paths; acquisition cost, SVF1 transfer, NVIDIA deposit, IPO proceeds, expenses, taxes, and intra-group effects remain distinct. 2024 CEO message, Arm discussion
  • Counterexample — OpenAI/Stargate · [researcher inference · strong]. Invested tranches, a decided follow-on, expected ownership, bridge finance, partner roles, and announced project capacity are commitments and structures, not one check or a realized return. OpenAI follow-on, sections 1–5 Stargate release
• Evidence. Independent CFO/CEO reports, audited filings, transaction notices, fund tables, and the denominator packet support the perimeter distinctions. The searched record has 534 canonical keys but only eleven sufficiently reconstructed families, so no Son-level all-history return is computable. [researcher inference · strong] exact denominator reconciliation FY2022 report, fund structure and economics, pp. 91–103 2023 CFO message, Alibaba monetization FY2025 SVF1 table, p. 78
• Limits, failure modes, and exception ledger. Missing private cash flows, taxes, fees, hedges, LP statements, dilution, and realization schedules can defeat reconstruction. Management-defined tables can be internally correct yet outcome-selected or incomparable. Precision must not be fabricated by dividing a known headline numerator by an unknown denominator. [researcher inference · strong] dominant-outlier tests FY2025 SVF1 table, p. 78
• Evolution and contradictions. The scorecard shifted from segment profit/FCF/debt to NAV and gross fund/company rows, while current communications again mix strategic commitments with asset value. The measures answer different questions; repetition of a high return under changing definitions is not corroboration. [researcher inference · strong] writings contradiction ledger FY2009 report, printed pp. 17–20 2019 AGM presentation, pp. 43–44 and 102 FY2022 report, pp. 91–103 OpenAI follow-on, sections 1–5
• Modern VC translation (as of 2026-08-22). What transfers is a four-ledger discipline: operating proof, marked NAV, gross security return, and realized net cash. What fails now is reconstructing portfolio- or fund-level IRR from incomplete public aggregates; structured rounds, secondaries, token-like rights, project finance, and cross-vehicle transfers make that shortcut less defensible. Signals are distributions, ownership bridges, independent marks, audited waterfalls, and reconciled fees/tax; misuse is reporting a commitment or last-round valuation as return. OpenAI is the grounded current example: model every closing and vehicle separately until cash realization exists. [researcher inference · strong] denominator reconciliation FY2025 SVF1 table, p. 78 FY2026 report, printed pp. 23–25

Underwrite duration and exits separately from entry conviction

• Rule. Approve a holding horizon, liquidity hierarchy, re-underwriting cadence, hedge/collar policy, and thesis-completion exit rule separately from the entry thesis; neither “hold forever” nor “take profit at 2x” is universal. [researcher inference · strong] portfolio exit analysis 2021 CEO message, capital-provider section
• Evidence status. Son's long-horizon rhetoric is investor-stated. [investor-stated · strong] 2021 CEO message, capital-provider section The security-specific duration/exit framework is researcher inference strongly supported by Alibaba, Yahoo Japan, NVIDIA, Arm, and standardized fund exits. [researcher inference · strong] 2001 Yahoo sale release FY2019 NVIDIA note, printed p. 66 / PDF p. 68 2023 CFO message, Alibaba monetization 2024 CEO message, Arm discussion FY2025 SVF1 exit rows, p. 78
• Provenance and origin. Early Yahoo positions combined staged follow-ons, partial sales, and residual ownership. Alibaba's prospectus documents early ownership and governance; the 2023 CFO message documents later staged monetization. The public sources describe actions, not a contemporaneous universal sell discipline. [contemporaneous record · strong] 2001 Yahoo sale release Alibaba prospectus 2023 CFO message, Alibaba monetization
• Mechanism and reasoning. Power-law outcomes may require holding through volatility, but solvency, fund life, collateral, concentration, governance, thesis maturity, taxes, and better uses of capital constrain duration. Separating exit from entry forces the IC to specify why future option value belongs to this security and this owner. [researcher inference · strong] FY2019 NVIDIA note, printed p. 66 / PDF p. 68
• Historical and vehicle context. Corporate holdings could persist for decades; SVF1 has a defined fund life and outside LPs; public shares can support collars or asset-backed finance; control assets carry operating/regulatory obligations. The feasible horizon is a vehicle property, not merely Son's conviction. [researcher inference · strong] FY2022 report, pp. 91–93
• When to use. Use it at entry, every financing, listing, merger, large mark change, collateral event, and strategic alternative. Do not hold solely because the company is admired or sell solely because an arbitrary multiple is reached. [researcher inference · strong] NVIDIA lab FY2019 report, NVIDIA note, printed p. 66
• How to apply it. Write thesis horizon and completion signals; map fund life, cash needs, collateral, lockups, governance, tax, and concentration; schedule re-underwriting; compare hold/sell/hedge/distribute/secondary/M&A; size a residual option explicitly; record counterfactual capital use and the facts that force exit. [researcher inference · strong] operational rules FY2025 SVF1 table, p. 78
• Examples.
  • Documented application — Alibaba · [documented behavior · strong]. SBG held across multiple financings, governance changes, IPO, and staged monetization; the company-reported result is exceptional but the original security-level cash ledger remains incomplete. 2023 CFO message
  • Counterexample — NVIDIA · [researcher inference · strong]. The January 2019 exit produced a reported roughly 2.0x gross result in the later standardized table but sacrificed enormous later upside; Son's COVID explanation conflicts with the contemporaneous date. FY2019 report, NVIDIA note TIME 2026, paragraphs beginning “When it comes to selling”
  • Boundary pair — Coupang versus Uber/Paytm · [researcher inference · strong]. Retained Coupang exposure preserved substantial marked value, while Uber and Paytm were fully exited at reported 1.2x and 0.7x. The table documents outcomes, not the original sell rationales. FY2025 SVF1 table, p. 78
• Evidence. Independent sale releases, annual reports, TIME's correction, filings, and fund tables establish diverse exit modes and the duration miss. They do not reveal full authorization, liquidity needs, tax, hedge constraints, or a Son-authored exit algorithm for most cases. [researcher inference · strong] uncertainty ledger 2001 Yahoo sale release FY2019 report, NVIDIA note, printed p. 66 TIME 2026, paragraphs beginning “When it comes to selling” FY2025 SVF1 table, p. 78
• Limits, failure modes, and exception ledger. Hindsight makes every sold winner look obvious; a feasible hold may have violated fund life, leverage, collateral, concentration, or information available then. Conversely, “patient capital” can excuse stale theses. Counterfactual hold returns are not actual losses and should never be booked as such. [researcher inference · strong] NVIDIA case packet FY2019 report, NVIDIA note, printed p. 66
• Evolution and contradictions. Early corporate duration coexisted with asset sales during crises; the Vision Fund added fund-life and distribution constraints; SBG later repurchased Arm exposure and now emphasizes retaining core AI assets. Long horizon and forced liquidity remain a structural tension, not a contradiction solved by rhetoric. [researcher inference · strong] 2001 Yahoo sale release 2020 CEO message, asset-sale program FY2022 report, fund life and distributions, pp. 91–93 2024 CEO message, Arm discussion
• Modern VC translation (as of 2026-08-22). What transfers is prewriting duration and exit governance; what fails is imitating corporate permanence in a finite fund. Signals are thesis completion, marginal return, remaining option value, fund life, concentration, liquidity, and governance. Misuse is “diamond hands” or mechanical MOIC selling. For liquid AI winners, use scheduled technical/competitive re-underwriting and a collateral/liquidity hierarchy before collars or sales; NVIDIA is the historical warning, not a rule to never sell. [researcher inference · strong] FY2019 NVIDIA note

Treat internal data and experiments as evidence only after outcome measurement

• Rule. Use internal telemetry, proposals, pilots, patents, and deployments to bridge vision to operations, but promote them to investment evidence only after a defined baseline, counterfactual, total cost, error/safety record, adoption, and realized operating outcome. [researcher inference · strong] SoftBank World 2024, 00:34:20–00:41:03
• Evidence status. High-throughput internal experimentation is documented behavior with moderate support. [documented behavior · moderate] 2023 AGM deck, p. 8 Causal productivity and investment efficacy remain insufficient in the public record because disclosed counts are mostly activity measures. [researcher inference · moderate] SoftBank World 2024, 00:34:20–00:41:03 SoftBank World 2025, 00:17:03–00:37:40
• Provenance and origin. The accessible operating form begins with 2013 network-data optimization and continues through 52,000 internal AI proposals in 2023, 188 patent submissions discussed in 2024, and the planned Cristal deployment in 2025. [documented behavior · moderate] SoftBank World FY2013 deck, PDF pp. 33–42 and 47–58 2023 AGM deck, p. 8 SoftBank World 2024, 00:34:20–00:41:03 SoftBank World 2025, 00:17:03–00:37:40
• Mechanism and reasoning. An investor-controlled company can become a design partner, shortening feedback and revealing integration, reliability, and customer-workflow constraints. Activity becomes evidence only when the experiment changes a decision or produces a measured outcome relative to a credible counterfactual. [researcher inference · strong] talks framework SoftBank World 2024, 00:34:20–00:41:03
• Historical and vehicle context. Most examples are SBG or SoftBank Corp. operating programs, not portfolio-company investment returns. Captive deployment may be subsidized and influenced by management, while fund companies have distinct data rights, boards, customers, and economics. [researcher inference · strong] vehicles map FY2022 report, pp. 91–103
• When to use. Use it when the investor or opco has legitimate access to representative workflows and can run a bounded, reversible test. Do not use proposal count, patents filed, agents created, or internal seats as evidence of external willingness to pay or venture-scale return. [researcher inference · strong] SoftBank World 2025, official caption route 00:17:03–00:37:40
• How to apply it. Freeze task baseline and owner; define test/control, success, safety, cost, and stop thresholds; include integration, compute, supervision, and incident cost; measure retention and repeat use; test external customers; audit cherry-picking; release capital only when the experiment changes probability-weighted value. [researcher inference · strong] 2025 CEO message, “A historical turning point”
• Examples.
  • Documented experiment — network telemetry · [documented behavior · moderate]. The 2013 framework used network and operational data to optimize decisions; the deck does not disclose model lift, data rights, cost, or causal return. FY2013 deck, pp. 33–42
  • Documented experiment — 52,000 AI proposals and 188 patent submissions · [documented behavior · moderate]. These counts show participation and pipeline throughput, not granted IP, product adoption, or cash benefit. 2023 AGM deck, p. 8 SoftBank World 2024, 00:34:20–00:41:03
  • Boundary case — Cristal/cognitive agents · [investor-stated · moderate]. The planned group rollout and nominal per-agent figure omit utilization, model/compute, integration, supervision, security, incidents, and verified productivity; adoption targets are not outcomes. SoftBank World 2025, 00:17:03–00:37:40
• Evidence. Official decks, provider-generated captions, annual messages, and the talks synthesis independently establish the programs and their measurement gaps. No retained model claims three verified causal operating wins; the support grade therefore remains moderate rather than padded with generic AI examples. [researcher inference · moderate] talks source guide, 2023–2025 sections SoftBank World FY2013 deck, PDF pp. 33–42 2023 AGM deck, p. 8 SoftBank World 2024, 00:34:20–00:41:03 SoftBank World 2025, 00:17:03–00:37:40
• Limits, failure modes, and exception ledger. Internal demand can be captive; employees may comply without valuing the product; activity measures invite Goodhart's law; patents can be low quality; unsafe or error-prone automation can shift costs to supervision and incident response. No public control-group or all-in ROI series closes these gaps. [researcher inference · strong] SoftBank World 2025, 00:17:03–00:37:40
• Evolution and contradictions. The record progresses from telemetry to mass ideation, patent pipelines, and autonomous-agent ambition, but public evaluation sophistication does not visibly rise at the same pace. The strongest contradiction is between increasing automation claims and absent error, authorization, rollback, and productivity evidence. [researcher inference · strong] SoftBank World FY2013 deck, PDF pp. 33–42 SoftBank World 2024, 00:34:20–00:41:03 SoftBank World 2025, 00:17:03–00:37:40
• Modern VC translation (as of 2026-08-22). What transfers is using a portfolio or LP network as an opt-in design-partner pool; what fails is treating captive rollout as product-market fit. Signals are blinded task success, retention, external conversion, gross margin after inference/integration, security incidents, human-review load, and measured productivity. Misuse is funding on proposal or agent count. Cristal is the grounded current case: its evidentiary status should change only when comparable baselines and realized external economics are published. [researcher inference · moderate] 2025 CEO message

Institutionalize attribution, decision rights, and succession

• Rule. Record who sourced, sponsored, diligenced, challenged, approved, negotiated, governed, operated, financed, followed on, and exited each position; attach every right and economic exposure to its legal entity and vehicle; maintain a tested succession and delegated-authority plan. [researcher inference · strong] case-specific decision-role map FY2026 report, printed pp. 98–103
• Evidence status. The record strongly documents distributed sourcing, operating, governance, and committee roles. [contemporaneous record · strong] Forbes 1999, U.S. venture-team passages FY2015 report, printed pp. 20–25 FY2016 report, printed pp. 18–19 FY2017 report, Vision Fund governance, printed pp. 70–73 FY2026 report, governance and investment process, printed pp. 98–103 The integrated credit-and-succession rule is researcher synthesis rather than a published Son checklist. [researcher inference · strong] case-specific decision-role map FY2026 report, printed pp. 98–103
• Provenance and origin. Forbes reported that Gary Rieschel ran the $230M Softbank Venture Capital pool in 1999. [contemporaneous record · strong] Forbes 1999, paragraphs beginning “Masayoshi Son” and “Since setting up” Later reports identify operating deputies and a named successor, that successor's departure, a three-tier Vision Fund governance structure, and a five-member SBG Investment Committee. [contemporaneous record · strong] FY2015 report, printed pp. 20–25 FY2016 report, printed pp. 18–19 FY2017 report, Vision Fund governance, printed pp. 70–73 FY2026 report, investment process, printed pp. 98–103 The recurring lesson is that the lone-genius narrative omits the organization that converts a thesis into a security and an operating result. [researcher inference · strong] case-specific decision-role map
• Mechanism and reasoning. A role ledger makes decisions reproducible: it reveals whose thesis, diligence, price, dissent, operating plan, and outcome should update the institution. Explicit authority reduces key-person delay and moral hazard; attribution without entity and decision-right boundaries instead rewards narrative ownership and hides who bore the risk. [researcher inference · strong] people, capital, and influence map FY2026 report, printed pp. 98–103
• Historical and vehicle context. Son, SBG, SBGC, operating subsidiaries, SVF1/SVF2, the manager, outside LPs, portfolio founders, boards, and counterparties had different rights and economics. An SBG board decision, fund IC recommendation, opco operating result, or founder action cannot be collapsed into “Son invested.” [researcher inference · strong] FY2022 report, fund governance and related-party boundaries, pp. 91–103
• When to use. Use this model before diligence begins, at every IC and board decision, when a vehicle or related party changes, at executive handoffs, and when public retrospectives assign credit or blame. It is mandatory when founder authority, fund manager incentives, or key-person dependency can change the security's value. [researcher inference · strong] FY2015 report, succession and governance, printed pp. 20–25
• How to apply it. Maintain a dated RACI-style role map and decision log; identify the legal principal, capital source, votes, vetoes, board rights, conflicts, fees, carry, information rights, and escalation path; preserve dissent and the facts that would reverse a decision; run annual key-person simulations and name interim authorities before a crisis. [researcher inference · strong] FY2026 report, corporate governance, printed pp. 98–103
• Examples.
  • Documented distribution — US venture team · [contemporaneous record · strong]. Forbes reported that Rieschel ran the $230M U.S. pool and that it held 55 companies; this is evidence of delegated portfolio work, not proof that every deal or result was his or Son's alone. Forbes 1999, paragraphs beginning “Masayoshi Son” and “Since setting up”
  • Documented operating delegation — Vodafone Japan and Sprint · [documented behavior · strong]. Ken Miyauchi and later Marcelo Claure held operating roles distinct from Son's capital-allocation and chairmanship roles; company outcomes therefore require role-specific attribution. FY2016 report, leadership profiles and Sprint section, printed pp. 16–19 and 42–47
  • Counterexample — succession reversal · [researcher inference · strong]. The 2015 report presented Nikesh Arora as Son's successor, while the 2016 report recorded Arora's departure and Son's continued leadership. Naming a successor without durable authority transfer did not resolve key-person dependence. FY2015 report, printed pp. 20–25 FY2016 report, printed pp. 18–19
• Evidence. Annual reports, governance disclosures, transaction filings, contemporary profiles, and the case packets independently establish distributed roles and repeated succession risk. They rarely disclose complete individual votes, dissent, compensation allocation, side letters, or informal influence, so this model refuses person-level precision where the record stops. [researcher inference · strong] Forbes 1999, U.S. venture-team passages FY2015 report, printed pp. 20–25 FY2016 report, printed pp. 18–19 FY2017 report, printed pp. 70–73 FY2026 report, governance and investment process, printed pp. 98–103 vehicle ledger
• Limits, failure modes, and exception ledger. Formal titles can understate informal control; logs can become compliance theater; diffuse attribution can erase accountability; over-specific credit can expose confidential deliberation or invite hindsight scoring. Unknown votes, side letters, and undocumented influence stay unknown, and a committee label does not prove independent challenge. [researcher inference · strong] FY2026 report, governance and risk-management qualifications, printed pp. 98–107
• Evolution and contradictions. SBG evolved from founder-centered corporate investing to operating deputies, fund governance, committees, and more explicit risk functions, yet Son remains representative director, chairman and CEO and central public narrator. The institution has added process without publicly demonstrating a completed transfer of ultimate strategic authority. [researcher inference · strong] Forbes 1999, U.S. venture-team passages FY2017 report, Vision Fund governance, printed pp. 70–73 Masayoshi Son officer biography FY2026 report, printed pp. 98–103
• Modern VC translation (as of 2026-08-22). What transfers is a security-level credit, authority, conflict, and succession ledger; what fails is founder mythology or committee theater. Signals are named owners, preserved dissent, independent valuation, delegated limits, incident escalation, and tested interim authority; misuse is assigning all wins to a star and all losses to a team. For current OpenAI exposure, keep Son's sponsorship, SBG's capital and approvals, Yoshimitsu Goto's finance role, OpenAI's operating control, Stargate counterparties, lenders, and each vehicle's economics separate. [researcher inference · strong] OpenAI follow-on, sections 1–5 Stargate release

Evidence gaps and candidate models

• Inventory reconciliation. The sixteen retained models merge every distinct reusable rule supported across the frozen writings, talks, investment cases, vehicle ledger, and failure packet; their order and anchors are the definitive parity ledger for the companion philosophy. Similar slogans were merged where they produced the same decision test, and apparent rules were excluded when the record could not support a repeatable application. [researcher inference · strong] writings synthesis talks synthesis investment rules FY1998 report, printed pp. 2–5
• Regret minimization — excluded as a standalone model. Son's Amazon and NVIDIA regret accounts are retrospective and outcome-aware; the accessible corpus does not show a contemporaneous rule applied prospectively across three decisions. The useful decision content—option value, duration, and exit discipline—is retained above. [researcher inference · strong] TIME 2026, paragraphs beginning “When it comes to selling” NVIDIA duration lab
• Probability arithmetic and scenario trees — merged, not independently retained. Son uses large numerical forecasts, but the public record rarely supplies probabilities, base rates, sensitivity ranges, or forecast scoring. Falsifiability, scenario ranges, and the outside-view discipline therefore sit inside the backcasting model rather than being padded into a positive standalone doctrine. [researcher inference · strong] FII 8 executive report, p. 61 forecast-change ledger
• “Adversity is opportunity” — excluded as a universal rule. Broadband and Sprint support countercyclical action, while WeWork and repeated rescue rounds show how the same framing can rationalize escalation. The decision-useful content is retained in staged re-underwriting, financing survival, and precommitted stops. [researcher inference · strong] FY2019 Q2 presentation, pp. 6–19 failure and process-change evidence
• Standalone pricing discipline — candidate with insufficient public record. Price and security design matter throughout the record, but no stable Son-authored valuation method, hurdle rate, or forecast-error series survives across regimes and vehicles. Pricing is therefore embedded in the founder, staging, return-ledger, and duration models pending contemporaneous IC materials or complete security cash-flow schedules. [researcher inference · provisional] exact denominator boundary retrieval ledger FY2025 SVF1 table, p. 78
• Pure founder intuition and personal chemistry — merged, not retained alone. Access and founder rapport recur, but the observable record does not provide a stable scorecard or separate intuition from category selection, price, governance, and operating follow-through. The transferable version is model six's dual-key founder-and-contract test. [researcher inference · strong] Rubenstein interview synthesis founder-fit failure evidence Bloomberg Rubenstein interview, paragraph beginning “His name is Gary”
• The 300-year cluster as an empirically proven longevity model — not established. Mission continuity, ecosystem design, and succession are retained separately, but the public record cannot validate centuries of persistence or show a completed succession mechanism. The literal duration is aspiration, not observed evidence. [researcher inference · strong] 2010 Next 30-Year Vision digest succession evidence 2010 vision, PDF pp. 105–123
• Renewable-energy catalyst — candidate with insufficient repetition. Son's 2011 post-Fukushima advocacy is a grounded episode, but one policy intervention does not establish a repeated investment rule, security-level economics, or outcome-tested playbook. It remains biographical evidence rather than a mental model. [researcher inference · provisional] Asahi 2011 archive
• Self-reflection and course correction — distributed across models rather than retained alone. Public admissions and group-level governance changes exist, but the corpus lacks a complete case-specific causal chain from error to changed rule to improved base rate. The supportable parts are explicit re-underwriting, solvency controls, attribution, and decision rights; generic humility would be slogan padding. [researcher inference · strong] FY2019 Q2 presentation, pp. 6–19 failure and process-change evidence
• Highest-value unresolved route. The 1992 HBR interview remains only partially inventoried because the full body was not lawfully accessible; it could change the provenance of early horizon, founder, or operating rules but cannot be used to manufacture support. Private IC minutes, security cash-flow ledgers, LP reports, and forecast scorecards are the next highest-value evidence because they would test pricing, attribution, and probability discipline directly. [researcher inference · strong] HBR canonical route writings unresolved record talks access limits investments uncertainty ledger

What the evidence supports

The defensible conclusion is narrower—and more useful—than the legend. From roughly 1995 through 2016, Son and the institutions he led repeatedly combined unusual counterparty access, an early view of a technology transition, selection of a distribution or infrastructure control point, and a capital or control structure capable of accelerating adoption. That pattern is visible in Yahoo and Yahoo Japan, the broadband/mobile operating build, Sprint, and Arm. It is not evidence of a general venture-selection edge, a Son-only operating contribution, or a complete career return. [researcher inference · strong] Yahoo announcement, “Purpose of investment” Sprint SEC-filed remarks, lines 701–814 and 1,065–1,264 Arm acquisition announcement, sections 2–7 and 11 profile thesis

The mechanism is observable enough to test. Son forecast a regime shift, sought a scarce gateway or foundational layer, used SBG or another disclosed vehicle to acquire influence or control, added localization, financing, distribution, or operating intervention, and attempted to preserve duration. Yahoo Japan’s 60/40 structure, Vodafone Japan’s corporate acquisition, Sprint’s majority-control proposal, and Arm’s full acquisition are different legal instruments serving that recurring control-point logic. [researcher inference · strong] Yahoo Japan license agreement, recitals A–B Vodafone acquisition release, sections 1–5 Sprint transaction record, lines 701–814 Arm announcement, sections 2–7

Three qualifications travel with that conclusion. First, the strongest early cases mix investing with corporate strategy and operating execution. Second, the record is power-law shaped: Alibaba dominates the best published corporate scorecard, while a small set of additional wins coexists with ordinary securities and serious losses. Third, the scale-fund period expanded access and capital but also exposed the method’s failure modes—high entry prices, minority governance, correlated financing risk, repeated follow-ons, and reliance on marks. [researcher inference · strong] FY2015 selected-position table, printed p. 26 / PDF p. 28 FY2022 Q1 transcript, answers 1, 4–5, and 9 FY2025 SVF1 public-company table, p. 78

The favorable thesis survives ordinary outcomes, but only in reduced form. Sprint needed a merger exit and has a company-calculated mixed-perimeter 25% equity IRR rather than an independently reconstructed cash return. The FY2025 standardized public rows show Uber at 1.2x, Grab at 0.8x, DiDi at 0.3x, and Paytm at 0.7x. NVIDIA was a profitable 2.0x exit that forfeited later duration. These outcomes support an ability to identify important platforms more strongly than an ability to price, govern, and exit every security well. [researcher inference · strong] 2020 AGM presentation, Sprint, pp. 34–35 FY2025 results, p. 78 FY2019 NVIDIA note, printed p. 66 / PDF p. 68

The failure record is not an appendix to the thesis; it defines its boundary. WeWork shows that founder conviction without price and governance discipline can invite capital escalation. DiDi shows that category leadership does not remove jurisdictional and listing risk. OneWeb shows that ecosystem fit does not ensure financing survival. Greensill and the Katerra-linked allegations show why affiliated exposures, receivables, and vehicle boundaries require independent verification. Later monitoring, valuation, defense, consolidation, investment-risk review, and liquidity controls are documented, but the public record does not establish that any specific loss caused each control or that the controls improved an error rate. [researcher inference · strong] WeWork S-1, pp. 49–51 DiDi 20-F, risk factors OneWeb loss table, printed p. 14 / PDF p. 18 FY2026 report, investment-risk review, printed p. 103 / PDF p. 105

The package’s direct-voice conclusion is also appropriately bounded. Thirty-five writings are disposed across thirty-six routes; thirty-four full works were content-analyzed and the inaccessible HBR body was inventoried without inference. Forty-six appearances are reconstructed from official and credible routes, but only eleven are full access and thirty-five are partial. Decks are treated as written corporate evidence, not speech; reporter prose and witness memory are not converted into Son’s voice; verified short quotations remain under the twenty-five-word source limit. [researcher inference · strong] writings coverage talks coverage HBR access page

The intellectual record is internally coherent without being self-validating. Mission continuity, backcasting, control points, ambitious founders, capital scale, and ecosystems recur from the software and Internet era into the ASI stack. Repetition proves persistence of belief, not forecast accuracy or investment efficacy. The strongest synthesis therefore keeps the doctrine while adding price, security, governance, financing survival, duration, attribution, and denominators as independent keys. [researcher inference · strong] 2010 Vision, PDF pp. 89–123 FY2019 Q2 corrective deck, pp. 6–29 mental-model inventory

What could falsify the thesis

The edge thesis is falsifiable because it predicts decision-near evidence and relative outcomes, not merely compelling retrospective stories. Newly recovered 1995–2016 IC, board, or negotiation records should show a specific control point identified before the outcome, differentiated access, a vehicle and ownership choice matched to the work, an operating or localization plan, and a duration thesis. Matched cases with those features should outperform contemporaneous technology investments after controlling for price, leverage, stage, geography, team, and market regime. If the records show only generic optimism, if matched performance disappears, or if teams rather than Son supplied the decisive sourcing and analysis, the personal-edge thesis weakens. [researcher inference · moderate] Yahoo 1995 announcement, “Purpose of investment” Arm acquisition announcement, sections 1–7 profile predictions

The strongest skeptical explanation is that Son is an exceptional promoter and capital mobilizer with privileged institutional access, whose reputation is dominated by Alibaba, a favorable Internet/mobile era, and operating value created by founders and teams; marks and selected tables then make the residual record look cleaner than it is. That explanation fits the incomplete denominator, price admissions, ordinary public rows, WeWork, DiDi, OneWeb, and the fact that many major “investments” were corporate or operating combinations. [researcher inference · strong] FY2015 selected table, printed p. 26 / PDF p. 28 FY2022 Q1 price admission, A9 FY2025 public-company table, p. 78

A pure-luck version is nevertheless too strong. Yahoo Japan’s filed localization structure, the broadband and Vodafone operating record, Sprint’s decision-near control logic, and Arm’s transaction design show repeated strategic action across different control points. A broad “investing genius” conclusion is also too strong. The surviving claim is a domain- and period-specific access-plus-control-point capability, with price, governance, financing, and duration as unresolved independent variables. [researcher inference · strong] Yahoo Japan agreement, recitals A–B FY2007 Vodafone operating record, printed pp. 13–20 Sprint SEC record, lines 701–814 Arm announcement, sections 2–7

The outlier tests reinforce this narrower verdict. In SBG’s selected FY2015 table, Alibaba supplied about 75% of the reported value/return perimeter. Mechanical removal leaves roughly 7.1x; removing Yahoo Japan too leaves roughly 3.5x. Those figures are not IRRs or complete-history returns because the table is outcome/size selected and lacks a common cash-flow chronology. In the FY2025 SVF1 public table, removing Coupang leaves about 1.03x; removing Coupang and DoorDash leaves about 0.87x, again before fees, financing, private positions, and LP economics. [researcher inference · strong] FY2015 table, printed p. 26 / PDF p. 28 FY2025 table, p. 78 dominant-outlier tests

Attribution and denominator audit

The source and decision denominators reconcile, but they answer different questions. Route counts measure retrieval; work counts measure writings and appearances; canonical keys measure discovered company/fund-position families; sufficiently reconstructed cases measure decision-level evidence; case labs are teaching units. None is a career investment population or hit-rate denominator. [researcher inference · strong] source-map coverage deals reconciliation

The 534-key arithmetic was re-performed from the frozen packet: 337 live cards plus 39 historical-only named families plus eight anonymous historical SVF2 positions equals 384; adding 101 LatAm positions and removing six overlaps yields 479; adding 23 FY2015 positions yields 502; adding four net-new FY2014 positions yields 506; and 28 net-new report/transaction families yields 534. Eleven sufficient families leave 523 unresolved. The eleven excluded contexts are supporting counterparties or regulators, not portfolio-company keys. [researcher inference · strong] reconciliation and exclusions FY2022 held-position tables, printed pp. 98, 101–103, and 109–110

Independent biography and attribution cross-checks

The formative biography is evidentially uneven and is correctly written that way. Den Fujita’s advice, discrimination, the Berkeley period, and the translator commercialization are mainly Son recollection or later witness reporting; TIME supplies a counterweight on family wealth and Forrest Mozer’s technical contribution. They are context, not proven causes of alpha. The corporate chronology—SoftBank’s formation, operating acquisitions, named roles, Arora’s succession reversal, and the current five-member IC—is supported by issuer records. [researcher inference · strong] Bloomberg Fujita excerpt TIME, biographical sections official history FY2026 report, printed p. 54

Economic language passes only when its perimeter travels with the number. Alibaba’s selected scorecard is an SBG management table, Arm’s ¥24.6T/10x is company-reported and substantially retained/unrealized with internal and in-kind paths, Sprint’s 25% is company-calculated with contingent/marked elements, SVF rows are gross manager-reported security perimeters, and WeWork’s $6.71B is a stated fund-accounting loss perimeter rather than total support cash. No number is converted into Son’s personal return, LP net return, or an all-history IRR. [researcher inference · strong] FY2015 report, printed p. 26 / PDF p. 28 2024 Arm discussion 2020 Sprint presentation, pp. 34–35 FY2025 report, printed p. 91 / PDF p. 93

Transfer limits for modern VCs

As of 2026-08-22, the transferable content is a decision system, not Son’s scale, persona, or corporate toolkit. A conventional VC cannot assume access to a listed balance sheet, carrier distribution, acquisition debt, sovereign LP capital, a controlling operating subsidiary, public-market monetization, or project-finance capacity. It can still use the method to identify a regime and control point, stage conviction, match rights to the work, precommit follow-on and stop rules, preserve a common return denominator, and keep credit and authority attached to legal entities. [researcher inference · strong] vehicle analysis FY2022 fund terms, printed pp. 91–93 FY2026 capital policy, printed pp. 101–103

The current OpenAI/Arm/Ampere/Stargate program is a grounded live test, not a validated replay of Yahoo or Arm. The appropriate modern underwriting requires separate legal entities, committed and funded tranches, compute and power contracts, utilization, customer economics, collateral and nonrecourse boundaries, dilution, syndication, and realized cash. Until those facts exist, “integrated AI stack” is an investor-stated strategic thesis and the associated value is conditional or marked. [researcher inference · strong] OpenAI follow-on, sections 1–5 Stargate announcement FY2026 report, printed pp. 23–25 and 101–103

A useful hypothetical application is a seed fund evaluating an AI-agent company. It may borrow Son’s backcast and bottleneck questions, but it should not imitate his check size or assume ecosystem synergy. It should specify whether the bottleneck is proprietary workflow data, distribution, switching cost, or a contract; cap initial exposure; demand security and governance rights proportionate to the work; and define adoption, gross-margin, retention, financing, and safety milestones before the next check. That is a researcher-applied translation, not a documented Son decision. [researcher inference · moderate] mental-model transfer tests OpenAI current transaction boundary

Unresolved evidence

The package is exhaustive over its mapped public routes, not exhaustive over the underlying historical reality. The remaining gaps are decision-critical because each could change price, attribution, or return conclusions. [researcher inference · strong] source-map residual gaps investment retrieval ledger

The learning path is substantive and works without custom components: begin with the chronological biography, read the source map, then writings, talks, investment forensics, the integrated investment philosophy, and the mental-model workbook. The philosophy supplies a recommended order, six decision labs including WeWork, four reusable checklists, ranked primary and secondary sources, a glossary, and go-deeper questions. People/capital maps appear in both biography and investments. [researcher inference · strong] study guide profile people map investments people map

Acceptance evidence

Central high-risk claim audit trail

Reproducible ordinary-claim sample

The sample uses two deterministic claims from each of the seven pre-review core documents: the first suitable dated factual claim in an early section and a later governance, economics, or access-boundary claim. This avoids clustering on famous cases. [researcher inference · strong]

Twenty-point acceptance matrix

Final verdict — PASS. The evidence supports a narrow, falsifiable access-plus-control-point edge in a defined period, not a general hit-rate or Son-only return claim. The independent source/evidence reviewer returned PASS with no P0/P1 after reconciling the source, writing, talk, deal, model, peer, citation, and anchor gates. The independent CTO/package reviewer also returned PASS with no P0/P1 after the 528-link, quote-length, rendered-anchor, isolated-build, and diff gates. The package satisfies all twenty acceptance conditions on the frozen public record. [researcher inference · strong] source freeze denominator reconciliation

Coverage and route-status convention

Status marks are route-level: [x] means the route is represented and access-attempted in a frozen WRITINGS, TALKS, DEALS, or SYNTH control note; [ ] means an unprocessed discovery/control route. Priorities rank the route for this dossier, not the speaker's importance. Multiple checked routes may represent one underlying work and are deduplicated in the note/retrieval ledger.

WRITINGS — 35 works / 36 processed routes

• ★★★ | Harvard Business Review — Japanese-Style Entrepreneurship: An Interview with Softbank's CEO, Masayoshi Son | https://hbr.org/1992/01/japanese-style-entrepreneurship-an-interview-with-softbanks-ceo-masayoshi-son | 1992-01/02 | edited interview; reliability B; access partial | Public title, byline, and issue metadata only; body remained unavailable, so no interview claims or quotes were reconstructed.
• ★★★ | Forbes — Master of the Internet | https://tollbit.forbes.com/forbes/1999/0705/6401146a.html | 1999-07-05; metadata modified 2013-06-06 | publisher-delivered profile/interview; reliability B; access full | TollBit is the canonical full-body route; original Forbes discovery alias: https://www.forbes.com/forbes/1999/0705/6401146a.html. One work and one route row, not two.
• ★★ | WIRED/Reuters — Softbank's Son: “Cannibalize!” | https://www.wired.com/2000/02/softbanks-son-cannibalize/ | 2000-02-01 | reported World Economic Forum remarks; reliability B; access full | Locator: HTML paragraphs 74–100; reported remarks, not a prepared Son essay.
• ★★★ | SoftBank Annual Report 1998 — To Our Shareholders | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy1998_01_en.pdf | report year 1998; published 1998-07-15 | signed president/CEO letter; reliability A; access full | Printed pp. 2–5 / PDF pp. 4–7.
• ★★★ | SoftBank Annual Report 1999 — Dear Shareholders | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy1999_01_en.pdf | report year 1999; published 1999-08-11 | signed president/CEO letter; reliability A; access full | Printed pp. 2–6 / PDF pp. 3–6.
• ★★★ | SoftBank Annual Report 2000 — Message from the President | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2000_01_en.pdf | report year 2000; published 2000-08 | signed president/CEO message; reliability A; access full | Printed pp. 2–7 / PDF pp. 35–40.
• ★★★ | SoftBank Annual Report 2001 — Message from the President | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2001_01_en.pdf | report year 2001; published 2001-08 | signed president/CEO message; reliability A; access full | Printed pp. 2–5 / PDF pp. 63–66.
• ★ | SoftBank Annual Report 2002 — To Our Shareholders | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2002_01_en.pdf | report year 2002; published 2002-07 | signed president/CEO letter; reliability A; access full | Printed pp. 2–3 / PDF pp. 4–5.
• ★★ | SoftBank Annual Report 2003 — To Our Shareholders | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2003_01_en.pdf | report year 2003; published 2003-07 | signed letter plus company exhibits; reliability A; access full | Signed message printed pp. 2–3 / PDF pp. 4–5; company evidence remains separately attributed.
• ★ | SoftBank Annual Report 2004 — Message from the President | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2004_01_en.pdf | report year 2004; published 2004-07 | signed president/CEO message; reliability A; access full | Printed pp. 2–3 / PDF pp. 4–5.
• ★ | SoftBank Annual Report 2005 — Message from the President | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2005_01_en.pdf | report year 2005; published 2005-07 | signed president/CEO message; reliability A; access full | Printed pp. 4–5 / PDF pp. 6–7.
• ★★★ | SoftBank Annual Report 2006 — To Our Stakeholders | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2006_01_en.pdf | report year 2006; published 2006-07 | signed president/CEO letter; reliability A; access full | Printed pp. 6–10 / PDF pp. 8–12.
• ★★ | SoftBank Annual Report 2007 — CEO's Message | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2007_01_en.pdf | report year 2007; published 2007-07 | signed chairman/CEO feature; reliability A; access full | Printed pp. 13–21 / PDF pp. 15–23.
• ★ | SoftBank Annual Report 2008 — CEO's Message | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2008_01_en.pdf | report year 2008; published 2008-07 | signed chairman/CEO Q&A; reliability A; access full | Printed pp. 13–20 / PDF pp. 15–22.
• ★★★ | SoftBank Annual Report 2009 — To Shareholders and Investors | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2009_01_en.pdf | report year 2009 | signed chairman/CEO Q&A; reliability A; access full | Printed pp. 11–18 / PDF pp. 13–20; finance continuation is director-attributed.
• ★★★ | SoftBank Annual Report 2010 — To Shareholders and Investors | https://group.softbank/system/files/pdf/ir/financials/annual_reports/annual-report_fy2010_01_en.pdf | report year 2010 | signed chairman/CEO message and Q&A; reliability A; access full | Printed pp. 15–21 / PDF pp. 17–23; finance detail is director-attributed.
• ★★★ | SoftBank Next 30-Year Vision — presentation deck | https://group.softbank/system/files/pdf/philosophy/vision/next30/press_20100625_01_en.pdf | 2010-06-25 | chairman/CEO deck; reliability A; access full | PDF pp. 1–133; deck text is not a speech transcript.
• ★★★ | SoftBank — Announcement of “SoftBank's Next 30-Year Vision” | https://group.softbank/en/news/press/20100625 | 2010-06-25 | official announcement; reliability A; access full | Heading and paragraphs 7–9; second processed route to the same Vision work, not an extra work.
• ★★★ | SoftBank Annual Report 2011 — Philosophy, Vision, and Strategy for Sustained Growth | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2011_01_en.pdf | report year 2011; published 2011-07 | signed chairman/CEO message; reliability A; access full | Printed pp. 8–19 / PDF pp. 10–21.
• ★★ | The Asahi Shimbun — INTERVIEW: Masayoshi Son off on new quest to raise green profile | https://jdarchive.org/api/item/1531498 | archive timestamp 2011-09-01T00:00:00Z; UI 2011-08-31 EDT | archived Q&A by Yoshibumi Wakamiya; reliability B; access full | Exact UTC/UI boundary retained; original dateline not separately visible; locate by archive fields and paragraph beginnings.
• ★★ | SoftBank Annual Report 2012 — Goal-Oriented Management | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2012_01_en.pdf | report year 2012; published 2012-07 | annual-report CEO message; reliability A; access full | Printed p. 010 / PDF p. 12.
• ★★★ | SoftBank Annual Report 2013 — Global No. 1 Is the Starting Point | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2013_01_en.pdf | report year 2013; published 2013-07 | annual-report CEO message; reliability A; access full | Printed p. 007 / PDF p. 9.
• ★★ | SoftBank Annual Report 2014 — We Can Change the World | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2014_01_en.pdf | report year 2014; published 2014-07 | annual-report CEO message; reliability A; access full | Printed p. 003 / PDF p. 5.
• ★★ | SoftBank Annual Report 2015 — Transformation into “SoftBank 2.0” | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2015_01_en.pdf | report year 2015; published 2015-07 | annual-report CEO message; reliability A; access full | Printed pp. 005–007 / PDF pp. 7–9.
• ★★ | SoftBank Annual Report 2016 — Finding Opportunities in Adversity | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2016_01_en.pdf | report year 2016; published 2016-07 | annual-report CEO message; reliability A; access full | Printed pp. 009–011 / PDF pp. 11–13.
• ★★★ | Masayoshi Son — Leading the Information Revolution to the Next Paradigm Shift | https://group.softbank/en/news/press/20160906 | 2016-09-06 | signed official Arm statement; reliability A; access full | Title and HTML paragraphs 1–9; corporate transaction facts remain corporate voice.
• ★★★ | SoftBank Annual Report 2017 — On Our Journey to Make the World a Better Place Through Leading the Information Revolution | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2017_01_en.pdf | report year 2017; published 2017-07 | annual-report CEO message; reliability A; access full | Printed pp. 012–015 / PDF pp. 14–17.
• ★★★ | Masayoshi Son — Cluster of No.1 Strategy for 300 Years of Continuous Growth | https://group.softbank/en/ir/financials/annual_reports/2018/message/son | report year 2018; published 2018-07 | canonical signed CEO HTML; reliability A; access full | Section headings and visible paragraphs. The report PDF is an alternate corporate route and does not contain this message.
• ★★ | SoftBank Annual Report 2019 — Aiming to Orchestrate the AI Revolution | https://group.softbank/system/files/pdf/ir/financials/annual_reports/annual-report_fy2019_01_en.pdf | report year 2019; published 2019-07 | annual-report CEO message; reliability A; access full | Printed pp. 002–005 / PDF pp. 4–7.
• ★★ | SoftBank Annual Report 2020 — Leading the Digital Transformation While Increasing Shareholder Value | https://group.softbank/system/files/pdf/ir/financials/annual_reports/annual-report_fy2020_01_en.pdf | report year 2020; published 2020-07 | annual-report CEO message; reliability A; access full | Printed pp. 002–005 / PDF pp. 6–9.
• ★★ | Masayoshi Son — Creating the Future with AI Entrepreneurs as a “Vision Capitalist” | https://group.softbank/en/ir/financials/annual_reports/2021/message/son | report year 2021; published 2021-07 | signed CEO message/Q&A; reliability A; access full | Stable section/question headings and numbered footnotes.
• ★ | Masayoshi Son — Investing in Technology Evolution While Adhering to Our LTV Commitment | https://group.softbank/en/ir/financials/annual_reports/2022/message/son | report year 2022; published 2022-07 | signed CEO message; reliability A; access full | Stable section headings and numbered footnotes.
• ★★★ | Masayoshi Son — Going on the Offensive to Lead the Information Revolution | https://group.softbank/en/ir/financials/annual_reports/2023/message/son | report year 2023; published 2023-07 | signed CEO message; reliability A; access full | Stable section headings and numbered footnotes.
• ★ | Masayoshi Son — Realizing Artificial Super Intelligence (ASI) for the Evolution of Humanity | https://group.softbank/en/ir/financials/annual_reports/2024/message/son | report year 2024; signed 2024-06-21; page information 2024-07-29 | signed CEO message; reliability A; access full | Stable headings, figure notes, and numbered footnotes.
• ★ | Masayoshi Son — Becoming the World's No. 1 ASI Platform Provider | https://group.softbank/en/ir/financials/annual_reports/2025/message/son | report year 2025; signed 2025-06-27; page information 2025-07-28 | signed CEO message; reliability A; access full | Stable headings, figure notes, and numbered footnotes.
• ★ | Masayoshi Son — Leading ASI's Four Key Areas to Become the No. 1 Platform Provider | https://group.softbank/en/ir/financials/annual_reports/2026/message/son | report year 2026; signed 2026-06-24; page information 2026-07-27 | signed CEO message; reliability A; access full | Stable section headings and numbered footnotes.

TALKS — AGMs 2008–2016 (nine works; closed-webcast archive routes)

• ★★ | SoftBank — 28th AGM year page | https://group.softbank/en/ir/investors/shareholders/2008 | 2008-06-25 | official archive page; reliability A; access partial | VOD/Materials table says webcast closed; no deck, notice, resolutions, Q&A, captions, or transcript exposed.
• ★ | SoftBank — FY2007 results release linked from the 28th AGM record | https://group.softbank/en/news/press/20080508_1 | 2008-05-08 | official results control; reliability A; access full | Release headings and tables; company disclosure, not AGM speech.
• ★ | SoftBank — 28th AGM convocation notice release | https://group.softbank/en/news/press/20080520/ | 2008-05-20 | official governance control; reliability A; access full | Release heading and meeting metadata; no meeting speech.
• ★★ | SoftBank — 29th AGM year page | https://group.softbank/en/ir/investors/shareholders/2009 | 2009-06-24 | official archive page; reliability A; access partial | Notice/VOD/Materials headings; webcast closed and no transcript/timestamps.
• ★★ | SoftBank — 29th AGM notice and business report | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2009/shareholders-meeting_29_01_en.pdf | 2009 | official notice PDF; reliability A; access full | PDF pp. 1–3 and business-report pages; board/company text, not verbatim Son speech.
• ★ | SoftBank — 29th AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2009/shareholders-meeting_29_02_en.pdf | 2009-06-24 | official resolutions PDF; reliability A; access full | PDF p. 1; outcome control only.
• ★★ | SoftBank — 30th AGM year page | https://group.softbank/en/ir/investors/shareholders/2010 | 2010-06-25 | official archive page; reliability A; access partial | Archive headings; webcast closed and no transcript/timestamps.
• ★★ | SoftBank — 30th AGM / Next 30-Year Vision webcast page | https://group.softbank/en/news/webcast/20100625_01_en | 2010-06-25 | official event/VOD route; reliability A; access partial | Page metadata and player route; speech unavailable as captioned transcript.
• ★★ | SoftBank — 30th AGM notice and business report | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2010/shareholders-meeting_30_01_en.pdf | 2010 | official notice PDF; reliability A; access full | PDF pp. 1–5 and named business-report pages; company text only.
• ★ | SoftBank — 30th AGM resolutions and voting result | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2010/shareholders-meeting_30_03_en.pdf | 2010-06-25 | official resolutions PDF; reliability A; access full | PDF p. 1 and voting-result table.
• ★★ | SoftBank — 31st AGM year page | https://group.softbank/en/ir/investors/shareholders/2011 | 2011-06-24 | official archive page; reliability A; access partial | VOD closed; links governance documents and presentation, with no transcript/Q&A text.
• ★★ | SoftBank — 31st AGM notice and business report | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2011/shareholders-meeting_31_01_en.pdf | 2011 | official notice PDF; reliability A; access full | Notice/business-report PDF pages; corporate authorship.
• ★ | SoftBank — 31st AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2011/shareholders-meeting_31_03_en.pdf | 2011-06-24 | official resolutions PDF; reliability A; access full | Resolution page; no speech.
• ★★★ | SoftBank — 31st AGM business-strategy presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2011/shareholders-meeting_31_05_en.pdf | 2011-06-24 | official deck; reliability A; access full | Exact PDF slide pages; visible deck text/graphics only, not spoken words.
• ★★ | SoftBank — 32nd AGM year page | https://group.softbank/en/ir/investors/shareholders/2012 | 2012-06-22 | official archive page; reliability A; access partial | VOD closed; documents/deck accessible; no transcript, captions, or Q&A.
• ★★ | SoftBank — 32nd AGM notice and business report | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2012/shareholders-meeting_32_01_en.pdf | 2012 | official notice PDF; reliability A; access full | Notice/business-report pages; corporate text.
• ★ | SoftBank — 32nd AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2012/shareholders-meeting_32_03_en.pdf | 2012-06-22 | official resolutions PDF; reliability A; access full | Resolution page; no speech.
• ★★★ | SoftBank — 32nd AGM business-strategy presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2012/shareholders-meeting_32_05_en.pdf | 2012-06-22 | official deck; reliability A; access full | PDF slide/page locators; deck text is not transcript text.
• ★★ | SoftBank — 33rd AGM year page | https://group.softbank/en/ir/investors/shareholders/2013 | 2013-06-21 | official archive page; reliability A; access partial | Closed VOD; governance and deck routes only.
• ★ | SoftBank — 33rd AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2013/shareholders-meeting_33_03_en.pdf | 2013-06-21 | official resolutions PDF; reliability A; access full | Resolution page.
• ★★★ | SoftBank — 33rd AGM business-strategy presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2013/shareholders-meeting_33_05_en.pdf | 2013-06-21 | official deck; reliability A; access full | PDF page/slide locators; no inferred speech.
• ★★ | SoftBank — 34th AGM year page | https://group.softbank/en/ir/investors/shareholders/2014 | 2014-06-20 | official archive page; reliability A; access partial | Closed VOD; governance/deck routes; no transcript or Q&A text.
• ★★ | SoftBank — 34th AGM notice and business report | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2014/shareholders-meeting_34_01_en.pdf | 2014 | official notice PDF; reliability A; access full | Notice/business-report pages; corporate text.
• ★ | SoftBank — 34th AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2014/shareholders-meeting_34_03_en.pdf | 2014-06-20 | official resolutions PDF; reliability A; access full | Resolution page.
• ★★★ | SoftBank — 34th AGM business-strategy presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2014/shareholders-meeting_34_04_en.pdf | 2014-06-20 | official deck; reliability A; access full | PDF page/slide locators; deck evidence only.
• ★★ | SoftBank — 35th AGM year page | https://group.softbank/en/ir/investors/shareholders/2015 | 2015-06-19 | official archive page; reliability A; access partial | Closed VOD; governance/deck routes; no transcript/captions/Q&A.
• ★ | SoftBank — 35th AGM convocation release | https://group.softbank/en/news/press/20150511 | 2015-05-11 | official governance release; reliability A; access full | Meeting and proposal headings; no speech.
• ★ | SoftBank — personnel changes following AGM cycle | https://group.softbank/en/news/press/20160621_2 | 2016-06-21 | official personnel release; reliability A; access full | Exact title and named-role paragraphs; team/role control, not Son speech.
• ★ | SoftBank — 35th AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2015/shareholders-meeting_35_03_en.pdf | 2015-06-19 | official resolutions PDF; reliability A; access full | Resolution page.
• ★★★ | SoftBank — 35th AGM business-strategy presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2015/shareholders-meeting_35_04_en.pdf | 2015-06-19 | official deck; reliability A; access full | PDF page/slide locators; deck evidence only.
• ★★ | SoftBank — 36th AGM year page | https://group.softbank/en/ir/investors/shareholders/2016 | 2016-06-22 | official archive page; reliability A; access partial | Closed VOD; documents/deck and personnel controls; no transcript or captions.
• ★★ | SoftBank — 36th AGM notice and business report | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2016/shareholders-meeting_36_01_en.pdf | 2016 | official notice PDF; reliability A; access full | Notice/business-report pages; corporate voice.
• ★ | SoftBank — 36th AGM resolutions | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2016/shareholders-meeting_36_04_en.pdf | 2016-06-22 | official resolutions PDF; reliability A; access full | Resolution page.
• ★★★ | SoftBank — 36th AGM business-strategy presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2016/shareholders-meeting_36_05_en.pdf | 2016-06-22 | official deck; reliability A; access full | Exact PDF pages/slides; visible deck text only.

TALKS — AGMs 2017–2026 (ten works; year page + deck + archived player)

• ★★ | SoftBank — 37th AGM year page | https://group.softbank/en/ir/investors/shareholders/2017 | 2017-06-21 | official meeting hub; reliability A; access partial | Page headings and materials table; one 2017 meeting.
• ★★★ | SoftBank — 37th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2017/shareholders-meeting_37_05_en.pdf | 2017-06-21 | official deck; reliability A; access full | PDF page plus printed slide locators; deck text is not speech.
• ★★ | Brightcove — 37th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6270707408001 | 2017-06-21 | official player; reliability A; access partial | Video ID 6270707408001; no public transcript/caption track.
• ★★ | SoftBank — 38th AGM year page | https://group.softbank/en/ir/investors/shareholders/2018 | 2018-06-20 | official meeting hub; reliability A; access partial | Year-page headings and materials table.
• ★★★ | SoftBank — 38th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2018/shareholders-meeting_38_04_en.pdf | 2018-06-20 | official deck; reliability A; access full | PDF/printed slide locators; corporate deck voice.
• ★★ | Brightcove — 38th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6271422980001 | 2018-06-20 | official player; reliability A; access partial | Video ID 6271422980001; no public transcript/captions.
• ★★ | SoftBank — 39th AGM year page | https://group.softbank/en/ir/investors/shareholders/2019 | 2019-06-19 | official meeting hub; reliability A; access partial | Page headings and meeting materials.
• ★★★ | SoftBank — 39th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2019/shareholders-meeting_39_05_en.pdf | 2019-06-19 | official deck; reliability A; access full | PDF page plus printed slide locators.
• ★★ | Brightcove — 39th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6271430467001 | 2019-06-19 | official player; reliability A; access partial | Video ID 6271430467001; no exposed transcript/captions.
• ★★ | SoftBank — 40th AGM year page | https://group.softbank/en/ir/investors/shareholders/2020 | 2020-06-25 | official meeting hub; reliability A; access partial | Year-page and materials headings.
• ★★★ | SoftBank — 40th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2020/shareholders-meeting_40_07_en.pdf | 2020-06-25 | official deck; reliability A; access full | Exact PDF/printed slide locators; no inferred speech.
• ★★ | Brightcove — 40th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6166819798001 | 2020-06-25 | official player; reliability A; access partial | Video ID 6166819798001; no transcript/caption track.
• ★★ | SoftBank — 41st AGM year page | https://group.softbank/en/ir/investors/shareholders/2021 | 2021-06-23 | official meeting hub; reliability A; access partial | Year-page headings and material links.
• ★★★ | SoftBank — 41st AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2021/shareholders-meeting_41_05_en.pdf | 2021-06-23 | official deck; reliability A; access full | PDF page plus printed slide locators.
• ★★ | Brightcove — 41st AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6254917265001 | 2021-06-23 | official player; reliability A; access partial | Video ID 6254917265001; no transcript/caption track.
• ★★ | SoftBank — 42nd AGM year page | https://group.softbank/en/ir/investors/shareholders/2022 | 2022-06-24 | official meeting hub; reliability A; access partial | Year-page headings and material links.
• ★★★ | SoftBank — 42nd AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2022/shareholders-meeting_42_05_en.pdf | 2022-06-24 | official deck; reliability A; access full | PDF page plus printed slide locators.
• ★★ | Brightcove — 42nd AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6307232284112 | 2022-06-24 | official player; reliability A; access partial | Video ID 6307232284112; no native transcript/captions.
• ★ | SoftBank YouTube — 42nd AGM highlight | https://youtu.be/j3kZIdhI_nU | 2022-06-24 | official highlight; reliability A; access partial | Uncaptioned excerpt; no unsupported quotation.
• ★★ | SoftBank — 43rd AGM year page | https://group.softbank/en/ir/investors/shareholders/2023 | 2023-06-21 | official meeting hub; reliability A; access partial | Year-page headings and materials.
• ★★★ | SoftBank — 43rd AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2023/shareholders-meeting_43_04_en.pdf | 2023-06-21 | official deck; reliability A; access full | PDF/printed slide locators; deck claims only.
• ★★ | Brightcove — 43rd AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6327716164112 | 2023-06-21 | official player; reliability A; access partial | Video ID 6327716164112; no transcript/captions.
• ★★ | SoftBank — 44th AGM year page | https://group.softbank/en/ir/investors/shareholders/2024 | 2024-06-21 | official meeting hub; reliability A; access partial | One underlying 2024 meeting.
• ★★★ | SoftBank — 44th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2024/shareholders-meeting_44_05_en.pdf | 2024-06-21 | official alternate deck route; reliability A; access full | PDF/printed slide locators; alternate route to the meeting, not an extra work.
• ★★ | Brightcove — 44th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6354423982112 | 2024-06-21 | official player; reliability A; access partial | Video ID 6354423982112; no transcript/captions.
• ★★ | SoftBank — 45th AGM year page | https://group.softbank/en/ir/investors/shareholders/2025 | 2025-06-27 | official meeting hub; reliability A; access partial | One underlying 2025 meeting.
• ★★★ | SoftBank — 45th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2025/shareholders-meeting_45_05_en.pdf | 2025-06-27 | official alternate deck route; reliability A; access full | PDF/printed slide locators; alternate route to the meeting, not an extra work.
• ★★ | Brightcove — 45th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6374034241112 | 2025-06-27 | official player; reliability A; access partial | Video ID 6374034241112; no transcript/captions.
• ★★ | SoftBank — 46th AGM year page | https://group.softbank/en/ir/investors/shareholders/2026 | 2026-06-24 | official meeting hub; reliability A; access partial | Year-page headings and materials.
• ★★★ | SoftBank — 46th AGM presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/investors/shareholders/2026/shareholders-meeting_46_05_en.pdf | 2026-06-24 | official deck; reliability A; access full | PDF/printed slide locators, including AI-stack appendix; deck text is not speech.
• ★★ | Brightcove — 46th AGM archived webcast | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6397875166112 | 2026-06-24 | official player; reliability A; access partial | Video ID 6397875166112; no transcript/captions.

TALKS — SoftBank World FY2013–FY2018 (six works; 39 distinct processed routes)

• ★★ | SoftBank Corp. — SoftBank World 2013 Japanese event page | https://www.softbank.jp/corp/news/conference/fy2013/ | 2013-07-23 | official event hub; reliability A; access full | Event title/date and material links.
• ★★★ | SoftBank — SoftBank World 2013 “Challenge the World” deck | https://www.softbank.jp/corp/set/data/news/conference/pdf/material/20130723_01.pdf | 2013-07-23 | official 64-page deck; reliability A; access full | PDF page locators; deck text only.
• ★★ | SoftBank Corp. — SoftBank World 2013 English event page | https://www.softbank.jp/en/corp/news/conference/fy2013/ | 2013-07-23 | official English hub; reliability A; access full | Heading and linked materials; duplicate language route to one event.
• ★★ | SoftBank Corp. — SoftBank World 2013 VOD | https://www.softbank.jp/en/corp/news/conference/webcast/?videoId=5813050797001 | 2013-07-23 | official 52-minute VOD; reliability A; access partial | Player metadata; no public transcript/captions/stable text timestamps.
• ★ | SoftBank News — SoftBank World 2013 recap | https://www.softbank.jp/sbnews/entry/20130808_01 | 2013-08-08 | official recap; reliability A; access full | Headings and attributed summary; not verbatim transcript.
• ★ | ITmedia — SoftBank World 2013 coverage | https://www.itmedia.co.jp/news/articles/1307/23/news123.html | 2013-07-23 | contemporaneous press recap; reliability B; access full | Article headings/paragraphs; secondary control.
• ★★ | SoftBank Corp. — SoftBank World 2014 Japanese event page | https://www.softbank.jp/corp/news/conference/fy2014/ | 2014-07-15 | official event hub; reliability A; access full | Event heading and material links.
• ★★★ | SoftBank — SoftBank World 2014 “The Stage of the Information Revolution Goes Global” deck | https://www.softbank.jp/corp/set/data/news/conference/pdf/material/20140715_01.pdf | 2014-07-15 | official 66-page deck; reliability A; access full | Exact PDF pages; visible deck evidence.
• ★★ | SoftBank Corp. — SoftBank World 2014 English event page | https://www.softbank.jp/en/corp/news/conference/fy2014/ | 2014-07-15 | official English hub; reliability A; access full | Heading and linked materials.
• ★★ | SoftBank Corp. — SoftBank World 2014 VOD | https://www.softbank.jp/en/corp/news/conference/webcast/?videoId=5813055525001 | 2014-07-15 | official 64-minute VOD; reliability A; access partial | No public transcript, captions, or stable text timestamps.
• ★ | SoftBank News — SoftBank World 2014 recap | https://www.softbank.jp/sbnews/entry/20140731_01 | 2014-07-31 | official recap; reliability A; access full | Recap headings and paragraph locators; not transcript text.
• ★ | SoftBank Group — SoftBank World 2014 webcast mirror | https://group.softbank/en/news/webcast/y5xk38ht | 2014 | official mirror route; reliability A; access partial | Page/player metadata; same keynote, not an additional work.
• ★★ | SoftBank Corp. — SoftBank World 2015 Japanese event page | https://www.softbank.jp/corp/news/conference/fy2015/ | 2015-07-30 | official event hub; reliability A; access full | Event title/date and material links.
• ★★★ | SoftBank — SoftBank World 2015 “Information Revolution: Now, to the Next World” deck | https://www.softbank.jp/corp/set/data/news/conference/pdf/material/20150730_01.pdf | 2015-07-30 | official 60-page deck; reliability A; access full | PDF page locators; not spoken transcript.
• ★★ | SoftBank Corp. — SoftBank World 2015 English event page | https://www.softbank.jp/en/corp/news/conference/fy2015/ | 2015-07-30 | official English hub; reliability A; access full | Heading and links; alternate language route.
• ★★ | SoftBank Corp. — SoftBank World 2015 VOD | https://www.softbank.jp/en/corp/news/conference/webcast/?videoId=5813067939001 | 2015-07-30 | official 75-minute VOD; reliability A; access partial | No public captions or official transcript.
• ★ | SoftBank News — SoftBank World 2015 recap | https://www.softbank.jp/sbnews/entry/20150814_01 | 2015-08-14 | official recap; reliability A; access full | Heading/paragraph locators; summary voice.
• ★★ | Logmi — SoftBank World 2015 transcript part 1 | https://logmi.jp/main/social_economy/80202 | 2015 | speaker-labeled third-party transcript; reliability B; access full | Article heading and paragraph openings; attribution retained.
• ★★ | Logmi — SoftBank World 2015 transcript part 2 | https://logmi.jp/main/social_economy/80303 | 2015 | speaker-labeled third-party transcript; reliability B; access full | Continuation route to same keynote.
• ★ | SoftBank Mobile — 2015 personnel/company control | https://www.softbank.jp/corp/group/sbm/news/press/2015/20150518_01/ | 2015-05-18 | official release; reliability A; access full | Named-role/context control; not keynote voice.
• ★★ | SoftBank Corp. — SoftBank World 2016 Japanese event page | https://www.softbank.jp/corp/news/conference/fy2016/ | 2016-07-21 | official event hub; reliability A; access full | Event title/date and material links.
• ★★★ | SoftBank — SoftBank World 2016 “The Largest Paradigm Shift in Human History” deck | https://www.softbank.jp/corp/set/data/news/conference/pdf/material/20160721_01.pdf | 2016-07-21 | official 50-page deck; reliability A; access full | Exact PDF page locators.
• ★★ | SoftBank Corp. — SoftBank World 2016 English event page | https://www.softbank.jp/en/corp/news/conference/fy2016/ | 2016-07-21 | official English hub; reliability A; access full | Heading/material links.
• ★★ | SoftBank Corp. — SoftBank World 2016 VOD | https://www.softbank.jp/en/corp/news/conference/webcast/?videoId=5813076730001 | 2016-07-21 | official 77-minute VOD; reliability A; access full | Used with captioned official YouTube duplicate and deck.
• ★★★ | SoftBank official YouTube — SoftBank World 2016 | https://www.youtube.com/watch?v=bewYm7LB9p4 | 2016-07-21 | official video with automatic captions; reliability A; access full | Timestamped caption route; machine-caption caveat retained; citation alias https://www.youtube.com/watch?v=bewYm7LB9p4&t=123s resolves to the same video at 02:03 and is not an extra route.
• ★ | SoftBank News — SoftBank World 2016 recap | https://www.softbank.jp/sbnews/entry/20160721_01 | 2016-07-21 | official recap; reliability A; access full | Heading/paragraph locators; recap voice.
• ★ | SoftBank Group — Execution of Bridge Loan Agreement for ARM Acquisition | https://group.softbank/en/news/press/20160718_0 | 2016-07-18 | official acquisition-financing release; reliability A; access full | Locator: opening and section 1. Bridge-loan agreement and acquisition-financing control; not the Arm transaction announcement or keynote speech.
• ★★ | SoftBank Corp. — SoftBank World 2017 English event page | https://www.softbank.jp/en/corp/news/conference/fy2017/ | 2017-07-20 | official event hub; reliability A; access full | Event heading, speaker roster, and material links.
• ★★ | SoftBank Corp. — SoftBank World 2017 VOD | https://www.softbank.jp/en/corp/news/conference/webcast/?videoId=5813085909001 | 2017-07-20 | official 132-minute VOD; reliability A; access full | Used with official captioned duplicate; guest speakers separately attributed.
• ★★★ | SoftBank — SoftBank World 2017 “A New World Led by the Information Revolution” deck | https://www.softbank.jp/en/corp/set/data/news/conference/pdf/material/20170720_01.pdf | 2017-07-20 | official 47-page deck; reliability A; access full | Exact PDF page locators.
• ★ | SoftBank News — SoftBank World 2017 recap | https://www.softbank.jp/sbnews/entry/20170721_02 | 2017-07-21 | official recap; reliability A; access full | Named guests and recap paragraphs; not transcript voice.
• ★★ | SoftBank official YouTube — SoftBank World 2017 Japanese upload | https://www.youtube.com/watch?v=Zew9RU99NFg | 2017-07-20 | official duplicate video; reliability A; access full | Duplicate-language/player control; one keynote work.
• ★★★ | SoftBank official YouTube — SoftBank World 2017 English upload | https://www.youtube.com/watch?v=z7kHvHKElQc | 2017-07-20 | official video with automatic captions; reliability A; access full | Timestamped caption route; platform-generated captions and guest boundaries disclosed; citation alias https://www.youtube.com/watch?v=z7kHvHKElQc&t=1340s resolves to the same video at 22:20 and is not an extra route.
• ★★ | SoftBank Corp. — SoftBank World 2018 English event page | https://www.softbank.jp/en/corp/news/conference/fy2018/ | 2018-07-19 | official event hub; reliability A; access partial | Event rail and duration; no official deck or caption track found.
• ★★ | SoftBank Corp. — SoftBank World 2018 VOD | https://www.softbank.jp/en/corp/news/conference/webcast/?videoId=5974645263001 | 2018-07-19 | official 167-minute VOD; reliability A; access partial | Player metadata; no official text timestamps or speaker-labeled Q&A.
• ★★ | Logmi — SoftBank World 2018 transcript series | https://logmi.jp/main/technology/301878 | 2018 | speaker-labeled third-party transcript; reliability B; access partial | Headings/paragraphs; series boundary and nonofficial status retained.
• ★ | AI NOW — SoftBank World 2018 recap | https://ainow.ai/2018/07/19/143163/ | 2018-07-19 | contemporaneous Japanese recap; reliability B; access full | Heading/paragraph locators; secondary evidence.
• ★ | Business Network — SoftBank World 2018 coverage | https://businessnetwork.jp/article/6213/ | 2018 | contemporaneous recap; reliability B; access full | Article heading/paragraph controls.
• ★ | Robot Start — SoftBank World 2018 keynote recap | https://robotstart.info/2018/07/19/sbw-ceo-keynote01.html | 2018-07-19 | contemporaneous recap; reliability B; access full | Headings/paragraphs; not verbatim Son speech.

TALKS — official briefings, dialogues, interviews, and current AI appearances (21 works; 40 processed routes)

• ★★ | FII Institute — FII 8 Executive Report, “The Importance of Investing in AI” | https://fii-institute.org/wp-content/uploads/2024/11/FII-8-Executive-Report.pdf | 2024-11-20; interview at FII 8 | institutional recap; reliability B; access partial | Printed/PDF p. 61 quotation blocks; not a full transcript or verified full-video route.
• ★ | SoftBank — Next 30-Year Vision event notice | https://group.softbank/en/news/info/20100621 | 2010-06-21 | official event notice; reliability A; access full | Event heading/date; metadata route to the 2010-06-25 appearance.
• ★★ | SoftBank — FY2019 Q3 earnings webcast notice | https://group.softbank/en/news/info/20200106_01 | event 2020-02-12 | official event page; reliability A; access partial | Event paragraph and Brightcove ID 6132102091001; no captions/transcript.
• ★★ | SoftBank — FY2019 full-year earnings webcast notice | https://group.softbank/en/news/info/20200424 | event 2020-05-18 | official release-package route; reliability A; access partial | Release-package row and player metadata; no captions/transcript.
• ★★★ | SoftBank — Consolidated Financial Report for FY2019 / results release | https://group.softbank/en/news/press/20200518 | 2020-05-18 | official results control; reliability A; access full | Release headings and tables; company disclosure, not Son speech.
• ★★★ | SoftBank and NVIDIA — termination of NVIDIA's acquisition of Arm | https://group.softbank/en/news/press/20220208 | 2022-02-08 | joint official release; reliability A; access full | Opening and attributed Son paragraphs; transaction counterparties separated.
• ★★★ | SoftBank — notice concerning termination of sale of Arm shares | https://group.softbank/en/news/press/20220208_0 | 2022-02-08 | official transaction notice; reliability A; access full | Paragraphs 1–3 and numbered footnotes; corporate evidence.
• ★★ | SoftBank — Sozo Summit: Conversation with Masayoshi Son and Larry Fink | https://group.softbank/en/news/webcast/20190919_01_en | 2019-09-19 | official webcast page; reliability A; access partial | Correct work identity/date/speakers; Brightcove upload metadata is not event date.
• ★★ | SoftBank — FY2019 Q2 Earnings Briefing and WeWork Response | https://group.softbank/en/news/webcast/20191106_01_en | 2019-11-06 | official webcast; reliability A; access partial | Event-page paragraph and player ID 6127106634001; no caption track.
• ★★ | SoftBank — A Conversation with Masayoshi Son, Arm/NVIDIA termination | https://group.softbank/en/news/webcast/20220208_02_en | 2022-02-08 | official post-results interview; reliability A; access partial | Page title and Brightcove ID 6294735281001; release controls supply text, not video speech.
• ★★ | SoftBank — A Conversation with Masayoshi Son, FY2021 results | https://group.softbank/en/news/webcast/20220512_02_en | 2022-05-12 | official post-results interview; reliability A; access partial | Page title/player ID 6304260229001; transcript control is the global call, not necessarily identical wording.
• ★★★ | SoftBank — A Conversation with Masayoshi Son, FY2022 Q1 “defense mode” | https://group.softbank/en/news/webcast/20220808_02_en | 2022-08-08 | official post-results interview; reliability A; access full | Paired with official global-call Q&A transcript; vehicle and speaker boundaries retained.
• ★★ | SoftBank World 2025 — Masayoshi Son and Sam Altman special session | https://group.softbank/en/news/webcast/20250901_01 | held 2025-07-16; published 2025-09-01 | official webcast; reliability A; access partial | Brightcove ID 6377401878112, 18:23; no exposed transcript/captions.
• ★★ | SoftBank World 2025 — Masayoshi Son keynote | https://group.softbank/en/news/webcast/20250901_02 | held 2025-07-16; published 2025-09-01 | official webcast; reliability A; access partial | Brightcove ID 6377403344112; official YouTube automatic-caption route is separate.
• ★★ | SoftBank — Digest Version of “Next 30-Year Vision” | https://group.softbank/en/news/webcast/next30_en | 2010-06-25 | official six-minute digest; reliability A; access partial | Page paragraph and player ID 6314097694112; no exposed caption track.
• ★★★ | SoftBank — FY2019 Q3 earnings presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/presentations/2019/earnings-presentation_q3fy2019_01_en.pdf | 2020-02-12 | official deck; reliability A; access full | Exact slide/page locators; deck text is corporate presentation evidence.
• ★★★ | SoftBank — FY2019 full-year earnings presentation | https://group.softbank/media/Project/sbg/sbg/pdf/ir/presentations/2019/earnings-presentation_q4fy2019_01_en.pdf | 2020-05-18 | official deck; reliability A; access full | Exact slide locators; not a transcript.
• ★★★ | Masayoshi Son — Keynote Remarks at CCA Global Expo | https://group.softbank/system/files/pdf/ir/presentations/2014/press_20140327_02_en.pdf | 2014-03-27 | official eight-page transcript; reliability A; access full | PDF page locators; preferred quote-control route.
• ★★★ | SoftBank — FY2019 Q2 investor presentation / WeWork accounting | https://group.softbank/system/files/pdf/ir/presentations/2019/investor-presentation_q2fy2019_01_en.pdf | 2019-11-08 | official deck; reliability A; access full | Accounting slides 4–5 / PDF pp. 8–9; SBG subsidiary and SVF1 exposures remain distinct.
• ★★★ | SoftBank — FY2021 global earnings conference call transcript | https://group.softbank/system/files/pdf/ir/presentations/2021/earnings-conference-call_q4fy2021_01.pdf | 2022-05-12 | official speaker-labeled transcript; reliability A; access full | Question/answer number plus printed PDF page; distinguish call from separate interview.
• ★★★ | SoftBank — FY2022 Q1 global earnings conference call transcript | https://group.softbank/system/files/pdf/ir/presentations/2022/earnings-conference-call_q1fy2022_01.pdf | 2022-08-08 | official speaker-labeled transcript; reliability A; access full | Question/answer and printed-page locators.
• ★ | Milken Institute — Japan Symposium 2019 event page | https://milkeninstitute.org/events/japan-symposium-2019 | 2019-03-25 | official event metadata; reliability A; access full | Event date and agenda context; not an extra appearance.
• ★★★ | Milken Institute — Closing Plenary: A Conversation with Masayoshi Son, media 21333 | https://milkeninstitute.org/media/21333 | appearance 2019-03-25; video uploaded 2019-07-09 | official 43:12 video; reliability A; access full | Timestamped English automatic-caption route; event and upload dates kept distinct.
• ★★ | Omny/Bloomberg — A Conversation with Billionaire Masayoshi Son | https://omny.fm/shows/foundering/a-conversation-with-billionaire-masayoshi-son | recorded 2017-09-19 | publisher audio/transcript backup; reliability A; access full | Duplicate/edited backup to Bloomberg interview, not another work.
• ★★ | Brightcove — Sozo Summit Son/Fink player | https://players.brightcove.net/5706818909001/default_default/index.html?videoId=6110251805001 | page dated 2019-09-19 | official 47:53 player; reliability A; access partial | No subtitles/transcript; OpenAI API unavailable, local transcription prohibited and discarded.
• ★★ | TIME — Masayoshi Son Is Betting Everything on American AI | https://time.com/7380074/ai-masayoshi-son-japan-softbank | 2026-02-24 | reported profile/interview; reliability B; access full via lawful index | Canonical/AMP requests returned 406; paragraph openings from lawfully indexed full text; trailing-slash alias https://time.com/7380074/ai-masayoshi-son-japan-softbank/ is the same route, not an extra row.
• ★★★ | Bloomberg — The David Rubenstein Show: Masayoshi Son | https://www.bloomberg.com/news/videos/2017-10-11/the-david-rubenstein-show-masayoshi-son-video | recorded 2017-09-19; published 2017-10-11 | publisher video with CC transcript; reliability A; access full | Transcript paragraph openings used because exposed CC timestamp is 00:00 only.
• ★★ | Bloomberg — The Meeting That Changed Masayoshi Son's Life | https://www.bloomberg.com/news/videos/2017-10-11/the-meeting-that-changed-masayoshi-son-s-life-video | 2017-10-11 | publisher interview excerpt; reliability A; access full | Duplicate excerpt to same Rubenstein recording; retrospective Fujita account.
• ★★ | Forbes — Masayoshi Son Talks WeWork, Vision Fund and SoftBank Under Siege | https://www.forbes.com/sites/alexkonrad/2020/04/05/exclusive-interview-masayoshi-son-talks-wework-vision-fund-softbank | 2020-04-06 | edited interview/profile; reliability B; access full | Publisher JSON node/paragraph locators; observed summit and direct interview text separated; trailing-slash alias https://www.forbes.com/sites/alexkonrad/2020/04/05/exclusive-interview-masayoshi-son-talks-wework-vision-fund-softbank/ is the same route, not an extra row.
• ★★ | NVIDIA On-Demand — AI Summit Japan fireside chat, session SJP1047 | https://www.nvidia.com/zh-tw/on-demand/session/aisummitjp24-sjp1047 | 2024-11-14 | official duplicate session page; reliability A; access full | Duration PT5211S; duplicate to official captioned video, one work.
• ★★★ | SEC — SoftBank–Sprint Strategic Partnership corrected transcripts and slides | https://www.sec.gov/Archives/edgar/data/101830/000119312512423563/d424843ddefa14a.htm | 2012-10-15 | SEC-filed transcripts; reliability A; access full | EDGAR line, press-conference timestamp, and filed-slide locators; initial 70% structure only.
• ★★ | Tokyo Forum — Special Dialogue 2: Masayoshi Son and Jack Ma | https://www.tokyoforum.tc.u-tokyo.ac.jp/ja/event/2019_specialdialog2.html | appeared 2019-12-06; archive 2019-12-17 | official event/video route; reliability A; access full | Timestamped automatic captions; speakers separated.
• ★★ | University of Tokyo — Son and Jack Ma dialogue recap | https://www.u-tokyo.ac.jp/focus/en/features/z0508_00151.html | 2019-12-17 | official companion recap; reliability A; access full | Heading/paragraph openings; preferred control for short verified quotes.
• ★★ | White House — Masayoshi Son speaks about the Stargate Project | https://www.whitehouse.gov/videos/softbank-ceo-masayoshi-son-speaks-about-the-stargate-project-at-the-white-house | remarks 2025-01-21; page 2025-04-30 | official video page; reliability A; access partial | Official video timestamp and embedded YouTube automatic-caption identity.
• ★★ | SoftBank official YouTube — SoftBank World 2025 keynote | https://www.youtube.com/watch?v=2TMANDp9v3w | delivered 2025-07-16 | official duplicate video; reliability A; access partial | Japanese platform-generated WebVTT, duration 47:38; machine-caption caveat.
• ★★ | YouTube backup — CCA Global Expo keynote | https://www.youtube.com/watch?v=2XDW6R9EqMU | 2014-03-27 | video timestamp backup; reliability B; access full | Quotes verified against official transcript before timestamp use.
• ★★ | SoftBank official YouTube — SoftBank World 2024 ASI keynote | https://www.youtube.com/watch?v=BzJHh5IZV2o | uploaded 2024-11-09; delivered 2024-10-03–04 | official 58:53 video; reliability A; access partial | Japanese automatic captions; timestamped with machine-caption caveat.
• ★★★ | Milken Institute official YouTube — Japan Symposium closing plenary | https://www.youtube.com/watch?v=bU0EWl9-9NA | uploaded 2019-07-09; event 2019-03-25 | official duplicate video; reliability A; access full | English automatic-caption timestamps; same work as media 21333.
• ★★★ | Tokyo Forum official YouTube — Son/Jack Ma dialogue | https://www.youtube.com/watch?v=g9id48oj0Bg | event 2019-12-06; archive 2019-12-17 | official 60:22 video; reliability A; access full | English automatic-caption timestamps; one underlying dialogue.
• ★★★ | NVIDIA official YouTube — Son and Jensen Huang at AI Summit Japan | https://www.youtube.com/watch?v=x8O6ChAWBxs | 2024-11-14 | official video with creator-provided English captions; reliability A; access full | Exact timestamps from creator WebVTT; Son/Huang/interviewer voices separated.

SCOUT — identity, corpus, vehicles, transactions, and current AI controls (30 high-priority routes; all processed, including 11 SYNTH controls)

• ★★★ | SoftBank Group — Biography: Masayoshi Son | https://group.softbank/en/about/officer/son | current as of 2026-06-24 | official biography; reliability A; access full; accessed 2026-08-22 | Birth, 1981 founding, current title, and dated career/directorship controls; promotional omissions remain possible.
• ★★★ | SoftBank Group — History | https://group.softbank/en/philosophy/history | 1981–present | official corporate chronology; reliability A; access full; accessed 2026-08-22 | Era/date control for operating entries, acquisitions, investments, and listings; not proof of Son-only authorship or return.
• ★★★ | SoftBank Group — Annual Reports archive | https://group.softbank/en/ir/financials/annual_reports | FY1998–FY2026 | official corpus index; reliability A; access full; accessed 2026-08-22 | Denominator control for 29 report years and canonical message/PDF routes.
• ★★★ | SoftBank Group — Annual Report 2018 PDF | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2018_01_en.pdf | report year 2018 | official report alternate; reliability A; access full; accessed 2026-08-22 | Corporate alternate route retained for context; does not contain the canonical Son-message HTML and is not an extra WRITINGS work.
• ★★★ | SoftBank Group — Annual Report 2021 | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2021_01_en.pdf | report year 2021 | audited/corporate report; reliability A; access full; accessed 2026-08-22 | SB Northstar at PDF p. 86 / printed p. 82; SBG 67%, Son 33% indirectly, pro-rata repayment obligation.
• ★★★ | SoftBank Group Report 2022 | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2022_en.pdf | report year 2022 | audited/corporate report; reliability A; access full; accessed 2026-08-22 | SVF1/SVF2/LatAm structure, SBIA/SBGA, MASA USA, preferred 8% distribution, scoped 17.25% program interest, and team/vehicle controls.
• ★★★ | SoftBank Group Report 2023 | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2023_en.pdf | report year 2023 | audited/corporate report; reliability A; access full; accessed 2026-08-22 | CFO message printed p. 18: Alibaba voting ownership 24.4% to 13.7%, primarily physical settlement of prepaid forwards.
• ★★★ | SoftBank Group Report 2025 | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2025_en.pdf | report year 2025 | audited/corporate report; reliability A; access full; accessed 2026-08-22 | WeWork table PDF p. 93 / printed p. 91: $6.71B reclassified losses, $3.18B SVF1 plus $3.53B SVF2; no $4.4B total.
• ★★★ | SoftBank Group Report 2026 | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2026_en.pdf | report year 2026 | audited/corporate report; reliability A; access full; accessed 2026-08-22 | Current AI holdings, commitments, financing, risks, governance, and group/vehicle denominator.
• ★★★ | SoftBank — Completion of Acquisition of Sprint | https://group.softbank/en/news/press/20130711 | closing 2013-07-10; release 2013-07-11 | official completion release; reliability A; access full; accessed 2026-08-22 | Approximately 78% of New Sprint fully diluted, excluding out-of-the-money options; supersedes initial 70% structure.
• ★★★ | U.S. DOJ — settlement permitting T-Mobile/Sprint merger subject to remedy package | https://www.justice.gov/archives/opa/pr/justice-department-settles-t-mobile-and-sprint-their-proposed-merger-requiring-package | 2019-07-26 | official regulator release; reliability A; access full; accessed 2026-08-22 | Regulatory/remedy control; does not establish operating or investment success.
• ★★★ | FCC — T-Mobile/Sprint transaction approval with conditions | https://www.fcc.gov/document/fcc-approves-t-mobilesprint-transaction-conditions | 2019-11-05 | official regulator order route; reliability A; access full; accessed 2026-08-22 | Communications-regulatory condition and dissent record; separate from DOJ and state litigation.
• ★★★ | T-Mobile — T-Mobile and Sprint are one company | https://www.t-mobile.com/news/un-carrier/t-mobile-sprint-one-company | 2020-04-01 | official closing announcement; reliability B; access full; accessed 2026-08-22 | Counterparty closing-date/management control; use filings for securities and consideration.
• ★★★ | Arm Holdings — final 2023 prospectus (424B4) | https://www.sec.gov/Archives/edgar/data/1973239/000119312523235320/d550931d424b4.htm | 2023-09-14 | SEC filing; reliability A; access full; accessed 2026-08-22 | SBG expected ownership 90.6%, or 89.9% with full option; controlled-company and offering-risk evidence.
• ★★★ | Arm Holdings — registration statement (F-1) | https://www.sec.gov/Archives/edgar/data/1973239/000119312523211673/d393891df1.htm | 2023-08-21 | SEC filing; reliability A; access full; accessed 2026-08-22 | Pre-offering business, ownership, related-party, risk, and historical financial control.
• ★★★ | Alibaba Group — registration statement (F-1) | https://www.sec.gov/Archives/edgar/data/1577552/000119312514184994/d709111df1.htm | 2014-05-06 | SEC filing; reliability A; access full; accessed 2026-08-22 | Offering-era ownership, governance, variable-interest-entity, and risk control; not Son's cash-return ledger.
• ★★★ | Alibaba Group — final IPO prospectus (424B4) | https://www.sec.gov/Archives/edgar/data/1577552/000119312514333674/d709111d424b4.htm | 2014-09-19 | SEC filing; reliability A; access full; accessed 2026-08-22 | Final offering terms and principal-shareholder control; distinguish market value from realization.
• ★★★ | The We Company — registration statement (S-1) | https://www.sec.gov/Archives/edgar/data/1533523/000119312519220499/d781982ds1.htm | 2019-08-14 | SEC filing; reliability A; access full; accessed 2026-08-22 | Decision-near governance, losses, related parties, capitalization, and risk counterevidence before aborted IPO.
• ★★★ | Kroll Restructuring Administration — WeWork Chapter 11 cases | https://restructuring.ra.kroll.com/wework/ | 2023–2024 | court-appointed claims/docket portal; reliability A for filed records; access full; accessed 2026-08-22 | Route to petitions, plan, disclosure statement, confirmation, emergence, and security cancellation.
• ★★★ | SoftBank — establishment of SoftBank Vision Fund | https://group.softbank/en/news/press/20161014 | 2016-10-14 | official fund announcement; reliability A; access full; accessed 2026-08-22 | Initial SBG/PIF capital plan and vehicle purpose; planned terms are not final close.
• ★★★ | SoftBank Vision Fund — first major close announcement PDF | https://www.softbank.jp/corp/d/group_news/press_20170520_01_en.pdf | 2017-05-20 | official fund release; reliability A; access full; accessed 2026-08-22 | More than $93B commitments and named LPs; fund capital not Son personal capital.
• ★★★ | SoftBank — completion of Vision Fund first closing | https://group.softbank/en/news/press/20170522 | 2017-05-22 | official corporate release; reliability A; access full; accessed 2026-08-22 | Corporate route to first close and LP roster; distinct publisher artifact, same close event.
• ★★★ | SoftBank — Launch of SoftBank Vision Fund 2 | https://group.softbank/en/news/press/20190726 | 2019-07-26 | official planned-fund release; reliability A; access full; accessed 2026-08-22 | $108B expected commitments based on MOUs and $38B intended SBG contribution; intentions are not final subscribed capital.
• ★★★ | OpenAI/SoftBank — advanced enterprise AI and Cristal intelligence partnership | https://group.softbank/en/news/press/20250203_0 | 2025-02-03 | joint official release; reliability A; access full; accessed 2026-08-22 | $3B annual SBG deployment claim, proposed 50:50 JV structure, and distinct roles for OpenAI/SBG/SoftBank Corp./Arm.
• ★★★ | SoftBank/OpenAI — Announcing the Stargate Project | https://group.softbank/en/news/press/20250122 | 2025-01-22; announcement 2025-01-21 | joint official release; reliability A; access full; accessed 2026-08-22 | $500B four-year intention, initial equity funders, SBG financial/OpenAI operational split, Son chair; not funded-spend evidence.
• ★★★ | SoftBank — Announcement Regarding Investment in OpenAI | https://group.softbank/media/Project/sbg/sbg/news/press/2025/20250401/pdf/20250401_0_en.pdf | agreement 2025-03-31; release 2025-04-01 | official transaction PDF; reliability A; access full; accessed 2026-08-22 | Up-to-$40B structure, syndication, conditions, delegation, borrowing, FVTPL accounting; commitment is not realized return.
• ★★★ | SoftBank — Completion of Additional $22.5 Billion Investment in OpenAI | https://group.softbank/en/news/press/20251231 | 2025-12-31 | official completion release; reliability A; access full; accessed 2026-08-22 | Tranche/vehicle/co-investor and approximately 11% aggregate ownership control.
• ★★★ | SoftBank — Follow-on Investments in OpenAI | https://group.softbank/en/news/press/20260227 | 2026-02-27 | official transaction announcement; reliability A; access full; accessed 2026-08-22 | Three planned $10B tranches via SVF2 and bridge-finance disclosure; Yoshimitsu Goto delegation preserved.
• ★★★ | SoftBank — Acquisition of Ampere Computing Holdings LLC | https://group.softbank/en/news/press/20250320_0 | 2025-03-20 | official transaction notice; reliability A; access full; accessed 2026-08-22 | $6.5B price, SB6/Summit Bright structure, 8.08% prior holding, financing, advisers, fairness opinion, board/Son delegation, regulatory conditions.
• ★★★ | Wayve — $1.05B Series C led by SoftBank | https://wayve.ai/press/series-c/ | 2024-05-07 | company transaction release; reliability B; access full; accessed 2026-08-22 | Names NVIDIA/Microsoft and stated embodied-AI use of proceeds; company voice and amount require financing records for return analysis.

SCOUT — company, fund, team, and AI context (17 medium-priority routes; all processed, including 14 SYNTH controls)

• ★★ | SoftBank Group — Corporate Data | https://group.softbank/en/about/profile | current; accessed 2026-08-22 | official company profile; reliability A; access full | Legal identity, headquarters, capital, fiscal year, and group description; current-state control only.
• ★★ | SoftBank Group — Corporate Philosophy, Vision & Values | https://group.softbank/en/philosophy | current; accessed 2026-08-22 | official doctrine hub; reliability A; access full | Corporate mission/values route; institutional voice is not automatically Son-sole authorship.
• ★★ | SoftBank Group — Investor Relations | https://group.softbank/en/ir | current; accessed 2026-08-22 | official IR index; reliability A; access full | Routing hub for financials, NAV, debt, presentations, governance, shareholders, and securities documents.
• ★★ | SoftBank Group Report 2024 | https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/annual_reports/annual-report_fy2024_en.pdf | report year 2024 | audited/corporate report; reliability A; access full; accessed 2026-08-22 | Interim AI-stack, Arm, fund, financing, and risk control between the 2023 IPO and 2025 concentration.
• ★★ | SoftBank Group — SoftBank Vision Funds segment | https://group.softbank/en/segments/svf | current; accessed 2026-08-22 | official segment overview; reliability A; access full | Manager/vehicle overview; marketing framing requires audited fund tables.
• ★★ | SoftBank Vision Fund — Portfolio | https://visionfund.com/portfolio | current; accessed 2026-08-22 | manager-controlled portfolio index; reliability B; access full | Company/status discovery denominator for SVF1/SVF2; not cost, realized-return, or Son-decision evidence.
• ★★ | SoftBank Group — Segments and fund-entity map | https://group.softbank/en/segments | current; accessed 2026-08-22 | official segment/group routing page; reliability A; access full | Identifies reportable segments and routes the Latin America Fund; keeps SBLA/SBGA and teams separate from SVF1/SVF2.
• ★★ | Graphcore — Graphcore joins SoftBank Group to build next generation of AI compute | https://www.graphcore.ai/posts/graphcore-joins-softbank-group-to-build-next-generation-of-ai-compute | 2024-07-11 | company acquisition announcement; reliability B; access full; accessed 2026-08-22 | Company/team rationale and completion claim; price and SBG economics require filings.
• ★★ | OpenAI/Oracle/SoftBank — Stargate expands to five new U.S. sites | https://group.softbank/en/news/press/20250924 | 2025-09-24; announcement 2025-09-23 | joint official release; reliability A; access full; accessed 2026-08-22 | Planned capacity/investment progress; partner claim, not audited spend or return.
• ★★ | Global Tech Alliance — Stargate UAE | https://group.softbank/en/news/press/20250522 | 2025-05-22 | multi-party official release; reliability A; access full; accessed 2026-08-22 | G42-build/OpenAI-Oracle-operate/SBG-partner/NVIDIA-supply role boundaries and planned 2026 first cluster.
• ★★ | SoftBank/OpenAI — launch of SB OAI Japan JV | https://group.softbank/en/news/press/20251105 | 2025-11-05 | joint official release; reliability A; access full; accessed 2026-08-22 | Separate legal entity, ownership, operating and go-to-market roles; current-state control.
• ★★ | SoftBank/Ampere — SoftBank Group to Acquire Ampere Computing | https://group.softbank/en/news/press/20250320 | 2025-03-20 | joint short-form announcement; reliability A; access full; accessed 2026-08-22 | Management quotations and $6.5B headline; detailed transaction notice is the economic control.
• ★★ | Yoshimitsu Goto — CFO Message, SoftBank Group Report 2025 | https://group.softbank/en/ir/financials/annual_reports/2025/message/goto | report year 2025 | signed CFO message; reliability A; access full; accessed 2026-08-22 | Team-credit and financing control for OpenAI follow-ons, LTV/liquidity, and NAV; not Son voice.
• ★★ | SEC — Arm Holdings issuer filings index | https://www.sec.gov/edgar/browse/?CIK=1973239&owner=exclude | 2023–present | official filing index; reliability A; access full; accessed 2026-08-22 | Route to 20-F, 6-K, prospectus and ownership disclosures; use exact filing/date for later claims.
• ★★ | SEC — T-Mobile US issuer filings index | https://www.sec.gov/edgar/browse/?CIK=1283699&owner=exclude | merger era–present | official filing index; reliability A; access full; accessed 2026-08-22 | Route to Sprint-merger consideration, SoftBank stake changes, contingent shares and later sales.
• ★★ | SoftBank — FY2024 Q3 earnings presentation | https://group.softbank/system/files/pdf/ir/presentations/2024/investor-presentation_q3fy2024_01_en.pdf | 2025-02 | official investor deck; reliability A; access full; accessed 2026-08-22 | Investment-count and Wayve/OpenAI timing controls; IC-approved-but-unfunded items remain distinct.
• ★★ | SoftBank Group — Biography: Yoshimitsu Goto | https://group.softbank/en/about/officer/goto | current; accessed 2026-08-22 | official executive biography; reliability A; access full | CFO/team attribution route for delegated financing and investment decisions; not Son biography.