Martin Casado

1. Snapshot and testable edge thesis

Predictions. [researcher inference · strong] If the edge thesis is true, unseen dated market maps, memos, customer-reference files, IC records, and board materials should show the same questions before outcomes: the unstable technical boundary; user, buyer, and budget; alternative architectures; paid control point; contribution margin; platform/supplier power; reversible versus sunk cost; founder competency gaps; scenario probabilities; and explicit kill/next-check milestones. The pattern should recur in ordinary and failed cases, not only selected winners, and should predict stronger outcomes after controlling for access, team, stage, price, ownership, sector, vehicle, and vintage. deal sourcing forensics · missing-record ledger

Falsifiers. [researcher inference · strong] The thesis weakens materially if the missing cohort shows that (1) sourcing came mainly from generic warm access without a distinctive technical map; (2) usage/customer evidence did not change rankings, prices, or passes; (3) successful cases were partner-led while Casado-attributed ordinary cases did no better than a matched baseline; (4) category boundaries were redrawn after outcomes; (5) follow-ons ignored the published unit-economics and no-new-capital tests; or (6) complete net ex-Cursor cash flows do not outperform appropriate vintage/stage benchmarks. dominant-outlier test

2. Core philosophy

What Casado actually says

• Architecture creates the search space. [investor-stated · strong] Across infrastructure talks and essays, Casado treats changes in abstraction—programmable networks, APIs, microservices, cloud specialization, data stacks, foundation models, and agents—as openings for new control points and companies. Infrastructure… Is Everything, Takeaways · Emerging Architectures, opening through “Blueprints” · Who Owns the Generative AI Platform?, stack map
• Usage can overturn a beautiful theory. [documented behavior · moderate] Casado recounts that Hugging Face usage challenged an internally reasoned view. The account supports a correction loop, but no dated memo, opportunity, terms, or vote proves that the correction produced an investment result. Past, Present, and Future of Digital Infrastructure, 06:49–34:18
• Product adoption and economic ownership are separate gates. [investor-stated · strong] Developer or user love can insert a product, while a buyer, budget, procurement path, commercial control point, margin, and ultimately fund ownership determine whether adoption can become a venture outcome. Growth+Sales, “New buyer, new motion” and Failure Modes #1–#6 · Monopolies vs. Oligopolies in AI, 53:38–57:09
• Uncertainty should be preserved long enough to learn. [investor-stated · strong] The written corpus favors simple early packaging, flexible cloud/API capacity, founder-led market discovery, and deliberate specialization only after workload, buyer, and repetition become clearer. Bottom Up Pricing & Packaging, user-journey framework · Market Annealing, operating stages · The Cost of Cloud, repatriation framework
• Non-consensus is conditional, not ceremonial. [investor-stated · strong] Casado distinguishes product/customer evidence that contradicts current investor belief from ignoring informed objections or the financing audience a company will soon need. His stated Series A and total-exposure framework makes the rule stage- and vehicle-dependent. Is Non-Consensus Investing Overrated?, 00:00:54–00:11:04, 00:28:20–00:32:46, 00:52:10–00:53:06
• Models must be revised when economics change. [investor-stated · strong] The 2020 warning that AI companies can carry compute, data, review, and deployment burdens becomes a 2023 conditional discontinuity and a 2025 margin-bridge argument using routing, tiering, enterprise mix, and product feedback. The durable rule is full-cost measurement, not permanent AI pessimism or optimism. 2020 AI essay, “Gross Margins” and “Scaling AI systems” · 2023 economic case · 2025 margins essay

Synthesis

[researcher inference · strong] Casado's operating system is best understood as disciplined translation: turn a technical change into falsifiable company economics, then allow customer evidence to change the map. Its strength is that it connects architecture, product, go-to-market, cost, and capital rather than stopping at technical novelty. Its weakness is that almost every public example comes from an interested, positively selected record; coherence is not causal proof. written core · spoken evolution · denominator boundary

The method has five gates:

1. Map the discontinuity. [researcher inference · strong] Identify which abstraction is moving, what becomes newly possible, and which incumbents or platforms gain or lose bargaining power. written synthesis
2. Locate the indispensable workflow. [researcher inference · strong] Test whether the product reaches value, is retained, spreads, and controls a meaningful part of the user's work. Growth+Sales
3. Trace user to budget. [researcher inference · strong] Name the economic buyer, paid trigger, procurement/security bridge, pricing metric, and expansion mechanism. Bottom Up Pricing & Packaging
4. Underwrite the whole stack. [researcher inference · strong] Include cloud, inference, data, evaluation, support, services, supplier power, migration, and capital needs rather than assuming software economics. 2020 AI essay · AI compute essay
5. Match evidence to capital and governance. [researcher inference · moderate] Set the next check, ownership aspiration, leadership needs, conflicts, and board role to the remaining uncertainty. The public doctrine supports this gate; deal-specific rights and effects are mostly unknown. Tecton lab · vehicle boundary

3. Sourcing

How opportunities appear

Sourcing rule. [researcher inference · strong] For every opportunity, record the first-contact source, date, economic relationship, information rights, sponsor, independent disconfirming references, and whether the market map existed before the company appeared. Kill the “sourcing edge” claim if the map was reconstructed after the winner or if related portfolio references supplied all decisive evidence. sourcing forensics · source coverage

4. Picking

Selection and underwriting hierarchy

[researcher inference · strong] The public record supports a hierarchy, not a weighted formula. written decision playbook · spoken decision evidence

1. Technical discontinuity and control point. Can the product own an indispensable workflow, interface, standard, data path, or economic bottleneck as architecture shifts? RapidAPI · Isovalent · Cursor
2. Customer truth. Can users reach value, return, spread, and produce references without a sales organization manufacturing the signal? Growth+Sales, Failure Modes · Tackle lab
3. Buyer and budget. Who owns the budget, which trigger justifies paid conversion, and what procurement, compliance, security, or integration path stands between usage and revenue? Bottom-up pricing, framework
4. Whole-product economics. What is contribution margin per successful outcome after inference, human review, implementation, support, failure/retry, cloud, and supplier costs? The New Business of AI, cost taxonomy
5. Founder and leadership fit. Does the team have the technical depth, commercial learning speed, recruiting power, and domain competence for the current bottleneck? Casado's own CEO-transition account and robotics-domain admission argue against a universal technical-founder rule. On Fear and Leadership, Takeaways · Entering Uncharted AI Territory, Takeaways
6. Financing survival and consensus. Can the company reach the next independently valuable milestone within a no-new-capital scenario, and will the next financing audience understand the evidence? Non-Consensus, Takeaways · Tackle company update, overbuilt/runway paragraphs

Diligence questions

• Market. [researcher inference · strong] What exact submarket is being underwritten; what alternative definitions would reverse the conclusion; and what customer budget is forming or disappearing? architecture-map evidence
• Product. [researcher inference · strong] What independent value moment, retained cohort, within-logo spread, reliability threshold, and switching behavior demonstrate workflow control? Growth+Sales, failure modes
• Economics. [researcher inference · strong] What is quality-adjusted contribution margin; which costs are variable, delayed, subsidized, supplier-set, or hidden in services; and what must improve by the next check? New Business of AI, cost taxonomy
• Moat. [researcher inference · strong] Does the company own scarce data, integration, distribution, brand, marketplace liquidity, open-standard governance, or cost advantage after model/platform parity? Raw data volume and open-source stars are insufficient. Empty Promise of Data Moats, “A practical framework” · Software Chasms, Takeaways
• People. [researcher inference · moderate] Which job must the founder remain uniquely responsible for, what leadership role must change, and what evidence would support that transition? The public corpus supplies self-reported Nicira learning, not a general causal cohort. Bringing in an External CEO, “The technical founder maze”
• Decision. [researcher inference · strong] Name sponsor, coauthors, objector, alternatives, vote/veto, conflicts, vehicle, check, security, ownership, reserves, board right, kill criteria, and next-check milestones. Public bylines and board rosters do not fill these fields. attribution rule

Picking boundary. [researcher inference · insufficient public record] No original IC memos, scoring thresholds, votes, price walk-aways, rejected companies, or matched outcome cohort establishes the actual weighting of these gates or Casado's legal decision rights. denominator boundary · missing-record ledger

5. Deal and portfolio mechanics

Price, ownership, securities, reserves, and concentration

Portfolio rule. [researcher inference · strong] Do not construct a portfolio from an architecture diagram. For each control-point thesis, cap correlated exposure to the same hyperscaler, model vendor, data source, buyer, regulation, or financing regime; map “mortal enemy” conflicts before investment; and re-underwrite every follow-on against the best new opportunity and a no-new-capital case. The need for this rule is strong; evidence that Casado/a16z actually applies it company by company is insufficient. conflict convention

Return proof boundary. [researcher inference · strong] Gross MOIC needs invested cost plus realized and residual value. DPI/TVPI need LP paid-in capital, distributions, and NAV after fees, expenses, and carry. IRR also needs dated calls and distributions. Apart from Yubico's proposed 6.1% combined-company closing snapshot, the record lacks complete ownership histories; it also lacks deal-to-vehicle mapping, securities, dilution paths, tenders, residual NAV, fee/carry allocation, recycling, subscription-line effects, and dated cash flows. Holding Cursor's $60B merger-implied company value constant, both an immaterial high-cost diluted stake never distributed and a material low-cost stake promptly distributed fit the evidence. Therefore no unique MOIC, DPI, TVPI, or IRR is publicly computable. Yubico/ACQ Bure release, “Ownership structure” table · SpaceX 8-K, Item 1.01 · formal non-computability proof

6. Vehicle and incentive fit

Authority, compensation, stakeholders, and continuity

• Practice leadership. [contemporaneous record · strong] The current profile says Casado leads Infrastructure. a16z profile, opening role
• Rounded oral authority claim. [investor-stated · moderate] Casado describes running an approximately $1.2B pool; this is a rounded self-description, not legal-control evidence. Monopolies vs. Oligopolies, 54:51–57:09
• Legal control boundary. [contemporaneous record · strong] Fund VII and IX/X filings and the ADV name Marc Andreessen and Ben Horowitz in the GP/control chain. ADV, Schedules A/B [researcher inference · strong] “Leads” does not prove unilateral allocation, legal vote, veto, ownership, or carry; absence from ownership schedules does not prove Casado lacks an IC vote, delegated budget, informal influence, or economics.
• Fees and carry. [contemporaneous record · strong] Fund IX/X Forms D say a GP or affiliate is entitled to a management fee under governing documents. They do not provide a usable rate, base, duration, offsets, hurdle, or carry split. Any Casado carry or compensation claim is therefore insufficient public record. Fund IX, Item 15 · Fund X, Item 15
• LP incentive. [researcher inference · strong] LPs supply pooled capital and bear vehicle-level fees, valuation, duration, and liquidity risk. Industry cumulative negative LP cash flow of $196.9B from 2022 through 2025-06-30 raises the value of realized-cash evidence; it does not reveal any a16z vehicle's DPI. Q4 2025 Venture Monitor, Executive Summary and liquidity sections
• Founder incentive and conflicts. [investor-stated · strong] Casado describes a “one mortal enemy” convention and acknowledges that pivots can create later conflicts. This may protect founder candor while restricting the fund's future search set. No general recusal, information-barrier, valuation, cross-vehicle allocation, or LPAC exception record is public. Monopolies vs. Oligopolies, 57:09–59:35
• Geography and regulation. [researcher inference · insufficient public record] The public corpus does not establish decision-date domicile, geographic authority, cross-border allocation, or company-specific regulatory restrictions consistently enough to infer a geographic mandate. The ADV and Form Ds establish U.S. adviser/vehicle facts, not deal-level authority. ADV, Item 5.F and Schedules A/B · vehicle boundary
• Duration and liquidity. [researcher inference · moderate] Large successor pools can finance longer company-building and follow-ons, while private marks, cross-vehicle allocations, and weak exit markets can delay LP cash. Without governing documents and cash flows, “patient capital” is an untested label. vehicle analysis · historical context
• Succession and key person. [researcher inference · insufficient public record] The package supplies no key-person clause, succession plan, replacement rule, or test of whether the Infrastructure process persists without Casado, Andreessen, Horowitz, or the current specialist team. formal-authority boundary · missing-record ledger

7. Team attribution and work with founders

People, capital, and influence

Working with founders

• Board posture. [investor-stated · moderate] Casado describes a board member's role as bringing market maps and scarce functions rather than issuing operating commands. That is a stated boundary, not a board-effect study. 20VC, 31:10–33:20
• Leadership fit. [retrospective witness · moderate] His Nicira account of declining the CEO role, regretting an interim choice, and hiring Steve Mullaney supports the principle that the required leader changes with company phase. It lacks board, employee, customer, and Mullaney corroboration and cannot establish the counterfactual. On Fear and Leadership, biography evidence
• Founder-led sales. [retrospective witness · provisional] Casado says founders should remain involved in strategic early sales and category learning. Public evidence does not identify the deals, bookings, counterfactual close rate, or board actions that caused later operating results. On Fear and Leadership, Takeaways · Market Annealing
• Recruiting and intervention. [researcher inference · insufficient public record] The 25 board claims create plausible access to recruitment and intervention, but no company-by-company record establishes who recruited, evaluated, backed, replaced, or constrained management. people map
• Ethical governance. [researcher inference · strong] Product use, customer relationships, board information, competitor mapping, co-investment, valuation, and public advocacy can improve diligence and bias it. Each deal should disclose related roles, permissible information use, recusals, competitor lanes, affected stakeholders, and who owns the decision. No public record proves that a16z's stated conflict conventions achieved those outcomes. allocation and conflict evidence · policy evidence boundary

8. Views on the craft

Craft maxim. [researcher inference · strong] A good venture narrative says what could be true; a good investment process specifies what would make it false, who owns the decision, what capital is at risk, and which evidence earns the next check. Non-Consensus note, takeaways · uncertainty ledger

9. Signature bets and decision process

These are cross-links to the no-hindsight labs, not outcome-based reconstructions of private IC logic. Every proposed alternative, kill test, and next-check rule is a researcher inference unless the cited decision-date source states it; support is graded claim by claim below.

Shared decision-rights boundary. [researcher inference · strong] None of the seven cases discloses the complete alternative set, dissent, IC vote, deal-specific vehicle, check, security, ownership, reserves, board conduct, exit decision, or documented post-outcome process change. Calendar year must not be used to guess the vehicle. seven decision labs · missing-record ledger

10. Mental models and maxims

This is the reconciled navigable inventory for the companion field guide. Seven combined rules survive the evidence bar; eight attractive candidates are explicitly absorbed or excluded in the field guide rather than promoted from slogans, forecasts, or too-few applications. [researcher inference · strong] candidate-model ledger

11. Evolution and contradictions

Dated evolution

Contradictions to preserve

Evolution boundary. [researcher inference · strong] Public changes in rhetoric do not establish changed IC or board behavior. The only defensible claim is that Casado's expressed framework widened from technical architecture toward customer evidence, full-stack economics, ownership, conflicts, capital intensity, and policy. Process-change claims require dated internal records. written evolution · spoken evolution

12. Failures and limits

Named adverse evidence

Failure handling protocol

1. Name the unit that failed. [researcher inference · strong] Product experiment, operating plan, technical milestone, whole company, security, fund position, and LP outcome are different. failure forensics
2. Reopen the decision-date thesis. [researcher inference · strong] Compare original signals, unknowns, alternatives, dissent, price, ownership, and kill criteria with the reveal; do not use later information to rewrite the entry rationale. failure forensics
3. Separate response from result. [researcher inference · strong] Cost cuts, management changes, product integration, design fixes, follow-ons, and acquisitions are actions, not proof they restored value. failure forensics
4. Record capital and governance decisions. [researcher inference · strong] Identify who authorized bridge, reserve, hiring reset, leadership change, sale, shutdown, or refusal—and what stakeholder obligations remained. failure forensics
5. Change the process only when documented. [researcher inference · strong] No public case establishes a Casado/a16z postmortem that changed an IC, reserve, or board rule. The downside re-underwrite and other process changes here are analyst rules, not investor-stated reforms. failure forensics

Failure boundary. [researcher inference · strong] The actual loss, write-off, pass, reserve-refusal, and below-cost-exit denominator is absent. “No adverse event surfaced” for a private company is unknown, not clean. The positive announcement corpus cannot support a personal hit rate, failure rate, or claim of superior learning. failure forensics · denominator boundary

13. Historical operating environment

Base-rate boundary. [researcher inference · strong] These industry observations are not matched to the Casado-linked companies by vintage, stage, geography, ownership, risk, or cash-flow timing. They explain alternative causes and changed constraints; they do not establish alpha. Contemporaneous decision analysis must use the data available then, while later statistics remain explicitly retrospective.

14. Modern VC translation — as of 2026-08-20

Each row carries a claim-local two-axis label. “Grounded example” means a current fact in the completed package, not a claim that Casado applied the rule or earned a return.

15. Comparison with completed peers

Comparator boundary. [researcher inference · strong] Only three completed repository peers are defensible. Trae Stephens is the closest current operator/founder-investor comparison for technical access, conflicts, private marks, and missing returns. Georges Doriot and Reid Dennis supply long-duration vehicle, active-governance, technical-translation, and outlier tests, but are not technology-era or modern private-fund matches. Marc Andreessen, Ben Horowitz, and Peter Thiel are not completed canon peers and are excluded from comparison.

16. The strongest case against greatness

Skeptical thesis. [researcher inference · provisional] Casado's visible reputation may be the product of exceptional technical and institutional access, a16z's team/platform/capital, a favorable infrastructure-and-AI regime, selected public narratives, and one dominant company event—without enough personal roles, ownership, failures, or net cash flows to establish repeatable personal outperformance. denominator · people map · dominant-outlier test

1. One dominant visible company event. [researcher inference · strong] Cursor is the only priced person-linked company outcome above $1B and dwarfs the other disclosed values. Ex-Cursor visible company-event value is $1.132729B across Yubico and Parsec; unpriced events cannot enter the arithmetic. This says nothing about stake, cost, or return. dominant-outlier test
2. Access may precede selection. [researcher inference · strong] Stanford/Nicira credibility, a16z's brand, portfolio customers, board rooms, internal product use, and specialist colleagues can supply opportunities and information unavailable to an ordinary investor. No matched cold-source or access-controlled cohort exists. sourcing forensics
3. Team and vehicle misattribution. [researcher inference · strong] Cursor has four named thesis authors; newer AI deals are frequently coauthored; Andreessen and Horowitz appear in legal control records; founders and boards do the operating work. A Casado byline or practice-lead title does not prove sole sponsorship, vote, ownership, or outcome causality. people map
4. Missing actual denominator. [researcher inference · strong] Forty-six public-attribution companies, 43 reconstructed, and 31 without qualifying primary outcomes do not reveal the actual investments, passes, losses, write-offs, reserve refusals, or private positions. Career-wide coverage and hit rate are not computable. denominator
5. No compatible economics. [researcher inference · strong] Company consideration, listing, private financing, operating metrics, investor proceeds, and LP distributions remain distinct. No case supplies cost, security, ownership, dilution, dated proceeds, and distributions together; MOIC, DPI, TVPI, and IRR are unidentified. non-computability proof
6. Survivorship and narration. [researcher inference · strong] Investment announcements explain why the firm invested; failed theses, losing categories, passes, diluted winners, and reserve refusals are less publishable. Flexible categories can make later winners look inevitable. official announcement corpus · denominator boundary
7. Doctrine may exceed behavior evidence. [researcher inference · strong] The public corpus coherently teaches customer truth, full-stack economics, market maps, and no-new-capital thinking, yet supplies no IC cohort showing that these tests actually controlled decisions, prices, or follow-ons. Tackle's overbuilding is direct counterevidence to easy claims that the doctrine always governed capital. written core · Tackle adverse lab
8. Era effects. [researcher inference · strong] Cloud programmability, open source, cheap 2021 capital, 2022 reset, AI capability jumps, platform competition, and 2026 AI funding concentration can create both opportunity and extreme marks independent of personal skill. historical environment
9. Conflicts and interested advocacy. [researcher inference · strong] Portfolio/customer/board access, competitor lanes, related valuations, public AI advocacy, and regulation positions can sharpen information and bias diligence or narration. Stated separate teams and the “one mortal enemy” convention are not audited controls. allocation and conflict evidence · policy evidence boundary
10. Ordinary and adverse cases remain economically unresolved. [researcher inference · strong] Tackle, Rapid, Orbit, Astranis, Resend, and Fly.io add real operating/technical evidence but not fund losses, recovery, governance causality, or returns. “No adverse event found” for the rest is not success. failure and adverse record
11. Generalization beyond core competence is unproved. [researcher inference · moderate] Space, robotics, physical AI, policy, and model-market claims may benefit from systems thinking but require domain, manufacturing, regulatory, safety, and affected-party evidence not supplied by infrastructure experience alone. Astranis competence-boundary lab · Uncharted note
12. Succession and transferability are unknown. [researcher inference · strong] The proposed edge may depend on Casado, a16z's current coauthors, the portfolio network, and Andreessen/Horowitz-controlled structures. No public key-person or post-Casado continuity test exists. vehicle and authority boundary · current institutional identity

What would decide between favorable and skeptical theses

[researcher inference · strong] The favorable case strengthens if signed role ledgers, dated maps/memos, pass logs, votes, independent references, company-by-company conflict controls, board records, reserve decisions, and full fund cash flows repeatedly connect Casado-specific actions to superior ordinary-case and net ex-Cursor outcomes after controlling for access, coauthors, stage, price, ownership, sector, vehicle, and vintage. The skeptical case strengthens if those records show mainly warm-network sourcing, generic post hoc categories, partner-heavy decision/execution, customer references without independence, doctrine-inconsistent follow-ons, weak ex-outlier net economics, or stakeholder/conflict costs hidden by private marks. uncertainty and retrieval ledger

17. Analyst study guide, glossary, and go deeper

Recommended reading path

1. Formation and attribution: read the sourced chronology and learning path, then the Stanford/LLNL/Nicira sources before using operator experience as an investing cause.
2. Core operating system: read Software Chasms, 20VC — Venture Model, and Past, Present, and Future of Digital Infrastructure.
3. Written economics: read Data Moats, The New Business of AI, Growth+Sales, Cost of Cloud, and Market Annealing.
4. Replay decisions without hindsight: do Tackle first as the failure/miss lab, then Cursor, Astranis, and the remaining four labs. Answer invest/pass, kill criteria, ownership/terms, and next-check milestones before reading each reveal.
5. Audit economics and credit: read the denominator, vehicle boundary, people map, and outcome forensics.
6. Stress the modern doctrine: read Entering Uncharted AI Territory, Non-Consensus, Monopolies vs. Oligopolies, and Capital Flywheel alongside the skeptical case above.
7. Compare, do not canonize: compare only the completed Trae Stephens, Georges Doriot, and Reid Dennis packages.

Investor-specific glossary

Reusable sourcing and diligence checklist

Provenance. [researcher inference · strong] These fields consolidate the supported sourcing, picking, model, role, and vehicle boundaries; they are analyst procedures, not a recovered a16z checklist. sourcing · picking · mental models · people and vehicle map

• Source — [researcher inference · strong]: first-contact person/date; map predating company; related portfolio/customer/board/angel role; permissible information; independent references; cold-source comparison.
• Problem/workflow — [researcher inference · strong]: user, job, frequency, severity, incumbent, independent value moment, retention, spread, switching, failure tolerance.
• Buyer/budget — [researcher inference · strong]: buyer, budget source, paid trigger, security/procurement/integration path, pricing metric, expansion, churn after repricing.
• Architecture/moat — [researcher inference · strong]: control point, alternatives, open/proprietary boundary, data marginal lift/rights, platform capture, supplier dependence, portability.
• Economics — [researcher inference · strong]: quality-adjusted contribution margin, services, review, inference/training, cloud/egress, CAC/payback, working capital, depreciation, no-new-capital runway.
• Team — [researcher inference · strong]: technical, commercial, operational, regulatory, and domain evidence; missing competency; recruiting; succession; role-change trigger.
• Decision — [researcher inference · strong]: sponsor, coauthors, objector, alternatives, vote/veto, prediction log, kill date, stakeholder impacts, conflicts/recusal.
• Terms — [researcher inference · strong]: vehicle, check, security, price, fully diluted ownership, preferences, dilution, pro rata, reserves, board/observer rights, information, downside, exit rights.

Reusable IC, follow-on, governance, and exit checklist

Provenance. [researcher inference · strong] These steps consolidate the seven no-hindsight labs, vehicle/return boundaries, failure handling, and model exception ledgers; they are analyst procedures, not documented Casado process changes. decision labs · vehicle boundary · return proof · model exception ledgers

1. Freeze the information set. [researcher inference · strong] Record what was known, unknown, believed, disputed, and available at the decision date; do not import later market or company facts.
2. Write both cases. [researcher inference · strong] Probability-weight at least two market structures and the strongest pass; name evidence and dates that invalidate each.
3. State invest/pass and kill tests. [researcher inference · strong] Include product, buyer, economics, technical, financing, regulatory, conflict, and ethical thresholds.
4. Price the security, not the story. [researcher inference · strong] Model ownership, dilution, preferences, reserves, follow-ons, downside, tender, and distribution; a company valuation is not fund return.
5. Release the next check on changed evidence. [researcher inference · strong] Compare against the best new opportunity, same-date market conditions, and a no-new-capital plan; do not use sunk cost or current mark as proof.
6. Govern legibly. [researcher inference · strong] Predefine board purpose, independent voice, recruiting/leadership triggers, information barriers, conflicts, recusal, incident escalation, stakeholder remedy, key-person continuity, and stop conditions.
7. Separate exit layers. [researcher inference · strong] Company agreement/consideration → holder allocation → lockup/escrow/tax → sale or in-kind security → fund proceeds → LP distribution. Do not collapse the chain.
8. Run the postmortem. [researcher inference · strong] Distinguish thesis error, execution, timing, financing, governance, and luck; record capital response, stakeholder cost, counterfactual, and actual process change.

Best sources, ranked for learning value

1. Software Chasms, full transcript note — densest bridge from technical architecture to product, open source, delivery, integration, and enterprise sales.
2. 20VC — Venture Model, full note — strongest direct account of maps, debate, stage/vehicle design, board boundary, and scaled-platform advocacy.
3. Past, Present, and Future of Digital Infrastructure, full note — clearest usage-over-theory correction and workload/cloud/open-standard analysis.
4. The New Business of AI — foundational full-cost taxonomy and the best baseline for testing later AI optimism.
5. Market Annealing — strongest founder-led discovery/scale mechanism, with historical-threshold caveats.
6. Is Non-Consensus Investing Overrated?, direct interview — most falsifiable account of informed objections, financing consensus, stage, ownership, and promised cohort testing.
7. Monopolies vs. Oligopolies in AI, direct podcast — strongest current ownership, conflict, vehicle, and alternative-market-structure source.
8. Complete deal forensics — the essential denominator, vehicle/role/economics boundaries, seven no-hindsight labs, outlier test, and adverse record.
9. VMware 2012 Form 10-K, Note B — best demonstration that announced value, net consideration, post-combination compensation, and personal proceeds are different measures; Nicira remains operator history only.
10. A16Z ADV, Schedules A/B and Schedule D §7.B.(1) — strongest public control/vehicle/GAV/valuation-process boundary; not performance evidence.

Highest-value unresolved evidence

Bottom line. [researcher inference · moderate] The record supports a teachable, falsifiable operating system: map technical discontinuities, let customer use correct theory, translate workflow into buyer/budget and full-stack economics, preserve option value, and match capital/governance to the remaining uncertainty. It does not yet support the stronger claim that Casado personally generated superior net returns. The decisive evidence is not another selected-winner story; it is the dated role/pass/loss/ownership/cash-flow/conflict record, tested on ordinary cases and ex-Cursor. written core · spoken evolution · return boundary

Pareto 80/20 — what drove the record

[researcher inference · strong] No literal 80/20 split, personal hit rate, DPI, TVPI, MOIC, IRR, or Casado-generated dollar total can be calculated from the public record. The searchable record combines 43 investor-authored or coauthored deal announcements, 25 current/former board claims, four Casado-announcement repeat-financing histories plus one issuer-reported seed-and-Series-A history, company acquisition consideration, an all-stock public-company transaction, private financing marks, and issuer-reported operating metrics. It does not disclose compatible a16z cost, security, ownership, dilution, seller allocation, dated proceeds, or LP distributions. The rankings below therefore answer three different questions and never add those measures together. a16z author page, current/former board paragraphs · a16z portfolio disclosure, disclosure below portfolio table · searched denominator

Economic contribution — observable company-outcome rank, not investor returns

Ranking rule. [researcher inference · strong] The eligible universe is the sufficiently reconstructed person-linked venture ledger, excluding Nicira and Illuminics because they are founder/operator outcomes. No eligible case supplies verified Casado- or a16z-specific proceeds, cost, ownership, dilution, and dated cash flows together. The ranking therefore uses a reproducible proxy: (1) a completed, priced company transaction outranks an unpriced exit or private financing; (2) cash consideration is identified separately from stock consideration but does not automatically outrank a much larger priced all-stock transaction; (3) stronger public Casado/team attribution breaks ties inside an outcome class; and (4) capital raised, revenue, and an unclosed mark are not substitutes for company value. This ranks observable company evidence, not MOIC, DPI, IRR, or dollar contribution. ledger · vehicle boundary

Dominant-outlier test. [researcher inference · strong] Cursor is the only publicly priced person-linked company outcome above $1 billion in the searched record and is roughly two orders of magnitude larger than the two other disclosed transaction measures. Removing Cursor leaves $1.132729 billion of disclosed visible company-event value: Yubico is 70.626% and Parsec 29.374%, so both are required to cross 80%. Unpriced acquisitions, the Fivetran–dbt all-stock merger with undisclosed exchange economics, and private companies cannot enter that arithmetic. This is a share of visible company-event value only, not a return series or a calculated share of Casado's or a16z's economic contribution. Yubico transaction · Parsec consideration

Historical importance — separate from economics

Ranking rule. [researcher inference · strong] Eligible cases require a direct Casado-authored/coauthored investment record. Rank lexicographically by (1) evidence that the case introduced a new public practice or doctrine, (2) persistence of that practice in later cases, (3) later institutional/company significance, and (4) stronger Casado-specific attribution; earlier direct records break remaining ties. This is a chronology-and-doctrine rank, not economics.

Decision-learning value — ranked for replayability

Ranking rule. [researcher inference · strong] Eligible cases require both a decision-date record and a later reveal. Rank lexicographically by (1) clarity and severity of the decision tension—including documented error before extreme positive-outcome or technical-adverse cases—(2) strength of the later reveal, (3) transferability across deals, and (4) attribution quality. Thus Tackle's admitted operating error ranks ahead of Cursor's extreme outcome, which ranks ahead of Astranis's narrower technical failure-and-response record. This ranks learning value, not success.

Evidence boundary and searched denominator

Unit of analysis. [researcher inference · strong] The reproducible unit is a public deal-linked claim record, then a company lineage. The search found 70 candidate claim records across 48 candidate companies. Two company candidates were excluded from the venture denominator: Nicira is Casado's founder/operator outcome, and Illuminics lacks a reconstructed a16z investment record. The included universe is therefore 68 deal-linked public claims across 46 companies: 43 Casado-authored or coauthored a16z investment announcements plus 25 current/former board claims, with 22 companies appearing in both sets and three board-only companies. Multiple financings disclosed inside one announcement remain subfacts rather than new denominator units. a16z author archive, announcement index · a16z profile, current and former board paragraphs

Attribution rule. [researcher inference · strong] A Casado byline proves public participation in the published thesis; a joint byline remains team work. A board roster proves board service, not sourcing, sponsorship, diligence, IC approval, check allocation, ownership, follow-on authority, exit approval, or proceeds. Company and SEC records can establish financing or transaction facts but not silently fill person-level roles. Where the record is silent, the ledger says unknown; it never upgrades a firm lead to a Casado lead. a16z portfolio disclosure, disclosure below portfolio table · current author/profile record

Economics rule. [researcher inference · strong] A financing amount is capital raised; a post-money valuation is a private-company mark; revenue and annualized revenue are operating measures; acquisition consideration is company-level transaction value; a listing is a liquidity venue. None is a16z or Casado proceeds. Publicly observed person- or fund-specific proceeds are unknown, not zero. The record cannot support a hit rate because public claims are selective, losses and passes are underdisclosed, follow-on depth varies, and the underlying actual-deal denominator is unavailable.

Historical and base-rate context

Base-rate boundary. [researcher inference · strong] These statistics contextualize selection, price, fundraising, exits, and liquidity. They are not matched by vintage, stage, geography, ownership, risk, or cash-flow timing and cannot establish alpha. The public Casado record is positively selected: investor announcements explain why a16z invested, while passes, write-offs, reserve refusals, and ordinary private outcomes are far less likely to be published.

Vehicle economics, incentives, and decision rights

Formal authority and compensation boundary

• Practice leadership. [contemporaneous record · strong] The firm profile says Casado leads a16z Infrastructure; in a direct interview he describes running an approximately $1.2B Infrastructure fund, consistent with a rounded account of the announced $1.25B pool. a16z profile, opening biography · “Monopolies vs. Oligopolies in AI,” 53:38–57:09
• Legal control record. [contemporaneous record · strong] Fund VII and Funds IX/X filings name Andreessen and Horowitz in the GP managing chain. The current ADV identifies A16Z Holdings as a 75%+ direct owner/control person and Andreessen and Horowitz as indirect owners/control persons; Casado was not located in those ownership/control schedules. A16Z ADV, Schedules A/B
• Inference boundary. [researcher inference · strong] Absence from those schedules does not prove Casado lacks an IC vote, delegated budget, informal influence, or economics. Conversely, “leads” or “runs” does not prove a legal vote, veto, unilateral allocation, carry, or ownership. Partnership agreements, IC charters and minutes, delegation schedules, and compensation records would be needed.
• Fee evidence. [contemporaneous record · strong] Fund IX/X Forms D state that a GP or affiliate is entitled to a management fee under governing documents. They do not supply a usable rate, base, duration, offsets, hurdle, or carried-interest split. Fund IX, Item 15 · Fund X, Item 15
• Valuation and RAUM caution. [contemporaneous record · strong] The ADV reports 0% of Fund IX/X assets valued by a non-related person and warns that multiplexer RAUM double-counts uncontributed commitments and component-fund assets. That supports a valuation-process caution, not an allegation of misvaluation. GAV and RAUM are not NAV, proceeds, or LP distributions. A16Z ADV, Schedule D §7.B.(1) and Miscellaneous explanation

Allocation, ownership, and conflict incentives

• Ownership incentive. [investor-stated · strong] Casado says early-stage ownership is binding because a median outcome must return a meaningful share of a fund; price is market-set, while large markets and growth-stage checks alter the constraint. This is a general framework, not evidence of a company-specific stake. “Monopolies vs. Oligopolies in AI,” 53:38–57:09
• Conflict convention. [investor-stated · strong] He describes a “one mortal enemy” convention under which a founder may identify one rival the investor will not back and acknowledges that pivots can create later conflicts. same interview, 57:09–59:35
• Stage-team separation. [investor-stated · moderate] Casado says seed, venture, and growth teams are separated partly so founders can speak candidly to a board sponsor while another group assesses a later round. 20VC, 14:15–15:33
• Conflict implication. [researcher inference · moderate] Mapping a market and reserving a competitor lane can focus diligence and support a portfolio company, but it can also reduce later learning and optionality. The record does not disclose a general recusal, information-barrier, valuation, or cross-vehicle allocation protocol.

Why returns remain non-computable

[researcher inference · strong] Gross MOIC requires realized proceeds, residual value, and invested cost. Net TVPI/DPI require LP distributions, ending NAV, and paid-in capital after fees, expenses, and carry. IRR additionally requires every dated capital call and distribution. Apart from Yubico's proposed 6.1% combined-company closing snapshot, the public record lacks complete ownership histories; it also lacks company-to-fund mapping, initial and follow-on cost, security and preference terms, reserves, tender or secondary sales, residual NAV, fee/carry allocation, recycling, subscription-line effects, and dated LP cash flows. Yubico/ACQ Bure release, “Ownership structure” table

Hold every observed fact constant, including Cursor's $60B merger-implied value. One compatible scenario gives a relevant fund an immaterial diluted stake bought at high cost and never distributed; another gives it a material low-cost stake whose SpaceX shares are promptly distributed. Both fit the public record, yet their MOIC, DPI, and IRR differ by orders of magnitude. Therefore the record does not identify a unique answer. Report not publicly computable, never zero and never a point estimate.

People, capital, and influence map

Comprehensive public person-linked deal ledger

Key. [researcher inference · strong] A is an official Casado-authored/coauthored a16z deal announcement; BC/BF is a current/former board claim in the live a16z profile, current/former board paragraphs; R/I is sufficiently reconstructed/incomplete. Roles are S source, P public sponsor/narrator, D diligence, Ap formal approval, B board, F follow-on, and X exit; ? means unknown. An R record reconstructs the public claim—not the security, economics, or return. Unless stated otherwise, the capital entity is an a16z-managed vehicle whose exact fund, check, ownership, reserves, fee, and carry are unknown.

Ledger arithmetic and observable patterns

• Stage partition. [contemporaneous record · strong] The 43 announcement records divide into 7 seed-only, 12 Series-A-only, 5 Series-B-only, 4 explicit multi-check/multi-round, and 15 stage-unknown pages: 7 + 12 + 5 + 4 + 15 = 43.
• Follow-ons. [contemporaneous record · strong] SigOpt, Tecton, Coactive, and Ideogram explicitly disclose multiple checks; Metronome's company post separately says seed+A. No complete reserve schedule exists.
• Governance. [contemporaneous record · strong] The profile supplies 25 board claims: 17 current and eight former. Eleven announcements contain board-seat language, but three coauthored posts leave the designee unassigned. Historical Rasa/Tackle seats are absent from today's roster, so present tenure is unknown.
• Outcomes. [researcher inference · strong] Fourteen distinct primary-record corporate events affect 15 of 46 lineages: three priced transactions/listing events, ten unpriced acquisitions/combinations (the Fivetran–dbt event affects two lineages), and one bounded Rapid asset/R&D-unit sale. The remaining 31 lineages have no qualifying primary corporate-outcome record in this sweep; that is a search status, not a solvency or return conclusion.

Ex-ante decision labs

Each lab freezes the public information set before revealing the later record. The proposed kill tests and modern decision rules are researcher inference unless a source expressly states them; they are not reconstructed IC memos. Shared record limit. [researcher inference · strong] None of the seven cases discloses a complete alternative set, formal dissent, IC vote, deal-specific vehicle allocation, or post-outcome board/IC process change. Every alternative below is therefore a researcher-generated decision branch, and silence means unknown. The 2018 cases sit inside the public Fund V platform context; the 2020 cases precede or closely bracket the Fund VII announcement; the 2021 Tackle decisions must not import later full-year boom statistics; and the 2024 Cursor freeze follows the Fund IX Infrastructure announcement but is not publicly mapped to that vehicle. Fund V, allocation paragraphs · Fund VII, opening/team paragraphs · Fund IX, allocation paragraph

Decision lab 1: Cursor — workflow pull, model dependence, and an outlier exit

Freeze: 2024-08-22. [contemporaneous record · strong] The four-person a16z Series A post described an AI-native VS Code fork, thousands of paying developers, adoption at named sophisticated companies, and users who rarely switched back. Cursor did not train a foundation model. Known: product, paid pull, technical team, familiar editor base. Unknown: ARR, retention cohorts, inference-adjusted margin, enterprise security, customer concentration, model-supplier power, terms, ownership, and individual a16z roles. Fund IX Infrastructure had been announced four months earlier, but no public source maps this deal to that vehicle. a16z Cursor, byline through founders/model-choice paragraphs · Fund IX announcement, allocation paragraph

Underwriting reconstruction. [researcher inference · moderate] The thesis was that high-frequency workflow ownership and context could matter more than owning the model. Kill if paid cohorts reverted, model cost made usage uneconomic, vendors or GitHub bundled the interface, security blocked enterprise use, or the fork became a maintenance trap. Next-check evidence: retained paid seats, expansion, cost per useful task, latency, security conversion, and supplier concentration. No public terms justify a specific ownership answer.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were to wait for retention/margin proof, back a model-owning competitor, build a thinner extension thesis, or pass. No public record identifies the alternatives actually considered, dissent, the board designee's intervention, a stop decision, or a changed a16z process after the outcome.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] Cursor later reported $900M C/$9.9B and $2.3B D/$29.3B with rapidly rising annualized revenue; those are issuer-reported financings/metrics. SpaceX's filed agreement used a $60B implied Cursor equity value in an all-stock merger, and Cursor said it joined SpaceX on 2026-08-14. Cursor C, opening financing/ARR paragraphs · Cursor D, opening financing/revenue paragraphs · SpaceX 8-K, Item 1.01

Interpretation. [researcher inference · moderate] Later model/infrastructure investment is consistent with responding to external-model dependence, but the causal motivation and any a16z intervention are not established. The company event is directionally exceptional as visible company-value evidence; a16z cost, stake, stock received, lockup, tax, distribution, MOIC, and IRR remain unknown.

Lesson/counterfactual. [researcher inference · moderate] Workflow pull and retention were the strongest early signals. A standalone Cursor might have compounded or faced platform commoditization; the merger prevents the public record from resolving that counterfactual. Preserve the four-person team attribution.

Decision lab 2: Parsec — consumer product to enterprise infrastructure

Freeze: 2020-12-15. [investor-stated · strong] Casado said he saw a demo in 2017, tracked Parsec for years, and observed a loyal gaming community plus enterprise customers in media, architecture, engineering, and game development. The $25M Series B was led by a16z during pandemic remote work, 25 days after the public Fund VII announcement; no source maps the check to that vehicle. Known: difficult low-latency technology, product community, named enterprise demand. Unknown: post-pandemic durability, margin/egress cost, security, concentration, sales cycle, terms, ownership, and competitive bundling. Parsec, 2017-demo/customer and Series B/board paragraphs · issuer Series B, opening financing paragraph · Fund VII announcement, opening

Underwriting reconstruction. [researcher inference · moderate] A demanding consumer use case could prove a horizontal enterprise primitive. Kill if latency advantage disappeared, enterprise cohorts did not expand after reopening, cloud cost blocked margin, security failed, or consumer/enterprise focus fractured. Next evidence: renewals, expansion, reliability, security, mix shift, and egress-adjusted contribution margin.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were to wait for post-pandemic enterprise cohorts, back a native enterprise remote-work product, license the technology, or pass. No public record identifies actual alternatives, dissent, acquisition-vote conduct, stop rules, or a process change after the sale.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] Unity acquired Parsec about nine months later. Its 10-K records $332.729M gross cash consideration, separately records $23.402M of acquired cash as an asset, and says the transaction was not material enough for separate pro-forma results. The company cash price is not fund proceeds; entry valuation, allocation, preference stack, dilution, waterfall, escrow, and distributions are absent. Unity 10-K, Note 6 Parsec table

Lesson/counterfactual. [researcher inference · moderate] Multi-year observation before investment is more replayable than the quick sale. Independence could have created more value or exposed pandemic normalization and bundling risk; the early acquisition cannot validate the whole standalone thesis.

Decision lab 3: Tackle — customer truth, preemptive capital, and overbuilding

Freeze 1: 2021-03-09; freeze 2: 2021-12. [investor-stated · strong] Casado found Tackle because another portfolio company used it, cited 200+ customers, references, growth and low burn, led the $35M B, and joined the board. The issuer later called its C preemptive and planned more aggressive scaling. Known: customer pull, marketplace volume, founder references. Unknown: NRR, software/services mix, take rate, gross margin, cloud concentration, policy risk, implementation labor, CAC payback, terms, ownership, board votes, and downside runway. The decision-date posts contain no matched market-financing baseline; later full-year 2021 boom statistics are context for review, not facts available at either decision. Tackle thesis, discovery/traction and B/board paragraphs · Series B, financing/metrics paragraphs · Series C, preemptive/scaling paragraphs

Underwriting reconstruction. [researcher inference · strong] Kill if marketplace volume failed to create durable software revenue, services were required to deploy, a hyperscaler internalized the layer, payback stretched, or costs assumed an unsupported curve. Before the C, re-underwrite cohort expansion, services content, bookings-to-cash, burn multiple, concentration, and a no-new-capital scenario.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were a smaller or delayed C, milestone-gated hiring, services partnerships, or no follow-on. No public record identifies actual dissent, Casado's board vote, reserve decision, stop rule, or an investor process change after the operating reset.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] In 2023 CEO John Jahnke said 75 employees would leave, the company had “overbuilt in 2021 and 2022,” and it was broadening toward services/sales-led work while retaining more than five years of runway. AppDirect announced an acquisition agreement in 2025, and Tackle's current company page identifies it as acquired; terms were not disclosed. Tackle update, overbuilt/runway/model paragraphs · AppDirect, opening transaction paragraphs · Tackle company page, acquired status

Interpretation. [researcher inference · strong] The public warning pattern is a preemptive round followed by headcount acceleration and admitted overbuilding; the company response was cost reduction, runway extension, and model change. The rule that every preemptive round needs a downside re-underwrite is researcher-generated. No documented investor process change, Casado board action, reserve vote, ownership, or acquisition waterfall was found.

Lesson/counterfactual. [researcher inference · strong] Tackle is a documented operating miss and unknown investment outcome—not a zero or proven loss. Slower hiring or a smaller C might have preserved focus and ownership, or ceded a forming category. Every preemptive round needs a downside re-underwrite.

Decision lab 4: Isovalent — open-source control point and an unpriced exit

Freeze: 2020-11-10. [investor-stated · strong] The thesis was that eBPF/Cilium could become a cloud-native networking, observability, and security layer, supported by open-source production adoption and a team Casado knew through Nicira/Stanford. Terms, stage, board role, conversion metrics, concentration, and governance were unknown. Isovalent, eBPF/Cilium adoption/team passages

Underwriting reconstruction. [researcher inference · moderate] Kill if cloud vendors absorbed the control point, community neutrality failed, kernel security risk emerged, or open source did not convert into paid expansion. Next evidence: production adoption, commercial attach, retention, and ecosystem neutrality.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were to wait for paid conversion, back a closed commercial layer, rely on a cloud-vendor implementation, or pass. The public record does not identify actual alternatives, dissent, board designation, acquisition-vote conduct, a stop rule, or an investor process change.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] Cisco completed the acquisition in 2024 without disclosing consideration. Cisco completion, opening paragraphs

Interpretation. [researcher inference · moderate] The strategic fit is consistent with the technical thesis but does not validate it causally or identify a16z's return, board response, acquisition alternatives, or the value of remaining independent.

Lesson/counterfactual. [researcher inference · moderate] Open-source adoption is incomplete evidence until commercial attach, neutrality, and platform bargaining power are tested. Remaining independent could have preserved upside or exposed cloud-vendor capture; undisclosed consideration cannot resolve the branch.

Decision lab 5: Yubico — product truth to public transaction

Freeze: 2018-05-31. [investor-stated · strong] a16z used YubiKeys internally; Casado described reference checks among security teams, broad adoption, profitability, limited outside capital, and his board seat. Known: working hardware, open standards, practitioner use, capital efficiency. Unknown: check, price, ownership, hardware concentration, enterprise sales efficiency, consumer adoption, standards/platform risk, and exit rights. The public context is Fund V's broad multi-stage platform; the deal-to-vehicle mapping is unknown. Yubico, root-of-trust, diligence, capital-efficiency and board paragraphs · Fund V, allocation paragraphs

Underwriting reconstruction. [researcher inference · moderate] Kill if platform authenticators displaced keys without coexistence, standards fragmented, supply chain constrained growth, or enterprise use failed to expand. Next evidence: deployments, replacement/renewal, standards adoption, margins, and platform dependence.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were software-only authentication, waiting for platform-standard adoption, a smaller non-board position, or passing. Actual alternatives, dissent, transaction-election conduct, stop rules, and post-listing process changes are unknown.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] Yubico combined with ACQ Bure and listed in 2023 at an approximately $800M company equity value through mixed cash/stock mechanics. Before closing, the issuer expected a16z to own 6.1% of the combined company and disclosed a conditional 180-day lockup. The completion record does not establish a16z's actual closing shares, later compliance or sales, entry basis, fund allocation, distributions, or return. Listing is a venue, not realization. Yubico/ACQ Bure release, “Ownership structure” table and lockup paragraph · first trading day, opening completion paragraph

Lesson/counterfactual. [researcher inference · moderate] Direct product truth and practitioner references are strong early evidence, but company transaction, holder liquidity, and LP distribution remain three separate gates.

Decision lab 6: Astranis — outside-competence learning and capital intensity

Freeze: 2018-03-01. [investor-stated · strong] Casado said a16z had not invested in space, described his own initial context gap, and backed a lower-cost geostationary architecture after technical/founder work; he joined the board. Known: prototype path, differentiated architecture, founders, first-space-deal status. Unknown: launch reliability, supplier risk, customer contracts, regulatory timing, debt capacity, constellation economics, check, ownership, and team decision allocation. The public platform context is Fund V; no deal-level vehicle is identified. Astranis, unfamiliarity/architecture and board paragraphs · Forbes, opening financing/board and prototype passages · Fund V, allocation paragraphs

Underwriting reconstruction. [researcher inference · moderate] Kill if launch/on-orbit milestones failed without credible correction, contracts lacked financing value, unit economics required uneconomic scale, or debt/equity needs outran milestones. Next evidence: tests, launch, customer deposits, regulatory approvals, supplier redundancy, on-orbit performance, and capital per deployed capacity.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were to wait for flight proof, back a different orbit/architecture, syndicate a smaller option position, or pass. Actual alternatives, dissent, board responses to the anomaly, stop rules, and process changes are not public.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] Astranis reported a mission-impacting Arcturus solar-array-drive anomaly, reproduced a supplier-component failure, and proposed a fix. It later raised substantial equity/debt and reported five satellites/backlog. These are operating and financing facts, not valuation or return. No public source allocates a16z's governance response. Arcturus update, anomaly/reproduced-failure paragraphs · Series E, opening capital and operating paragraphs

Lesson/counterfactual. [researcher inference · moderate] An outside-domain deal should translate curiosity into technical, regulatory, customer, financing, and supplier milestones. Continued financing can reflect progress or capital dependence; without unit economics and ownership, it does not answer the return question.

Decision lab 7: Tecton — repeat capital and open-source commercialization

Freeze: 2018 seed and 2020-04-28 Series A. [investor-stated · strong] The post connected the founders' Uber Michelangelo experience to a production feature store, said a16z engaged top ML teams, and disclosed seed then A investment. Known: domain team, production problem, market conversations, repeat decision. Unknown: seed/A checks and vehicles, ownership, open-source/commercial conversion, retention, cloud competition, internal approval, and acquisition rights. The seed falls within the public Fund V era; the A predates the Fund VII announcement, and neither check is mapped to a vehicle. Tecton, Michelangelo/feature-store and seed/A closing · Fund V · Fund VII announcement

Underwriting reconstruction. [researcher inference · moderate] Kill if feature stores became a cloud-native commodity, data gravity favored platforms, deployment remained services-heavy, or paid use did not expand. Next evidence: production deployments, renewal/expansion, attach to ML workloads, time-to-value, and platform neutrality.

Alternatives, dissent, and governance. [researcher inference · insufficient public record] Plausible branches were seed-only observation, a cloud-native competitor, a broader MLOps platform, or passing on the A. Actual alternatives, dissent, the repeat-check approval, board role, acquisition conduct, stop rules, and process changes are unknown.

1. Would you invest then?
2. What would kill the deal?
3. What ownership/terms fit the risk?
4. What evidence unlocks the next check?

Reveal. [contemporaneous record · strong] Tecton later raised a $100M C, and Databricks announced the strategic acquisition without disclosing consideration. Series C, round/syndicate section · Databricks, summary and “Tecton + Databricks”

Interpretation. [researcher inference · moderate] A follow-on investor becoming acquirer is consistent with strategic fit but does not validate the original selection causally or reveal price, seller allocation, fund proceeds, governance response, or alternative standalone value.

Lesson/counterfactual. [researcher inference · moderate] Repeat capital is most informative when tied to a changed evidence set. The public post says seed then A, but not the internal gate. Platform acquisition may be the best path or evidence that standalone independence narrowed; public data cannot choose.

Portfolio forensics

Category exposure

[researcher inference · moderate] This mutually exclusive taxonomy is a research convenience, not an a16z classification. It shows that the public record is overwhelmingly software/infrastructure and that later AI additions extend rather than replace the earlier control-point thesis.

Geography, check size, and invested-capital boundary

• Geography. [researcher inference · insufficient public record] The announcement and mutable board/profile corpus does not consistently state incorporation, headquarters at investment date, operating footprint, or customer geography. A reproducible geographic partition was therefore not constructed; domain examples and current websites cannot be substituted for decision-date domicile.
• Check size and invested capital. [researcher inference · insufficient public record] Public posts often disclose total round size and sometimes that a16z led, but they rarely disclose the a16z check, vehicle, security, reserves, or follow-on allocation. No compatible invested-capital total exists across the 46 lineages, so round totals are not summed as firm or Casado capital. vehicle and economics boundary · source-map denominator boundary

Sourcing, diligence, and follow-on patterns

• Portfolio and customer loops. [investor-stated · strong] SigOpt surfaced through a portfolio executive briefing; Tackle through a portfolio-company board meeting; Fivetran through portfolio and public-company encounters; Distributional through SigOpt and interviews with 25+ AI leaders; ActionIQ through a prior Stanford relationship and a dozen marketer interviews. SigOpt, opening sourcing and closing follow-on · Tackle, discovery paragraph · Distributional, diligence/relationship paragraphs
• Product truth. [documented behavior · moderate] Braintrust and Yubico record direct a16z/Casado product use; Convex records earlier personal angel/adviser involvement. This improves access to product evidence but can also create confirmation, information-asymmetry, and conflict risks. Braintrust, lead/user sentence · Yubico, internal-use/diligence paragraphs · Convex company Series A, angel/adviser paragraphs
• Repeat-capital evidence. [contemporaneous record · strong] Only four Casado announcement pages explicitly disclose multiple checks—SigOpt, Tecton, Coactive, and Ideogram—while Metronome's issuer post separately says seed+A. Public reserve decisions, decline cases, and ownership maintenance are missing.
• Board coverage. [contemporaneous record · strong] Twenty-five of 46 lineages appear on the current profile's board roster, but only 11 deal announcements contain board-seat language and three of those are collective/unassigned. Board service increases access; it does not prove the initial investment role or later intervention.
• Attribution concentration. [researcher inference · strong] Early posts are more often Casado-only bylines; later AI/application posts more often name two to four a16z authors. Cursor's outcome should not be retroactively converted into solo Casado evidence, and coauthored recent deals are not independent proof that a personal edge scaled.

Outcome and realization forensics

Outcome reconciliation. [researcher inference · strong] The table contains 14 distinct corporate events affecting 15 lineages: 3 priced events + 10 unpriced company events + 1 asset event. Therefore 15 affected + 31 without a qualifying primary record = 46 included lineages. WaveForms and Deeptune remain secondary/provisional reports and do not enter the primary-event count.

Tender/secondary boundary. [researcher inference · insufficient public record] No reviewed primary source identifies an a16z tender or secondary sale in these 46 companies. The conclusion is “not observed in the bounded record,” not “no sale occurred.”

Failures, misses, and adverse evidence

No-result discipline. [researcher inference · strong] Searches also covered Fivetran/dbt, DeepMap, Metronome, Preset, WaveForms, Resend, Phota, Relace, Shizuku, QuiverAI, Deeptune, Runta, Fly.io, EngFlow, ActionIQ, Eclypsium, and the remaining ledger. “No adverse event surfaced” is unknown, not clean, because private-company disclosures are selective and identity/role coverage is uneven.

Misses and anti-portfolio

• Hugging Face recollection. [retrospective witness · moderate] Casado says the firm sometimes talked itself out of companies despite strong usage and uses Hugging Face in a discussion of open-source traction. The record supplies no dated opportunity, allocation, term sheet, IC vote, or proof the firm could have invested. Treat it as a process warning against over-intellectualizing—not a priced miss or counterfactual return. Colossus, 00:29:03–00:31:48
• “Nearly 200” is not a denominator. [investor-stated · strong] Casado's direct interview mixes investments he made with funds/teams he helped run; his Series A preference, $30–40M total-exposure framework, standalone-outcome preference, and informed-objection test are general process claims. a16z News, 00:00:54–00:11:04, 00:28:20–00:32:46, 00:52:10–00:53:06
• Missing anti-portfolio. [researcher inference · insufficient public record] No complete passes, rejected-deal chronology, allocation losses, write-offs, reserve refusals, or public/private-market opportunity-cost ledger was found. Therefore public announcements cannot yield a personal selection hit rate.

Uncertainty, missing records, and continuation routes

Bottom line

[researcher inference · strong] Casado's public record is most consistent with a technical-control-point and customer-truth investing practice amplified by a16z's portfolio, team, capital, and board access. Cursor is the dominant visible company-value event; Parsec is the cleanest cash-priced ordinary exit; Tackle is the strongest adverse operating lesson. Yet no public case identifies a person- or fund-specific return, and recent joint bylines weaken any solo-repeatability claim. The defensible conclusion is about observable decision evidence and company outcomes—not personal alpha, realized returns, or a complete hit rate.

Coverage and evidence base

Literal priority-source reconciliation

The WRITINGS inventory contains 16 priority works: 9 ★★★ and 7 ★★. Every work is processed here and has one source note. Fourteen canonical or authorized copies exposed the complete article body; two publisher routes exposed only an authenticated headline, deck or opening plus the a16z author-index attribution. The denominator therefore reconciles as 14 full + 2 partial + 0 metadata-only = 16, and analysis of the two gated essays stops at the visible boundary.

Evidence boundary — researcher inference | strong. The corpus is unusually useful for reconstructing a dated infrastructure-investing vocabulary: marginal data value, AI contribution margin, hybrid go-to-market, workload specialization, market creation, stack bargaining power, compute procurement, open-weight policy, and physical-AI industrial capacity. It is not a performance record. Coauthored essays are joint positions; unnamed portfolio observations lack a denominator; examples can be success-selected; and no writing supplies complete passes, checks, ownership, reserves, fund cash flows, or realized returns. a16z itself warns that mentioned investments are not representative and that some third-party information is not independently verified. a16z article disclosure · source map, candidate-deal boundary

Source-by-source analysis

1. “The Empty Promise of Data Moats” (2019, with Peter Lauten)

Context, argument, and reasoning. Casado and Lauten distinguish a true network effect—each additional node directly makes participation more valuable—from a data scale effect, where a larger corpus may improve a model without connecting users. Their anti-moat mechanism is two-sided: the next useful observation often becomes harder and costlier to acquire, clean, label, secure, and refresh, while overlapping or long-tail observations add less product value. A permission-shared Eloquent Labs customer-support distribution illustrates diminishing coverage, but it is one domain rather than a representative enterprise-AI base rate. investor-stated | strong for the framework; investor-stated | moderate for its empirical generalization. Essay, “Data + network effects ≠ data network effects” through “A practical framework”

Decision implications, boundary conditions, and evolution. The usable underwriting rule is to measure minimum viable corpus, marginal acquisition cost, incremental model lift, decay, rights, portability, error tolerance, and competitor catch-up—then test ordinary moats such as workflow, distribution, vertical knowledge, trust, brand, and account control. The authors preserve important exceptions: scarce or exclusive sources, regulated access, trust-mediated custody, material accuracy thresholds, and hard-to-reproduce domain learning can be defensible. The later robotics essay’s deployment-data flywheel and the 2025 margins essay’s product-feedback thesis therefore do not simply reverse 2019; they must prove the exact scarce-data and incremental-value conditions this article specifies. researcher inference | strong for the continuity test; no public deal record shows that Casado used it to approve or reject a named investment. Essay, “When IS data defensible?” · robotics essay, deployment-data discussion · 2025 margins essay, product-specific data discussion · source note

2. “The New Business of AI (and How It’s Different From Traditional Software)” (2020, with Matt Bornstein)

Context, argument, and reasoning. Before the generative-AI boom, Casado and Bornstein argue that many AI companies combine repeatable software with service-like production: inference and retraining consume infrastructure, changing customer distributions create deployment work, humans review or catch failures, and edge cases can require customer-specific handling. Their private-company observations put many gross margins around 50–60%, cloud at 25% or more of revenue, and ongoing manual data work at 10–15%, while warning that five to ten early customers can conceal a growing ML backlog. The source states that these estimates come from founder interviews and publishes no sampling frame, so they are dated hypotheses rather than base rates. investor-stated | strong for the cost taxonomy; investor-stated | provisional for the numeric generalization. Essay, opening list, both “Gross Margins” sections, “Scaling AI systems,” and “Sources”

Decision implications, boundaries, and later tension. The direct operating advice is to narrow the domain, minimize per-customer model forks, allocate compute and human failover to cost of revenue, expose edge-case backlog, preserve stack optionality, and accept a coherent services model when the task truly requires it. For investment analysis, recurring revenue is insufficient: contribution margin must include inference, evaluation, implementation, support, and successful-task completion. The 2023 economic essay argues that foundation models and more error-tolerant creative/copilot work can escape this “mediocrity spiral,” while the 2025 margins essay adds routing, tiering, and enterprise mix as repair mechanisms. Those are testable updates, not permission to ignore the 2020 accounting discipline. researcher inference | strong for this reconciliation; the corpus provides no matched cohort showing which companies crossed from services-heavy delivery to scalable product. 2020 essay, “Building, scaling, and defending great AI companies” · 2023 economic case · 2025 margins essay · source note

3. “Growth+Sales: The New Era of Enterprise Go-to-Market” (2020, with Peter Lauten)

Context, argument, and reasoning. Lauten and Casado describe “growth+sales” as two linked funnels: a self-serve product first creates end-user adoption, retention, word of mouth, and within-logo spread; enterprise sales then converts that pull into security review, procurement, broad rollout, premium features, and expansion. Their six failure modes make the thesis falsifiable: a product may not reach value unaided; paid acquisition may buy transient users; company-level revenue can hide weak engagement; sales can arrive before customer pull; free use may lack a buyer or premium bridge; and a successful sales motion can cause product neglect. The examples are selected winners, so they establish plausibility rather than comparative superiority. investor-stated | strong for the operating sequence; investor-stated | moderate for the market generalization. Essay, “New buyer, new motion” and Failure Modes #1–#6

Decision implications, boundaries, and evolution. The diligence rule is to reconstruct the full user-to-budget path: independent time-to-value, cohort engagement, persona, organic retention, within-account propagation, buyer identity, security/procurement bridge, premium conversion, CAC by motion, and product velocity after sales hiring. Customer requests for wider rollout—not management’s desire for faster revenue—are the timing signal. The model does not fit centralized systems of record, regulated deployments, or integration-heavy products whose users cannot bypass authorization. It also complements rather than replaces 2023 market annealing: bottom-up pull can reveal a market, while founder-led sales still has to shape product, buyer, and category when no stable market exists. researcher inference | moderate for this cross-source reconciliation; no public decision record shows which Casado deal passed because of the dual-funnel test. Essay, Failure Modes #1, #3–#6 · Market Annealing · source note

4. “Emerging Architectures for Modern Data Infrastructure” (2020, with Matt Bornstein and Jennifer Li)

Context, argument, and reasoning. The original October 2020 version—not the later update now foregrounded by the shorter landing route—maps a data-infrastructure layer whose spending, tooling, and operating complexity were rising. Bornstein, Casado, and Li report a two-year research process involving hundreds of founder and expert conversations plus more than 20 practitioner interviews. Their core mechanism is architectural: warehouses and lakes were acquiring overlapping storage, transaction, semi-structured-data, scale, and SQL capabilities, while analytics and operational/ML workloads still differed enough to support parallel systems. Three blueprints then match control and complexity to customer maturity: modern BI, multimodal processing, and a full AI/ML stack. investor-stated | strong for the blueprint trade-offs; investor-stated | provisional for the market totals and category breadth, which rely on selected third-party data and an undisclosed practitioner sample. Original essay, opening, “Analytics, AI/ML, and the Great Convergence?” and “Blueprints”

Decision implications, boundaries, and evolution. The investable rule is to locate an unstable but necessary control point: identify the customer blueprint, indispensable workload, expertise burden, integration cost, and whether convergence strengthens or erases the vendor. This is a sourcing taxonomy, not a portfolio or return model. The essay simultaneously celebrates tool proliferation and predicts convergence; an investor must therefore distinguish temporary exploration from a durable independent market. By 2026, the right use is to rebuild the architecture from observed deployments and ask whether a product reduces total system complexity after people, integration, and egress—not to carry forward 2020 vendor boxes or market estimates. researcher inference | moderate. Original essay, unified architecture, blueprint limitations, and “Looking ahead” · version-bound source note

5. “Bottom Up Pricing & Packaging” (2021, with Jennifer Li)

Context, argument, and reasoning. Li and Casado derive a pricing system from the user journey rather than a copied SaaS page. A casual user can become a professional, team, and enterprise buyer; each transition changes the value delivered, buying authority, packaging, and service burden. Their four-tier pattern—free, prosumer, team, enterprise—comes from a 50-company map, while nine operating rules cover simple starting plans, journey continuity, monetization timing, qualification, price metrics, unit price, and experimentation. The mechanism is strongest where product usage can precede central authorization: organic value reveals demand, qualified accounts emerge, and sales completes security, procurement, administration, or broader rollout without severing the adoption path. investor-stated | strong for the decision sequence; investor-stated | moderate for generalization from selected companies. Essay, sections #1–#9 and four-tier framework

Decision implications, boundaries, and evolution. The IC test is a literal map of user, team, buyer, budget, trigger, tier, sales handoff, conversion, expansion, churn, support cost, and margin. Usage pricing should scale with value rather than punish expansion; free use should remain only while it creates measurable learning or qualified demand. The framework does not fit every regulated system of record, mandated deployment, or integration-heavy product, and waiting to monetize can conceal weak willingness to pay. It joins the earlier growth+sales model and later AI-margin argument: product pull is a distribution asset, but compute-intensive free use must earn its way into positive contribution margin. researcher inference | strong for the combined test; no named board decision or deal outcome validates it. Essay, sections #4–#9 · 2025 margins essay, conclusion · source note

6. “The Cost of Cloud, a Trillion Dollar Paradox” (2021, with Sarah Wang)

Context, argument, and reasoning. Wang and Casado argue that public cloud has high option value early—fast provisioning, flexible capacity, lower initial operating burden—but can become a structurally expensive cost of revenue at scale. Their mechanism is not “cloud bad”: once workloads stabilize and utilization becomes predictable, a company may capture part of the provider margin through owned or contracted infrastructure, while migration difficulty and engineering distraction create a lock-in tax. The essay’s market-cap and aggregate-value estimates are scenario arithmetic, not observed recoverable value, and its Dropbox example bundles several operational changes rather than proving repatriation alone caused the reported margin improvement. investor-stated | strong for the lifecycle framework; investor-stated | provisional for the $100 billion / $500 billion aggregate extrapolations. Essay, “The paradox,” “The cloud cost equation,” Dropbox case, and market-value estimate

Decision implications, boundaries, and evolution. Underwrite portability before scale: workload unit cost, utilization volatility, data gravity, egress, reserved pricing, migration capital, required reliability staff, depreciation, facilities, and fallback capacity. The choice is workload-specific and staged, not a company-wide ideological switch. Dropbox’s FY2025 filing is useful counterevidence to any frozen “owned means cheap” rule: it reports gross margin declining from 83% to 80% amid infrastructure-refresh and facilities costs. By 2026, compare realistic hybrid scenarios and preserve bargaining power through architecture and contracts. researcher inference | strong. Essay, repatriation decision and objections · Dropbox FY2025 10-K, MD&A, pp. 57–59 · source note

7. “The Cloud Killed Infrastructure. Long Live Infrastructure.” (2022)

Context, argument, and reasoning. Casado argues that hyperscalers did not eliminate infrastructure companies; they abstracted general-purpose primitives and thereby made specialized workload layers easier to build. The opportunity exists where a slice is large enough and sufficiently different in performance, data, security, or operating requirements that a verticalized cloud or service can outperform a generic platform. His examples establish that specialization is possible, not that every named vertical is a durable venture market. Meanwhile, hyperscalers continue to extend upward, bundle adjacent capabilities, subsidize strategic products, and improve price/performance. investor-stated | strong for the enabling logic; investor-stated | provisional for broad investability across the selected categories. Essay, cloud abstraction, specialized infrastructure, and examples

Decision implications, boundaries, and evolution. A vertical-cloud thesis must identify the exact workload discontinuity, absolute market, utilization advantage, customer switching cost, distribution path, capital requirement, and credible hyperscaler response. Loose coupling to the lower stack and a large specialized slice are enabling conditions; dependence on one cloud, weak differentiation, or a bundlable feature can destroy the thesis. This complements rather than reverses the cost-of-cloud essay: public infrastructure can enable a new vendor while remaining its largest cost and bargaining-power risk. researcher inference | strong. Essay, conditions and closing framework · source note

8. “Market Annealing: Getting to $10M ARR in Very Early Markets” (2023, with Peter Levine)

Context, argument, and reasoning. Casado and Levine describe company and market as co-adapting before product, buyer, price, positioning, and sales motion have stabilized. Founder-led sales carries product and category information across early customers; two or three “renaissance reps” can generalize that learning; specialized functions arrive only after repeatable patterns emerge. The authors give rough ranges for founder involvement, sales-leader timing, rep economics, and elapsed time, but the observations come from unnamed company experience and lack a failed-company denominator. investor-stated | strong for the learning sequence; investor-stated | provisional for numeric thresholds such as founder involvement through roughly $4–10 million ARR, hiring a leader around $2–4 million, or quota/load heuristics. Essay, byline, market-annealing model, founder-led selling, renaissance reps, and scaling sequence

Decision implications, boundaries, and evolution. The operating question is whether early commercial work is reducing uncertainty: common painful use case, repeatable user, economic buyer, deployment path, price, and shorter cycle. Bespoke revenue that grows without convergence is not annealing; it is services drift. A modern milestone plan should time-box experiments and define kill conditions for stagnant retention, falling reuse, rising customization, no common buyer, and no referenceable wins. The essay itself supplies the learning mechanism but not those stop rules. Bottom-up adoption can accelerate annealing when users self-discover value, but it cannot replace founder interpretation in an undefined market. researcher inference | strong. Essay, early customers through transition to scale · source note

9. “Who Owns the Generative AI Platform?” (2023, with Matt Bornstein and Guido Appenzeller)

Context, argument, and reasoning. In January 2023, Bornstein, Appenzeller, and Casado divide generative AI into infrastructure, foundation-model, and application layers. They observe infrastructure receiving much of the early dollar flow, model providers bearing training and serving costs, and applications showing rapid experimentation but uncertain retention, margins, and differentiation. The point is diagnostic rather than a settled winner forecast: no layer yet had an obvious systemic moat, and bargaining power could move as models commoditized, cloud providers integrated, or applications accumulated workflow and distribution. Their flow-of-funds and margin ranges are rough early-market estimates with incomplete samples. investor-stated | strong for the three-layer framework; investor-stated | provisional for 20–40%, 10–20%, 80–90%, and related quantitative generalizations. Essay, market map, value-flow, moat, and “What next?” sections

Decision implications, boundaries, and evolution. Diligence each layer’s gross profit after compute, supplier concentration, switching cost, portability, distribution, data rights, and ability to move up or down the stack. The 2023 “no systemic moats” conclusion was an information-set boundary, not a timeless fact. The FTC’s 2025 cloud/AI partnership study documents equity, revenue-sharing, control, exclusivity, information access, cloud commitments, and switching-cost mechanisms that can concentrate power; it expressly does not decide their legality or market effect. By 2026, a stack thesis must model contractual control as well as technical architecture. researcher inference | strong. Essay, defensibility and integration discussion · FTC Office of Technology report, executive summary and §§4–5, pp. 1–3 and 16–37 · source note

10. “Navigating the High Cost of AI Compute” (2023, with Guido Appenzeller and Matt Bornstein)

Context, argument, and reasoning. Appenzeller, Bornstein, and Casado convert AI-compute anxiety into a workload model. Cost depends on training versus inference, model size, request volume, latency, utilization, hardware availability, engineering capacity, provider discounts, and how frequently the workload changes. Public cloud or hosted APIs maximize early option value; reserved capacity, specialized providers, model optimization, and owned infrastructure can become economical as utilization stabilizes. Their prices and the rough annual-spend threshold discussed in the article describe April 2023 market conditions, not enduring procurement laws. investor-stated | strong for the variable set; investor-stated | provisional for price points and the roughly $50 million scale heuristic. Essay, byline, workload matrix, build-versus-buy, provider, and cost sections

Decision implications, boundaries, and evolution. Require a per-successful-task cost tree, not “GPU spend”: model choice, tokens or runtime, batching, utilization, latency SLA, retry/evaluation, human review, failure rate, peak capacity, depreciation, networking, and portability. Stress at least hosted, reserved/multicloud, specialized-cloud, and owned scenarios with transition capital and service reliability. The rule evolves the cost-of-cloud thesis into AI, but model architectures, hardware, prices, and supply have changed too much to reuse the 2023 dollar thresholds. researcher inference | strong. Essay, optimization and decision sections · source note

11. “The Economic Case for Generative AI and Foundation Models” (2023, with Sarah Wang)

Context, argument, and reasoning. Casado and Wang revisit the 2020 warning that capability is not economics. Their discontinuity claim has three parts: foundation models broaden task coverage; many creative and copilot tasks tolerate probabilistic output and invite user feedback; and falling generation cost can expand consumption and new product behavior. This can escape the “mediocrity spiral” in which edge cases require ever more services work. The essay openly labels its calculations rough and non-scientific; illustrative output-cost comparisons do not establish equal quality, complete workflow substitution, or durable margins. investor-stated | strong for the conditional mechanism; investor-stated | provisional for adoption, revenue, cost, and labor-substitution examples. Essay, “Capability != economics,” “How generative AI…,” “Let’s run the numbers,” and defensibility

Decision implications, boundaries, and evolution. Underwrite reliable completion rather than nominal generation: task tolerance, evaluation, human review, integration, liability, full workflow cost, retention, willingness to pay, and who captures falling-cost surplus. “Marginal creation cost toward zero” is directional rhetoric, not accounting; training, serving, distribution, supervision, and capital remain nonzero. Snowflake’s FY2025 filing is a useful current boundary: product gross margin fell from 74% to 71% as cloud infrastructure and AI-related investment rose and newer features lacked scale. Falling unit prices do not guarantee company margin expansion. researcher inference | strong. Essay, numbers and caveats · Snowflake FY2025 10-K, MD&A, pp. 70–71 · source note

12. “AI Talks Leave ‘Little Tech’ Out” (2024, with Katherine Boyle)

Access-bounded analysis. The Wall Street Journal route exposes an authenticated headline-level record but not a complete article body in the retrieved record; the a16z author archive preserves Casado’s and Katherine Boyle’s joint attribution. The title supplies a bounded joint position that AI-policy conversations omit “Little Tech,” but the accessible record does not define the constituency, identify the conversations, or expose the article’s reasoning, evidence, prescriptions, or examples. investor-stated | insufficient public record. No sentence below treats a headline, index record, or later commentary as the full text. WSJ canonical route, headline-level access · a16z author archive, attributed entry

Decision implication and boundary. A present-day diligence translation is to model compliance fixed cost, licensing or compute thresholds, incumbent exceptions, access to standard-setting, and whether rules change entrant survival without improving measured safety. That is researcher inference | moderate, not a recovered Boyle/Casado checklist. The article’s gated body, examples, citations, and precise definition of “Little Tech” remain unresolved, so it cannot bear a stronger regulatory or investment conclusion. source note

13. “Keep the Code Behind AI Open” (2024, with Ion Stoica)

Access-bounded analysis. The Economist route exposes a headline-level record but not a complete article body; The Economist's public post identifies Martin Casado and Ion Stoica as the guest-essay writers and attributes a bounded claim that open source will power AI innovation and remain the most secure software-development path. The accessible record does not define “open,” distinguish weights from code or data, expose evidence, or show exceptions and safety boundaries. investor-stated | moderate for the excerpted joint position; investor-stated | insufficient public record for the hidden reasoning chain. The Economist publisher post, body sentence above the linked card · Economist canonical route, headline-level access

Decision implication and boundary. The NTIA’s July 2024 open-weights report offers the strongest contemporary boundary: open models can diffuse innovation, research, and access while also carrying misuse and distribution risks; the evidence did not support immediate restrictions, but monitoring, evaluation, transparency, and the option for future action remained necessary. A 2026 investment test should specify which artifact is open, who can modify or deploy it, what complements retain value, what misuse and provenance controls exist, and whether openness creates adoption without eliminating capture. researcher inference | moderate. NTIA report, executive summary and recommendations · source note

14. “Base AI Policy on Evidence, Not Existential Angst” (2024)

Context, argument, and reasoning. Casado argues that AI-specific regulation should target demonstrated marginal risk—incremental harm relative to existing computer systems—rather than fear attached to a technology category. The useful mechanism is a baseline test: define the existing risk, isolate new capability and exposure, identify the gap in current law, then choose an intervention proportional to the added harm. Yet the essay’s broad claims that feared events did not materialize and AI appears very safe do not publish a search method, denominator, or common outcome measure. investor-stated | strong for the marginal-risk framework; investor-stated | provisional for the safety generalization. Essay, “Avoiding spurious AI regulations” and “AI policy based on reality”

Decision implications, counterevidence, and self-interest. Investors should model both over-regulation and under-regulation: use case, jurisdiction, affected population, severity, probability, exposure growth, irreversibility, existing authority, and a measurable revision trigger. Newsom’s SB 1047 veto criticizes a compute proxy while still calling for proactive guardrails; the international safety report documents current harms, uncertain severe risks, and limited assurance. Evidence-based policy is therefore an adaptive measurement regime, not “wait for catastrophe.” Casado’s role leading an AI/infrastructure practice creates an institutional interest in lower compliance burdens and should raise the burden for independent evidence. researcher inference | strong. Newsom veto message, pp. 1–2 · International AI Safety Report, executive summary and §5 · source note

15. “Questioning Margins Is a Boring Cliché” (2025, with Sarah Wang)

Context, argument, and reasoning. Wang and Casado reject a static gross-margin veto for AI applications. A low-margin launch cohort can improve through tiering, rate limits, routing to cheaper models, inference-price declines, enterprise mix, and expansion into higher-value workflow. The conclusion still says persistently low or negative margins can be fatal and should affect valuation; the real disagreement is whether a current margin is destiny. Historical winners show possibility, not a base rate, and the authors’ private observations omit company identities, cohort tables, periods, and audited margin bridges. investor-stated | strong for the repair mechanisms; investor-stated | provisional for 10–100x cost decline, 80/20 revenue concentration, and broad application-economics claims. Essay, all five claim headings and conclusion

Decision implications, boundaries, and evolution. Underwrite a dated margin bridge: inference per successful task, support, rate-limit sensitivity, routing quality loss, conversion by plan, enterprise expansion, churn after repricing, supplier concentration, and time to positive contribution margin. This is an evolution of—not permission to discard—the 2020 AI-cost framework. Snowflake’s reported margin compression provides a counterexample to automatic improvement, while model competition and packaging remain genuine repair options. The authors are financially interested a16z general partners making the investability case and publish no failed-startup denominator. researcher inference | strong. 2025 essay, conclusion · 2020 AI essay, gross-margin taxonomy · Snowflake FY2025 10-K, MD&A, pp. 70–71 · source note

16. “America Cannot Lose the Robotics Race” (2025, with Anne Neuberger)

Context, argument, and reasoning. Casado and Neuberger describe a physical-AI flywheel: manufacturing clusters colocate component suppliers, integrators, and customers; proximity improves iteration and tacit process knowledge; deployment produces real-world data; better policies increase useful deployment. Harmonic reducers illustrate why precision parts and integration—not foundation models alone—can be system bottlenecks. The essay’s robot-density, component-share, price, automation, and productivity figures have different dates and denominators and were not reconciled into a causal estimate. investor-stated | strong for the stack mechanism; investor-stated | provisional for the comparative industrial and deregulatory claims. Op-ed, “The impossible,” “Micron tolerance,” “How the West lost…,” and “What we need to do”

Decision implications, contradictions, and self-interest. Diligence the complete learning system: bill of materials, country/supplier concentration, yield, failure rate, integration cycle, customer-site permission, utilization, teleoperation, repair, data rights, and regulatory milestones. The data flywheel does not contradict the 2019 anti-moat essay unless deployment data are scarce, reusable, representative, and incrementally valuable; simulation, shared models, customer ownership, or hardware reliability may break it. The call for both low-friction Chinese supply and resilient allied capacity also needs an explicit dependency boundary. The article names World Labs while Casado’s a16z biography lists him on its board, making the company example financially interested. researcher inference | strong. Op-ed, component and policy sections · a16z profile, current board list · source note

Cross-source synthesis

Core written philosophy

Casado’s written record is most coherent when treated as a sequence of mechanism maps, not slogans:

1. Identify the changing boundary. Data infrastructure, cloud specialization, model/application layers, and robotics clusters become interesting where a technical or institutional abstraction moves. researcher inference | strong. Data architecture · vertical cloud · GenAI platform
2. Translate the boundary into an economic mechanism. Data volume must create marginal learning; model capability must create reliable completion; cloud flexibility must beat lifecycle cost; product use must lead to a buyer and budget. researcher inference | strong. Data moats · AI economics · growth+sales
3. Preserve option value while uncertainty is high. Start with simple packaging and flexible infrastructure; learn through founder/customer contact; internalize, specialize, or add sales only after demand and workloads stabilize. researcher inference | strong. Bottom-up pricing · cost of cloud · market annealing
4. Re-test instead of canonizing the first model. The 2020 low-margin warning becomes a 2023 conditional discontinuity and a 2025 margin-bridge argument; the invariant is full-cost measurement, not a permanent bearish or bullish position. researcher inference | strong. 2020 AI · 2023 economic case · 2025 margins

This supports a plausible architecture-to-economics translation skill. It does not establish superior security selection or returns. The public writings omit complete opportunities, passes, failed theses, checks, ownership, reserves, cash flows, and realized proceeds. researcher inference | insufficient public record. Candidate-deal boundary

Evolution and contradiction matrix

What the writings do and do not establish by decision dimension

Reusable decision playbook

The following translations are dated 2026-08-19 and are researcher inference, unless a source explicitly states the underlying historical rule.

1. Measure marginal learning, not accumulated data

• Historical rule. [researcher inference | moderate] Ask whether the next observation is costly, scarce, legal to use, and materially improves the product. Data-moat framework
• Diagnostics. [researcher inference | moderate] Learning curve by corpus size; incremental acquisition/label cost; decay; rights; error-segment lift; competitor catch-up; simulator or synthetic-data substitution. Data-moat note
• Kill test. [researcher inference | moderate] If incremental data produce negligible customer or model value, “more data” is not a moat. Data-moat note
• Modern boundary. [researcher inference | strong] Robotics deployment data qualify only when they satisfy the same scarcity, reuse, representativeness, and lift tests. Robotics flywheel
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: gate a claimed data moat on measured marginal product lift and lawful reusable access; does not transfer: raw corpus size, the 2019 category examples, or an assumption that proprietary data are automatically scarce; changed assumptions: synthetic data, simulation, foundation-model priors, and multimodal collection can lower replication cost; observable signals: lift by added-data tranche, acquisition/label cost, rights, decay, and competitor catch-up; likely misuse: calling any private dataset a moat; current example: the 2025 robotics flywheel is a source-stated application of the test, not an independently demonstrated outcome. Data-moat note · robotics note

2. Underwrite contribution margin per successful outcome

• Historical rule. [researcher inference | moderate] Allocate inference, training, human review, implementation, support, and edge-case work to the delivered product. 2020 AI cost taxonomy
• Diagnostics. [researcher inference | moderate] Cost per attempted and successful task; failure/retry/evaluation; customer-specific work; gross profit after support; task tolerance and liability. 2020 AI note
• Kill test. [researcher inference | moderate] If scale increases exception labor or model forks faster than gross profit, growth is services drift. 2020 AI note
• Modern boundary. [researcher inference | strong] Permit low initial margins only with measured routing, tiering, conversion, retention, and time-bound improvement. 2025 margin bridge
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: underwrite quality-adjusted contribution margin per completed task; does not transfer: 2020 model prices, workload mix, or a fixed software-margin benchmark; changed assumptions: model routing, agent retries, multimodal workloads, and vendor price competition alter both cost and completion; observable signals: cost per attempt/success, human-review share, support load, routing quality, conversion, retention, and a dated breakeven path; likely misuse: treating future model deflation as permission for indefinite negative unit economics; current example: Snowflake's FY2025 gross-margin compression is a grounded company counterexample to automatic margin improvement, not a matched AI-startup cohort. 2025 margin note · Snowflake FY2025 10-K, MD&A, pp. 70–71

3. Reconstruct the user-to-budget system

• Historical rule. [researcher inference | moderate] Product use must connect continuously to a buyer, budget, enterprise trigger, and pricing metric. Growth+sales · Bottom-up pricing
• Diagnostics. [researcher inference | moderate] Time-to-value, cohort retention, within-logo spread, buyer identity, premium trigger, security/procurement path, CAC and payback by motion, churn after price changes. GTM note
• Kill test. [researcher inference | moderate] Free usage without a credible buyer/premium bridge, or sales growth disconnected from retained users, fails. Pricing note
• Modern boundary. [researcher inference | moderate] Usage billing must track customer value and include AI cost volatility; avoid bill shock masquerading as expansion. Pricing note
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: trace user value to the economic buyer, premium trigger, procurement path, and value-aligned metric; does not transfer: historical price points, named-winner examples, or a presumption that bottom-up use converts; changed assumptions: variable AI costs, security review, outcome pricing, and agent autonomy can separate usage from value; observable signals: retained activation, within-logo spread, free-to-paid conversion, buyer-approved expansion, payback, and churn after repricing; likely misuse: presenting free-user growth or token consumption as enterprise demand; hypothetical: an AI assistant with high daily use but no budget owner fails until a paid workflow and quality-adjusted margin are demonstrated. GTM note · pricing note

4. Separate discovery from scale in early markets

• Historical rule. [researcher inference | moderate] Founders carry product, buyer, pricing, and category information until patterns repeat; renaissance reps generalize before specialized teams scale. Market Annealing
• Diagnostics. [researcher inference | moderate] Common use case, reusable product, repeated buyer, cycle time, referenceability, implementation variance, salesperson ramp, quota contribution after full load. Annealing note
• Kill test. [researcher inference | moderate] Revenue rises while reuse falls, customization grows, and no common buyer or deployment emerges. Annealing note
• Modern boundary. [researcher inference | moderate] Treat the essay’s ARR and timing ranges as historical observations; set company-specific experiments and stop dates. Annealing note
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: separate founder-led discovery from repeatable sales and require convergence in product, buyer, and deployment; does not transfer: the essay's $10M/$30M/$100M markers, timing ranges, or selected-company outcomes; changed assumptions: AI-assisted building and selling can accelerate iteration without creating a budget or reusable product; observable signals: repeated use case/buyer, declining customization, stable cycle time, references, rep ramp, and fully loaded quota contribution; likely misuse: excusing permanent bespoke services as category creation; hypothetical: an industrial-AI vendor may keep founders in discovery until three repeat deployments share a product and buyer, then stop or scale on a dated test. Annealing note

5. Preserve infrastructure option value, then specialize deliberately

• Historical rule. [researcher inference | moderate] Flexible cloud/API capacity is valuable under uncertainty; predictable scaled workloads can justify reservation, specialization, or internalization. Cost of Cloud · AI compute
• Diagnostics. [researcher inference | moderate] Unit workload cost, utilization, egress, data gravity, provider discount, migration cost, reliability staff, depreciation, supplier leverage, multicloud portability. Cloud note
• Kill test. [researcher inference | moderate] An internalization plan whose savings disappear after migration capital, idle capacity, staffing, or reliability is rejected. Cloud note
• Modern boundary. [researcher inference | moderate] Compare hosted, reserved, specialized, and owned scenarios per workload; do not migrate an entire company on slogan economics. Compute note
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: preserve flexibility under demand uncertainty and compare lifecycle economics per stable workload before specializing; does not transfer: the 2021 multiplier, historical cloud discounts, or a company-wide repatriation prescription; changed assumptions: accelerator scarcity, specialized clouds, egress terms, sovereign requirements, and bursty AI demand reshape the frontier; observable signals: utilization, reserved-price capture, egress, reliability labor, migration capital, payback, and portability; likely misuse: turning repatriation or multicloud into ideology; current example: Dropbox's FY2025 filing grounds the continuing materiality of infrastructure cost but does not establish a universal owned-versus-cloud answer. Cloud note · Dropbox FY2025 10-K, cost-of-revenue and infrastructure discussion

6. Treat a stack map as a bargaining-power map

• Historical rule. [researcher inference | moderate] Identify the indispensable workload and which layer owns distribution, switching cost, compute, data, and customer relationship. Data architecture · GenAI platform
• Diagnostics. [researcher inference | moderate] Supplier/customer concentration, exclusivity, revenue share, control rights, information flow, cloud commitments, portability, platform-bundling response. Stack note
• Kill test. [researcher inference | moderate] A startup that is a replaceable feature with one dominant supplier and no workflow/distribution control lacks a venture-scale control point. Stack note
• Modern boundary. [researcher inference | strong] Include contractual power documented by the FTC, not only technical diagrams. FTC report, §§4–5
• Modern VC translation (as of 2026-08-19). [researcher inference | strong] Transfers: map technical dependencies together with distribution, contract, control, information, and switching rights; does not transfer: 2023 layer shares, vendor rankings, or the prediction that one layer will capture value; changed assumptions: model-provider partnerships, exclusivity, cloud credits, open weights, and platform bundling can move power without moving the diagram; observable signals: concentration, control rights, revenue share, portability tests, renewal leverage, and alternative suppliers; likely misuse: equating a visible stack position with durable economics; current example: the FTC's 2025 report documents contractual control mechanisms but does not establish final value capture or startup returns. Stack note · FTC report, §§4–5

7. Use architecture maps as hypotheses, not portfolio construction

• Historical rule. [researcher inference | moderate] Match products to customer maturity and the simplest adequate blueprint; more control is not automatically better. 2020 data blueprints
• Diagnostics. [researcher inference | moderate] Durable workload, operational necessity, integration burden, buyer, platform substitution, category survival, total system complexity. Architecture note
• Kill test. [researcher inference | moderate] If convergence or bundling removes the independent budget/control point, the mapped category is not a company. Architecture note
• Modern boundary. [researcher inference | strong] Rebuild maps from current deployments; do not infer position size, diversification, or return from sector coverage. Architecture note
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: use an architecture map to generate falsifiable workload and control-point hypotheses; does not transfer: the 2020 boxes, category boundaries, or any portfolio weight/return conclusion; changed assumptions: platform convergence, agents, open standards, and bundling can erase independent categories; observable signals: standalone budget, operational necessity, retention, integration burden, platform substitution, and category survival; likely misuse: funding every box or treating map coverage as diversification; hypothetical: a new data-observability vendor qualifies only if it retains an independent budget and workflow after the warehouse bundles adjacent features. Architecture note

8. Govern open models through an evidence-and-monitoring loop

• Historical rule. [researcher inference | moderate] Openness can diffuse innovation and competition; new restrictions should respond to measured marginal risk and gaps in existing law. Base AI Policy
• Diagnostics. [researcher inference | moderate] Artifact opened, licenses, capability, misuse exposure, provenance, evaluation, distribution reach, affected population, legal gap, severity, reversibility, monitoring trigger. Policy note
• Kill test. [researcher inference | moderate] “Open” without a definition or “safe” without a denominator does not support a policy or investment conclusion. Open-code note
• Modern boundary. [researcher inference | strong] Preserve the ability to tighten controls when monitored evidence changes; NTIA’s 2024 recommendation was continued evidence-building, not permanent nonregulation. NTIA report, recommendations
• Modern VC translation (as of 2026-08-19). [researcher inference | strong] Transfers: define the opened artifact, compare marginal risk with alternatives, identify legal gaps, and precommit monitoring/escalation triggers; does not transfer: a permanent presumption that openness is safe or that 2024 evidence settles later capabilities; changed assumptions: better distillation, tool use, distribution, and incident evidence can alter both benefit and misuse exposure; observable signals: evaluation deltas, exploit/misuse incidents, reach, severity, reversibility, affected populations, and enforcement gaps; likely misuse: treating portfolio interest or open status as societal-outcome evidence; current example: NTIA's 2024 evidence-building recommendation is a grounded policy process example, not proof that 2026 controls should remain unchanged. Policy note · NTIA report, recommendations

9. Diligence physical AI as an industrial system

• Historical rule. [researcher inference | moderate] Robotics performance depends on component precision, supplier proximity, integration, deployment access, and learning data—not models alone. Robotics essay
• Diagnostics. [researcher inference | moderate] Bill of materials, country/supplier concentration, yield, lead time, field failures, repair, teleoperation, utilization, customer-site approvals, data rights, working capital, policy milestones. Robotics note
• Kill test. [researcher inference | moderate] A demo whose production yield, component access, field reliability, or customer deployment loop cannot meet milestones is not validated by model capability. Robotics note
• Modern boundary. [researcher inference | moderate] Treat allied capacity and deregulation as scenarios; model cost, schedule, safety, labor, and security rather than assuming policy delivery. Robotics note
• Modern VC translation (as of 2026-08-19). [researcher inference | moderate] Transfers: diligence the full component-to-deployment learning system and finance its working-capital path; does not transfer: the essay's country statistics, component shares, price examples, or deregulatory outcome assumptions; changed assumptions: export controls, tariffs, automation, supplier localization, insurance, and customer-site rules affect scale; observable signals: qualified suppliers, yield, lead time, field reliability, repair, utilization, deployment permission, data rights, and cash conversion; likely misuse: treating a software benchmark or demo as manufacturing readiness; current example: the essay's World Labs reference is a financially interested illustration and supplies no independent production, deployment, or outcome record. Robotics note · a16z profile, current board list

Best five writings for reconstructing the method

1. [researcher inference | strong] “The New Business of AI” (2020) — clearest example of translating technical failure modes into a cost taxonomy and falsifiable operating economics. Note
2. [researcher inference | strong] “The Cost of Cloud” (2021) — strongest lifecycle and option-value framework, with enough assumptions to challenge rather than merely repeat. Note
3. [researcher inference | strong] “Market Annealing” (2023) — most explicit account of founder learning, sequencing, and the transition from discovery to repeatable go-to-market. Note
4. [researcher inference | strong] “The Empty Promise of Data Moats” (2019) — durable anti-slogan test that also supplies the boundary under which later data-flywheel claims can be true. Note
5. [researcher inference | strong] “Who Owns the Generative AI Platform?” (2023) — most useful stack-level bargaining-power map and a clean historical information set for testing later concentration. Note

The five are selected for decision transparency and cross-corpus leverage, not because public evidence shows they produced the best investments. researcher inference | strong.

Missing evidence and continuation routes

• [researcher inference | insufficient public record] Full licensed bodies or authorized archives for the WSJ and Economist essays; until then, their source notes and analysis remain partial. WSJ note · Economist note
• [researcher inference | insufficient public record] Original IC memos, deal calendars, sourcing records, passes, sponsor/approval roles, checks, security terms, ownership, reserves, and follow-on decisions tied to these frameworks. Source-map retrieval ledger
• [researcher inference | insufficient public record] A complete opportunity and outcome denominator, including companies that followed similar maps but failed, plus realized fund cash flows. Candidate-deal boundary
• [researcher inference | insufficient public record] Named company-level evidence for the 2020 AI cost sample, 2023 annealing benchmarks, 2023 platform economics, and 2025 margin bridge. AI note · annealing note · platform note · margin note
• [researcher inference | insufficient public record] Longitudinal tests of the 2020 data-infrastructure categories and predictions against acquisitions, bundling, shutdowns, and durable independent vendors. Architecture note
• [researcher inference | insufficient public record] Workload-level cloud/owned cost studies including engineering, utilization, reliability, migration, facilities, and capital—not gross-margin anecdotes alone. Cloud note
• [researcher inference | insufficient public record] Casado/a16z lobbying, policy, portfolio-exposure, and conflict-management records for AI openness, regulation, and robotics advocacy. Policy note · robotics note
• [researcher inference | insufficient public record] Robotics production evidence: yield, unit cost, field reliability, supplier concentration, deployment utilization, data rights, and working-capital/exit-duration fit. Robotics note

Until those records are public, the written corpus supports a testable analytical vocabulary and a plausible architecture-to-economics translation skill. It does not support a personal hit rate, causal portfolio advantage, audited governance regime, or realized-return claim. researcher inference | insufficient public record.

Coverage and evidence base

The TALKS corpus reconciles 31 priority appearance families: 11 ★★★, 19 ★★, and one ★. Every family has one notes/t-*.md record, and all 31 are substantively processed from a complete exact recording, transcript, or captions. [contemporaneous record | strong] for inventory/access; [researcher inference | strong] for family boundaries. Source map, priority talks and supporting routes

Several publisher versions are not independent talks. Turpentine E70 replays the June 2024 Cognitive Revolution interview; the December 2025 a16z “AI Boom” episode replays the October Generalist interview; the February 2026 capital-flywheel episode replays the Latent Space original; the March 2026 monopolies wrapper is the July 2025 20VC interview distributed by a16z in August; and the September 2023 generative-AI presentation is a spoken adaptation of the August Casado–Sarah Wang essay. Repetition establishes emphasis or a medium/version boundary, not independent confirmation. [researcher inference | strong]. E70 note, Contradictions · AI Boom note, Contradictions · Capital Flywheel note, Contradictions · Monopolies note, Contradictions · presentation note, Contradictions

Where no creator transcript existed, the exact publisher asset was bound by URL, GUID or platform ID, runtime, and clock. Third-party captions or ASR were navigation aids; uncertain speech was paraphrased, speaker turns were preserved, and no machine-only wording was promoted as a quotation. No raw audio, video, captions, or transcript artifact is stored in the investor directory. [contemporaneous record | strong]. Coverage map, coverage assessment · 20VC access note, Contradictions · Future Software note, Contradictions

The corpus is unusually good at mechanisms and self-correction: Casado explains how technical discontinuities become company searches, why usage can overturn a theory, why founder roles change across company phases, and why stage and vehicle shape an investment rule. It remains an interested record produced largely by a16z, its portfolio, or friendly interviewers. Repetition does not prove selection alpha, portfolio returns, AI market structure, policy welfare, or that an example caused an outcome. [researcher inference | strong]. Non-Consensus note, Contradictions · Future VC note, Contradictions

Source-by-source guide

Infrastructure… Is Everything (2016-02-25)

Casado connects LLNL simulation, post-9/11 intelligence work, Stanford networking, and Nicira to a search lens: newly programmable technical layers can create companies, but only when a concrete application, user, buyer, and delivery path exist. This is formation evidence, not proof that technical depth predicts returns. [retrospective witness | moderate]. Note, 01:19–10:07 and 32:13–34:38

Selling to Developers & Open Source Business Models (2016-04-13)

Developer adoption lowers discovery and insertion friction but does not identify the economic buyer or paid control point. [investor-stated | strong]. Casado's strongest direct behavior is an organizational choice at Nicira—separate open and proprietary teams with a senior arbiter—while Peter Levine supplies much of the broader sales and Red Hat argument. [documented behavior | moderate]. Note, 06:47–16:27 and 21:21–24:58

Pricing, Pricing, Pricing (2016-08-13)

Pricing is treated as market discovery: require paid trials, connect price to the sales motion, and learn whether a founder can move from early hunters to repeatable selling. Mark Cranney owns much of the packaging and contracting playbook; Casado's evidence is retrospective Nicira practice. [investor-stated | moderate]. Note, Takeaways

Trends in—and the Future of—Infrastructure (2016-08-21)

The initial sector map joins three shifts—hardware to software, installed product to service, and operator procurement to developer influence. The investable mechanism is removal of incumbent advantages; developer attention remains top-of-funnel, not revenue or defensibility. [investor-stated | strong]. Note, 00:45–20:38

All About Microservices (2016-09-01)

Microservices create richer application semantics and new security, observability, routing, and operational gaps, but a technical function is not automatically a company. Adrian Cockcroft owns the Netflix operating history; Casado owns the infrastructure-market interpretation and its TAM boundary. [investor-stated | moderate]. Note, Takeaways

Software Secures the World (2017-02-12)

Casado reframes security from a simple arms race to a workflow-misalignment problem: point tools create alerts faster than human teams can process them, so automation and integrated response may become the new product layer. Company examples and inbound figures are selected illustrations without a losing-company denominator. [investor-stated | moderate]. Note, 01:20–18:27

Going to Market When No Market Exists (2017-06-09)

A new market lacks a settled budget, comparison set, buyer, price, and repeatable process. Casado prescribes founder-led discovery, explicit narrative, high initial anchoring, and later sales repeatability; the talk is a direct precursor to “market annealing,” not evidence that category declaration causes adoption. [investor-stated | strong]. Note, 00:52–25:22

Navigating Uncertainty (2017-07-02)

The graduation talk adds a personal judgment rule: preserve optionality, act before uncertainty disappears, and update after failure. Its genre and survivor perspective make it useful biography but weak investment-efficacy evidence; the separate closing speaker is excluded from Casado attribution. [investor-stated | moderate]. Note, Casado 00:04–06:56

Creating a Category, from Pricing to Positioning (2018-02-24)

Casado separates category creation from proof of market: map the buyer's existing budget and mental model, delay price until technical value is established, then require production use, repeated use cases, and non-founder selling before calling founder-generated interest product-market fit. [investor-stated | strong]. Feaster's early-customer, product-marketing, and competitive-playbook claims remain separately attributed. [contemporaneous record | strong]. Note, Casado 02:06–07:38 and 13:41–22:30

The World Through an API (2018-03-09)

Casado generalizes disaggregation: as a technical ecosystem matures, smaller functions can support products, infrastructure vendors, and aggregators. Kong/Mashape, RapidAPI, and SigOpt are interested company illustrations, not evidence of Casado's sponsorship, check, ownership, or return. [investor-stated | strong]. Note, 01:01–16:28

Software Chasms (2019-01-28)

The most complete early operating framework is that technical companies cross repeated chasms—problem definition, product, delivery, integration, and enterprise sales. Founder-led sales is diagnostic; open source and services are tools with costs, not shortcuts. [investor-stated | strong]. Note, publisher transcript

Gross Margins, Early to Late (2020-06-26)

Gross margin is a phase- and architecture-dependent diagnostic rather than an early universal cutoff. Casado focuses on technical COGS, cloud reversibility, implementation, and AI/ML costs; later-stage benchmarks belong partly to co-speakers. [investor-stated | strong]. Note, Takeaways

On Fear and Leadership (2020-12-04)

Casado supplies rare decision evidence: he declined Nicira's CEO role, later regretted an interim arrangement, hired Steve Mullaney permanently, and describes the hire as probably his best company decision. [retrospective witness | strong]. The broader rule is to match product, sales, and operations leadership to the current bottleneck without making fear the decision criterion. [investor-stated | strong]. Note, 09:07–29:30

Cloud Cost (2021-06-25)

Cloud economics move from reversible startup tradeoff to potential COGS and strategic constraint as usage scales. The source supports scenario modeling, workload segmentation, negotiated-price and migration boundaries—not a universal repatriation threshold. [investor-stated | strong]. Note, transcript pp. 3–20

Past, Present, and Future of Digital Infrastructure (2022-06-07)

Casado combines cloud cost, workload specialization, fine-grained consumption, and open-standard boundaries with a valuable correction loop: reported Hugging Face usage challenged the team's theory. The case is process testimony; the episode supplies no investment memo, sponsor, price, ownership, or return. [documented behavior | moderate]. Note, 06:49–34:18

Cloud Wars, Company Wars, and Innovating Through Change (2022-09-07; recorded 2021)

The second cloud war moves infrastructure inside software COGS, while company transitions require deliberate cannibalization and organization change. The recording predates publication and includes Michael Dell and Steven Sinofsky; their operating claims must remain separate. [investor-stated | moderate]. Note, Takeaways

20VC — whether the venture model is broken (2022-12-05)

Casado argues for a barbell of scaled multi-product platforms and focused boutiques, with research performed before a fast deal and separate teams across risk buckets. He also defines a board boundary—supply market maps and scarce functions, not operating commands. [investor-stated | strong]. His admission of early overfitting is retrospective evidence about his own practice. [retrospective witness | moderate]. Note, 09:30–39:30

The Economic Case for Generative AI — presentation (2023-09-25)

The talk explains why earlier AI startups struggled with tail accuracy, human exception handling, narrow markets, hardware, and cheap human substitutes, then argues foundation models change cost and market breadth. It is a solo adaptation of the August Casado–Wang essay and therefore a processed spoken version, not independent corroboration. [investor-stated | moderate]. Note, 00:00–14:39

Why America Must Lead in AI Investment (2024-03-05)

Casado's questions favor investment, talent, and application-level regulation; Senator Todd Young supplies many of the specific policy prescriptions and acknowledges low-probability, high-cost risks. The useful boundary is cautious enablement, not a no-regulation position or Casado-only authorship. [investor-stated | moderate]. Note, Takeaways

Cognitive Revolution / Turpentine E70 — AI Innovation and AGI (original 2024-06-28; replay 2024-12-24)

Casado emphasizes heavy-tail reliability, a moving AGI definition, application-level regulation, and a falsifiable prediction of model-market fragmentation; Nathan Labenz supplies material counterarguments on abstraction, scaling, and unknown capability. E70 is one shortened replay, not a later forecast. [investor-stated | moderate]. Note, original clock

Entering Uncharted AI Territory (2024-07-09)

The interview combines model economics, market structure, overbuild risk, and technical-founder bias. [investor-stated | strong]. Casado's admission that drone/robotics boards revealed his lack of vertical expertise is retrospective evidence of a domain boundary changing the firm's specialist coverage. [retrospective witness | moderate]. Note, 23:13–30:31 and 42:28–58:41

AI's Hard Limits and Open Problems (2024-09-02)

Casado treats machine learning as a third programming primitive while stressing correctness, memory, state, information, and human-control limits for agents. He simultaneously favors public research compute, application-level rules, and M&A availability; the policy-incidence claims remain untested here. [investor-stated | strong]. Note, 20:12–42:42

Where Value Will Accrue in AI (2025-05-27)

Casado disaggregates AI into many markets and says model capability solves a bootstrap problem, not retention or moat. Sarah Wang—not Casado—supplies the sharpest warning about oversized pre-traction rounds and category conflicts. [investor-stated | strong]. Note, 02:45–20:55

AI Is Upending SaaS Pricing (2025-07-18)

Seat pricing breaks when AI reduces users while increasing work, so the metric must move toward output or value. The practical playbook is jointly shaped by Metronome CEO Scott Woody and is financially interested; the episode provides no matched retention, margin, or valuation cohort. [investor-stated | strong]. Note, Takeaways

The Future of Software Development (2025-07-21)

Casado calls models a new programming/infrastructure layer but rejects a simple replacement story: probabilistic logic adds verification, data, networking, latency, and guarantee requirements. Co-speakers own distinct product and workflow claims, and external productivity evidence is mixed. [investor-stated | strong]. Note, Takeaways

20VC / Monopolies vs. Oligopolies in AI (original 2025-07-28; later wrappers)

Casado starts from observed users, revenue, and value creation, then underwrites both monopoly and oligopoly paths by submarket. The same interview gives the clearest current ownership rule, conflict admission, approximately $1.2 billion infrastructure-fund reference, and a bounded 2×—not 10×—coding-productivity intuition. [investor-stated | strong]. Note, a16z feed-drop clock

State of American AI Policy: From Pause AI to Build (2025-08-15)

Casado characterizes a policy shift from precaution toward national building, while co-speakers supply distinct legal and political claims. The note tests the rhetoric against primary policy documents; “build” remains advocacy by an investor with AI exposure, not a welfare result. [investor-stated | moderate]. Note, Takeaways

Is Non-Consensus Investing Overrated? (2025-09-04)

Casado narrows his viral claim: products should often be non-consensus to customers, but investors must understand informed objections and future financing consensus. He invites a cohort test, admits it is unfinished, and makes stage/check size part of the rule. [investor-stated | strong]. Note, 00:00:30–11:04 and 43:06–53:06

Jack Altman and Martin Casado on the Future of VC (original 2025-09-03; a16z republication 2025-09-04)

The Uncapped feed drop describes specialist market coverage, media/distribution, conflict lanes, staying power, and a16z's multi-product design. It is valuable firm-process testimony but also explicit advocacy for the platform Casado helps lead. [investor-stated | strong]. Note, Takeaways

“This Feels Like 1996” (original 2025-10-07; a16z replay 2025-12-30)

Casado argues that enterprise adoption lags individual capability, AI has no inherent moat, and the boom is still early rather than late-dot-com overbuilt. “Markets in” is his current sequence—map structural demand, then select the team. [investor-stated | strong]. No failed-map cohort tests the claimed edge. [researcher inference | insufficient public record]. Note, Takeaways

AI's Capital Flywheel (2026-02-19)

Casado and Sarah Wang describe a capital→compute→capability→demand→revenue loop, vertical expansion risk, and two market futures: fragmented applications or model oligopoly. [investor-stated | strong]. GPU utilization is not independent customer cash, and current revenue may borrow against the next training run. [researcher inference | strong]. Note, Takeaways

Evolution, contradictions, and present translation

The dated evolution judgments in this table are [researcher inference | strong]; each row links the direct spoken records from which the comparison is reconstructed.

The main tensions and usable reconciliations are [researcher inference | strong] unless a row states otherwise:

1. Developer adoption versus the economic buyer. Developer love can discover and insert a product, while budget authority, procurement, security, and enterprise sales decide monetization. Treat usage as a falsification/learning signal, not revenue proof. Selling note, Contradictions · Digital Infrastructure note, Contradictions
2. Open source versus control and scale. Open source can create trust and distribution, but services burden, governance, cloud capture, and the paid control point determine venture economics. Separate community health from ownership and margin. Selling note, Contradictions · Software Chasms note, Contradictions
3. Infrastructure transformation versus cloud consolidation. Casado first predicts startup opportunity as layers become programmable, then documents hyperscaler leverage and cloud COGS, and later reopens specialized workloads. Underwrite the workload-specific cost/performance curve and portability rather than a universal rebundling cycle. Trends note, Contradictions · Cloud Cost note, Contradictions
4. Category creation versus trend alignment. Founder will and narrative can create a market, but a shrinking trend or absent financing path can turn the exercise into a decade-long death march. Require named buyers, budget formation, sales repetition, and runway. GTM note, Contradictions · 20VC note, Contradictions
5. Software margins versus irreversible COGS. Early software rhetoric coexists with later cloud, inference, implementation, supervision, and next-model-training costs. Gross margin is an architecture and phase diagnostic, not a categorical software entitlement. Gross Margins note, Contradictions · Flywheel note, Contradictions
6. Technical-founder preference versus commercial and domain leadership. Casado's systems formation sharpens technical diligence, while his own Nicira CEO decision and robotics-board admission show where sales, operations, or vertical expertise dominate. Record the missing competency and decision right rather than universalize founder type. Fear/Leadership note, Contradictions · Uncharted note, Contradictions
7. Market-first search versus retrospective category selection. Space mapping can improve preparedness, but flexible market definitions make winners look inevitable after the fact. Preserve dated maps, passes, ranking changes, and losing categories. 20VC note, Contradictions · AI Boom note, Contradictions
8. Contrarian product versus financing consensus. Non-consensus customer behavior can create the product opportunity; informed investor objections and the next financing audience are survival constraints. The rule is stage- and vehicle-specific and may favor a large networked firm. Non-Consensus note, Contradictions
9. Model fragmentation versus oligopoly. Distillation, diminishing returns, and specialized tasks support fragmentation; capital intensity, subsidy, platform distribution, and vertical expansion support concentration. Define submarket, share, switching, and loss thresholds ex ante. E70 note, Contradictions · Monopolies note, Contradictions
10. AI bootstrap versus durable moat. Model capability can create magical acquisition, but workflow, integration, data, marketplace, distribution, brand, or cost advantage must retain the user after model parity. Value Accrual note, Contradictions · AI Boom note, Contradictions
11. Agent productivity versus systems limits. Code generation can remove accidental complexity, while correctness, memory, state, verification, and human review limit autonomous deployment. Measure end-to-end release quality and intervention, not generated tokens. Future Software note, Contradictions · Hard Limits note, Contradictions · Monopolies note, Contradictions
12. Build policy versus harm and regulatory incidence. Casado favors investment and application-level rules; that does not establish which applications, liability regimes, market-power controls, export rules, or low-probability risks warrant intervention. America/AI note, Contradictions · Policy note, Contradictions
13. Scaled specialist platform versus conflicts. Specialist coverage and multiple capital products can improve access and stage fit; they also enlarge category conflicts, signaling power, fees, and internal allocation questions. Separate teams are a control claim, not an audited conflict outcome. 20VC note, Contradictions · Future VC note, Contradictions · Monopolies note, Contradictions
14. Capital flywheel versus circular financing. Used GPUs and rising model revenue show activity, not independent demand or return. Track end-customer cash, strategic recycling, training-adjusted margin, next-round dependence, and supplier credits. Flywheel note, Contradictions
15. “1996” supercycle versus overbuild. Enterprise adoption lag may extend the cycle, while calendar analogy and utilization cannot rule out subsidized capacity or weak ROI. Define demand quality, capacity utilization, retention, and financing dependence before using the analogy. AI Boom note, Contradictions

Biography, decisions, vehicles, and attribution leads

The disciplined use of each lead is [researcher inference | strong] unless otherwise labeled.

The spoken corpus also contains unreconciled Nicira/VMware operating figures—approximately $600 million in some accounts and approximately $1 billion in the 2017 go-to-market talk. They may refer to different dates, measures, or business scopes; none is an audited Casado-specific outcome. [retrospective witness | provisional]. GTM note, Biography and decision evidence · 20VC note, Biography and decision evidence

Five sources to study first

The ranking is [researcher inference | strong], based on source completeness, mechanism density, self-correction, and direct decision evidence—not prestige.

1. Software Chasms (2019), Takeaways — best bridge from technical-founder experience to repeated product, delivery, integration, open-source, and enterprise-sales tests; full publisher transcript.
2. 20VC — Venture Model (2022), Takeaways — best direct account of space mapping, team debate, stage-specific vehicles, board-role limits, scenario planning, and scaled-firm advocacy.
3. Past, Present, and Future of Digital Infrastructure (2022), Takeaways — best investor correction loop: usage over theory, workload specialization, consumption pricing, cloud cost, and open-standard boundaries.
4. Entering Uncharted AI Territory (2024), Takeaways — best AI-era combination of economics, technical moat, category risk, domain-limit admission, and explicit willingness to be wrong.
5. Is Non-Consensus Investing Overrated? (2025), Takeaways — best falsifiable decision-process discussion, with informed objections, financing consensus, stage/check-size boundary, and an unfinished promised cohort test.

On Fear and Leadership, Takeaways is the best biography-to-decision supplement; Monopolies vs. Oligopolies, Takeaways the best current ownership/conflict/AI-market supplement; and Capital Flywheel, Takeaways the best current capital/compute dependency supplement.

Evidence still missing

• No complete personal deal ledger, opportunity/pass denominator, dated market-map archive, IC memorandum set, partner vote/veto record, check/security/ownership schedule, reserve history, fund cash flow, loss series, or realized person-level attribution record surfaced. “Almost 200 investments” cannot support hit rate, MOIC, IRR, DPI, or alpha. [researcher inference | strong]. Non-Consensus note, Open questions · Source map, vehicle and deal-attribution records
• The corpus cannot show that specialization, technical formation, market mapping, usage sensitivity, consensus awareness, or a16z services caused better selection or returns. Needed evidence is an ex-ante scored cohort including passes, write-offs, missed winners, stage, access, price, ownership, and matched baselines. [researcher inference | strong]. 20VC note, Open questions · Future VC note, Open questions
• Developer adoption, open-source popularity, category narrative, and fast AI usage remain upstream signals. Buyer authority, paid control point, retention, direct-sales burden, conflict, gross margin, and capital intensity can break the conversion to venture outcomes. [researcher inference | strong]. Selling note, Open questions · Value Accrual note, Open questions
• No public cloud-repatriation crossover rule or AI gross-margin cohort supplies workload costs, discounts, staffing, depreciation, egress, migration, inference, retries, supervision, training allocation, or later outcome. [researcher inference | strong]. Cloud Cost note, Open questions · AI Pricing note, Open questions
• The model market can fragment, monopolize, or oligopolize depending on submarket definition, capital, subsidy, distillation, distribution, proprietary data, and switching. The talks supply competing mechanisms, not a settled outcome. [researcher inference | strong]. E70 note, Open questions · Monopolies note, Open questions
• “1996,” “no dark GPUs,” and AI capital-flywheel claims still need independent utilization, end-customer cash, revenue quality, strategic-credit, training-adjusted margin, and financing-dependence evidence. Private financing and infrastructure use are not investor proceeds. [researcher inference | strong]. AI Boom note, Open questions · Flywheel note, Open questions
• AI developer productivity needs task-level evidence across greenfield and mature codebases, time, defects, review, maintainability, deployment, and intervention. Casado's hands-on use and company anecdotes do not establish general productivity or labor displacement. [researcher inference | strong]. Future Software note, Open questions · Monopolies note, Open questions
• Policy advocacy lacks a complete application-risk taxonomy, causal incidence analysis, company-by-company financial-interest map, public-cost accounting, and triggers for changing position. Public investment and regulatory restraint are compatible only after objectives, externalities, authority, and accountability are specified. [researcher inference | strong]. America/AI note, Open questions · Policy note, Open questions
• The spoken record cannot establish that Nicira's permanent CEO hire, founder-led selling, open/proprietary split, cloud-cost monitoring, or a16z platform services caused success. Decision participants, contemporaneous documents, customers, employees, and counterfactuals remain thin. [researcher inference | strong]. Fear/Leadership note, Open questions
• The category-creation heuristics still lack a failed-category denominator, named portfolio outcome cohort, and evidence that delayed price, the $200,000–$300,000 direct-sales hurdle, or the three product-market-fit signals predict retention or returns. [researcher inference | insufficient public record]. Creating Category note, Open questions

[researcher inference | strong] The defensible spoken-corpus through-line is diagnostic rather than predictive: identify a technical discontinuity; map the workflow, user, buyer, and budget; let usage challenge the map; separate insertion from monetization; match leadership and sales motion to the bottleneck; document irreversible costs; and underwrite financing, ownership, conflicts, and supplier dependencies. The record supports that coherent operating system. It does not prove personal returns, complete attribution, policy welfare, or a settled AI market structure. Software Chasms, Takeaways · 20VC, Takeaways · Non-Consensus, Takeaways · Monopolies, Takeaways

Model 1: Map the market, then let observed use and relative diligence falsify the map

• Rule: A technical shift can reveal necessary workloads before the winning product is legible. A map narrows where to look; it cannot tell the investor which box becomes a durable company. Real deployments, retained use, budget, and relative company evidence must be allowed to overturn the taxonomy.
• Evidence status: [investor-stated · strong] for Casado's stated market-first search, “usage graph over theory” correction, and architecture-mapping practice; [researcher inference · moderate] for the combined procedure and any claim that it improved selection. No public comparative performance test exists.
• Provenance and origin: The intellectual root is Casado's network-systems work: begin with a technical discontinuity and the functions it makes possible, rather than a company name. In investing, the 2020 data-infrastructure essay turned that habit into explicit architecture blueprints after two years of research; the 2022 Past, Present and Future of Digital Infrastructure interview then supplied the important correction that Hugging Face's usage contradicted the team's theory; the 2022 venture-model and 2025 non-consensus interviews made market-first research and company-relative diligence explicit. 2020 architecture essay, opening through “Blueprints” · 2022 infrastructure interview, Casado at 00:29:03–00:31:48 · 2022 venture-model interview, 18:47–19:22 and 33:35–36:17
• Mechanism and reasoning: Mapping concentrates search and builds a common language for workload, buyer, incumbent response, and control point. Usage supplies a reality check. Comparing companies inside the same problem space makes team, product, go-to-market, and architecture differences more legible than a one-off “is this good?” judgment. Informed dissent reduces the risk that a senior partner's elegant map becomes the answer. Architecture essay, “Blueprints” and “Looking ahead” · Digital Infrastructure, Hugging Face correction at 00:29:03–00:31:48 · Venture-model interview, research/debate at 18:47–19:22 and 33:35–36:17 Assumptions: The discontinuity is real; the mapped workload is operationally necessary; public or diligence usage is not purely subsidized; comparable alternatives exist; the team can observe disconfirming evidence; and incentives allow the thesis to be revised rather than defended.
• Historical and vehicle context: This model spans Casado's operator formation in network systems, the 2020 architecture-mapping work, the 2022 Hugging Face correction, and 2025 investing-process interviews. 2020 architecture essay, opening and blueprints · 2022 correction, 00:29:03–00:31:48 · 2025 non-consensus debate, 00:06:38–00:11:04 Distributional, ActionIQ, and Astranis are public a16z thesis records, but none is mapped publicly to a specific fund, check, security, ownership target, reserve decision, or IC vote. Distributional · ActionIQ · Astranis Hugging Face is process testimony, not a disclosed deal or return.
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: production use rather than pilots; repeated workload across customers; retained or expanding usage; a named buyer and budget; referenceability; lower total system complexity; a product surviving adjacent-platform bundling; an outlier whose use persists despite contradicting the initial theory. Do not use it as a selection or sizing rule when no independent workload/budget exists, usage is subsidy-driven, or consumer taste and regulation dominate technical topology.
• How to apply it: (1) Draw the current workflow and stack from actual deployments. (2) Mark necessary but unstable functions, independent budgets, incumbents, and likely bundling paths. (3) Build a same-problem comparison set, including a pass/do-nothing alternative. (4) gather retained-use, production, reference, and willingness-to-pay evidence. (5) Write the evidence that would erase a box or elevate an outlier. (6) Debate the strongest disconfirming case before discussing valuation and ownership. (7) Rebuild the map at each follow-on; never use sector coverage as portfolio construction.
• Examples:
  1. Documented process correction — Hugging Face. [documented behavior · moderate] Casado said the usage graph conflicted with the team's theory and taught them to trust observed adoption more. The account does not say whether a16z invested at that moment, who changed the decision, or what return followed. Casado at 00:29:03–00:31:48
  2. Documented diligence — Distributional. [investor-stated · strong] The announcement connects an existing SigOpt relationship with more than 25 customer interviews; that is market- and company-level triangulation, not proof the process selected the best company. Distributional announcement, founder relationship and interview paragraphs
  3. Documented diligence — ActionIQ. [investor-stated · strong] Casado tied prior Stanford knowledge of the founder to interviews with roughly a dozen enterprises about the customer-data problem. No alternative set, pass record, or deal economics is public. ActionIQ announcement, founder and enterprise-interview paragraphs
  4. Documented learning boundary — Astranis. [investor-stated · strong] Casado said he began outside his expertise and learned the satellite, regulatory, and market system before investing. A later supplier-caused mission anomaly tests the map but does not establish that the map caused the decision. Astranis announcement, “When we first heard” through board paragraph · Astranis anomaly, reproduced-failure paragraphs
• Evidence: The architecture essay establishes the map; the digital-infrastructure interview documents the usage-over-theory correction; the venture-model interview documents pre-deal research and internal debate; and the company posts establish only the named diligence facts. Architecture essay · Digital Infrastructure, 00:29:03–00:31:48 · Venture-model interview, 18:47–19:22 and 33:35–36:17 They do not prove causal selection skill or investor return. Exact evidence gaps: No original category maps used in a named IC; no company-comparison scorecards; no list of alternatives or passes; no vote/dissent record; no pre/post-Hugging Face process document; no matched mapped-versus-unmapped investment outcomes; and no ownership or return data showing economic value from the method.
• Limits, failure modes, and exception ledger: Limits and exceptions: Strong early use can be free, curiosity-driven, or concentrated; architecture maps can be self-fulfilling firm marketing; same-space comparison may miss an orthogonal entrant; and the public record contains no complete passes or category cohorts. Hugging Face is evidence of one correction, not a measured correction rate. Architecture mapping is least useful where consumer taste or regulation, rather than technical topology, determines the market. Failure modes: funding every box; confusing category salience with a company; privileging a partner's theory over deployment data; mistaking subsidized or free usage for demand; allowing a map to freeze while platform boundaries move; or using “market-first” to hide that no comparative company work was documented. Falsifiers: the workload disappears into a platform bundle; customers will not allocate a standalone budget; retained use collapses after subsidy or fashion; total integration burden rises; or repeated company comparisons show the mapped variable has no relationship to adoption, economics, or outcomes.
• Evolution and contradictions: The method moves from a founder-engineer's system map, to a 2020 category taxonomy, to a 2022 admission that adoption can defeat theory, to a 2025 emphasis on non-consensus relative judgment. The update is meaningful: maps become falsifiable search instruments rather than predictions. Architecture essay · Digital Infrastructure, 00:29:03–00:31:48 · Non-consensus debate, 00:06:38–00:11:04 · Relative diligence, 01:00:32–01:05:27 The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: map workflows and contract power, then require a dated map-change memo. Does not transfer: static taxonomies or architecture-as-capture. Changed assumptions: concentrated model/cloud suppliers, faster bundling, regulation, and physical-system dependencies. Observable signals: retained production use, an accountable buyer/budget, independent references, and survival after bundling. Likely misuse: financing every attractive box. Current grounded example: the 2026 Runta thesis makes agent-compute architecture investable only as a public thesis; it does not establish capture or return. Runta, relationship and agent-compute thesis

Model 2: Trace a continuous user-to-buyer-to-budget-to-expansion chain

• Rule: Product use proves that somebody can obtain value; it does not prove a company has a budget owner, procurement path, premium trigger, or profitable sales motion. An investable bottom-up business needs a continuous causal path from activation to retained use, within-account spread, buyer-approved purchase, and economic expansion.
• Evidence status: [investor-stated · strong] for Casado's repeated distinction between technical adoption and the buying center, plus the coauthored growth+sales and bottom-up-pricing frameworks; [researcher inference · moderate] for deal-level application and portfolio efficacy.
• Provenance and origin: The 2016 developer-selling discussion treated open source as a technical trust and insertion mechanism, not a complete business model. The 2019 Software Chasms interview sharpened bottom-up entry plus enterprise conversion. The 2020 Growth+Sales essay supplied six failure modes, and the 2021 pricing essay mapped the user journey through free, professional, team, and enterprise tiers. 2016 podcast, 06:47–16:27 · 2019 interview, 00:26:48–00:30:39 · Growth+Sales, “New buyer, new motion” and Failure Modes #1–#6 · Bottom Up Pricing, sections #1–#9
• Mechanism and reasoning: Low-friction use bypasses incumbent distribution and creates qualified accounts. Product telemetry identifies where value concentrates. Enterprise sales converts security, administration, compliance, support, and broader rollout into paid expansion. Pricing and packaging preserve continuity between the user's value and the buyer's economic reason to pay. Selling to Developers, 06:47–16:27 · Growth+Sales, “New buyer, new motion” and Failure Modes #1–#6 · Bottom Up Pricing, sections #1–#9 Assumptions: Users can adopt or influence adoption; product value is experienced before a full centralized purchase; retained activity predicts organizational value; a premium boundary exists; the paid metric scales with customer value; and sales/support cost fits deal size.
• Historical and vehicle context: The doctrine predates and crosses several a16z fund eras. Yubico's 2018 decision sits in the Fund V platform period. Yubico thesis · Fund V, allocation paragraphs Parsec's 2020 Series B closely follows the Fund VII announcement. Parsec thesis · Fund VII, opening Tackle's 2021 rounds occurred in a permissive financing regime, while Fivetran is another portfolio-customer sourcing example. Tackle Series B · Tackle Series C · 2021 Venture Monitor, Executive Summary · Fivetran No public source maps these checks to a vehicle or supplies compatible ownership and return data.
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: retained activation; organic account penetration; users asking for administration/security/enterprise deployment; repeatable buyer identity; paid conversion without engagement collapse; expansion and referenceability; declining implementation per account; sales productivity after full ramp; value-aligned usage rather than bill shock. Do not use it as a mandatory bottom-up template for centralized systems of record, regulated or safety-critical deployments that require prior authorization, or any product with no credible premium/budget bridge.
• How to apply it: (1) Name user, champion, administrator, economic buyer, and procurement authority. (2) Measure independent time-to-value, cohort retention, and within-logo spread. (3) identify the event that creates enterprise value and the feature/control that can be charged for. (4) trace security, procurement, implementation, and contract steps. (5) segment CAC, payback, support, churn, and gross margin by self-serve and sales-assisted motions. (6) Test price and packaging with real buyers. (7) Hire sales because customers request wider deployment, not because free growth needs a revenue story.
• Examples:
  1. Documented application — Yubico. [investor-stated · strong] a16z used YubiKeys before investing, and Casado described broad practitioner references plus profitability/capital efficiency. This is unusually direct product truth, but public records omit conversion cohorts, price, and a16z economics. Yubico announcement, product-use, diligence, and board paragraphs
  2. Documented application — Parsec. [investor-stated · strong] Casado tracked the product from a 2017 demo and cited a loyal gaming community plus enterprise customers in several professional verticals before a16z led the 2020 Series B. The consumer-to-enterprise bridge was part of the public thesis; post-pandemic cohort and margin data were not. Parsec announcement, demo/customer and Series B paragraphs
  3. Documented application and adverse test — Tackle. [investor-stated · strong] A portfolio customer sourced the company, and the decision record cited 200-plus customers, references, growth, and low burn. Later company testimony that it overbuilt and separated 75 employees shows that customer pull did not by itself validate scalable economics. Tackle Series B thesis · Tackle company update, opening and “overbuilt” paragraphs
  4. Documented application — Fivetran. [investor-stated · strong] The announcement says repeated encounters with the product through portfolio companies helped establish demand. That supports portfolio-customer sourcing, not a complete user-to-budget cohort. Fivetran announcement, portfolio encounter and investment paragraphs
• Evidence: The developer-selling discussion establishes the user/buyer split; Growth+Sales and Bottom Up Pricing supply the funnel, failure modes, and packaging procedure; the four announcement records establish only their named sourcing and diligence facts. Selling to Developers, 06:47–16:27 · Growth+Sales · Bottom Up Pricing They do not prove causal selection skill or investor return. Exact evidence gaps: No deal-level funnel dashboards; no cohort retention, conversion, CAC/payback, expansion, or churn for the four cases; no record of rejected high-usage companies; no causal board actions; no matched bottom-up versus top-down outcomes; and no investment ownership/return record.
• Limits, failure modes, and exception ledger: Limits and exceptions: Centralized systems of record, safety-critical deployments, and regulated infrastructure may require top-down authorization before use. High-touch enterprise products can be sound businesses without bottom-up entry. Conversely, Yubico's hardware/security path and Parsec's acquisition do not prove the framework generalizes. The public examples are selected and understate products that achieved use but never monetized. Failure modes: free users without a budget owner; sales masking weak engagement; a paid pilot buying access to expertise rather than product; open-source popularity without a paid control point; services-heavy integration; pricing units disconnected from value; premature field sales; or product neglect after sales succeeds. Falsifiers: no common buyer after repeated deployments; paid cohorts retain no better than free cohorts; within-logo use does not expand; enterprise requirements force customer-specific product forks; gross profit cannot cover the required sales/support motion; or repricing destroys use because consumption was never economic value.
• Evolution and contradictions: The model develops from “developers are an adoption wedge” (2016), through a barbell enterprise-buying model (2019), into explicit funnel and packaging diagnostics (2020–2021), and then into 2025 AI pricing concerns where a unit of usage must include variable cost and outcome quality. Selling to Developers, 06:47–16:27 · Software Chasms, 00:26:48–00:30:39 · Growth+Sales · Bottom Up Pricing · AI pricing, Takeaways The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: trace active user → workflow owner → economic buyer → procurement → price → contribution. Does not transfer: seats, tokens, or raw task counts as value. Changed assumptions: agents add retries, human review, security approval, liability, and variable inference cost. Observable signals: paid expansion, budget transfer, stable completion quality, and declining intervention. Likely misuse: calling daily use a business. Current grounded example: the 2025 AI-pricing discussion treats seat-to-usage change as a live operating problem, not proof that a particular pricing unit works. AI Is Upending SaaS Pricing, 00:10–20:12 and 21:42–31:44

Model 3: Anneal product, buyer, and category before scaling the organization

• Rule: In a new market, product, user, buyer, budget, deployment, price, and category language are all uncertain and coupled. Scaling specialized functions before those variables converge amplifies noise. Founder-led customer work carries the most product context; a small number of generalists can then turn repeated patterns into a playbook.
• Evidence status: [investor-stated · strong] for the founder-led learning sequence and services-drift boundary; [investor-stated · provisional] for published ARR, timing, and staffing ranges; [researcher inference · moderate] for deal and outcome applications.
• Provenance and origin: Nicira supplied the operating lesson: OEM/partnership routes failed, and roughly two years of direct founder selling were required to learn a new category. The 2017 no-market and pricing talks converted that history into a discovery discipline. The 2019 Software Chasms interview separated execution failure from absent/premature product-market fit. The 2023 Market Annealing essay named the model and sequenced founders, a few generalist “renaissance reps,” and later specialization; the 2024 interview limited it explicitly to category creation. Software Chasms, 00:08:10–00:11:36 · Pricing podcast, 07:20–13:51 · Market Annealing, founder-led selling through scaling sequence · Uncharted, 00:42:28–00:45:45
• Mechanism and reasoning: Each live sale tests product and market hypotheses at once. Repeated use cases reduce the search space; reusable implementation and a common buyer shorten cycles; references lower education cost. Specialization becomes productive only after the information being handed off is stable enough to replicate. Software Chasms, 00:08:10–00:11:36 · Market Annealing, founder-led selling through scaling sequence Assumptions: The company is truly creating a category rather than entering a mature one; founders can sell and update product; customer work can converge on reusable functionality; early salespeople report bad news; and financing runway is long enough to learn without making endless exploration costless.
• Historical and vehicle context: Nicira is the founder/operator origin and must remain outside Casado's venture-selection denominator. Software Chasms, 00:08:10–00:13:19 The illustrations span Tecton's 2018 seed and 2020 Series A, Parsec's 2020 Series B, and Tackle's 2021 Series B and preemptive Series C. Tecton, customer and seed/Series A paragraphs · Parsec, Series B/board paragraph · Tackle Series B · Tackle Series C The Tackle follow-on occurred inside the 2021 financing boom. 2021 Venture Monitor, Executive Summary No public record assigns deal-specific vehicles, internal alternatives, reserve gates, or board votes.
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: three or more repeat deployments sharing a use case and buyer; declining customer-specific engineering; shorter and less founder-dependent cycles; repeatable price/value language; references; successful generalist-rep ramp; stable implementation; and fully loaded quota contribution after the playbook exists. Do not use it to excuse indefinite bespoke work, to delay ordinary execution in a mature category with known buyers, or after dated convergence tests repeatedly fail.
• How to apply it: (1) Freeze the unknowns: painful use case, user, buyer, budget, deployment, price, and category language. (2) Set a dated experiment for each. (3) Keep founders in sales/product feedback until repeated evidence emerges. (4) Add a small number of generalists who can create, not merely execute, the playbook. (5) Measure reuse, customization, referenceability, cycle time, implementation, and rep ramp. (6) Add specialized sales/marketing/customer-success capacity only after convergence. (7) Stop or change model if revenue grows while reuse falls.
• Examples:
  1. Documented origin case — Nicira. [retrospective witness · strong] Casado says OEM and partner distribution failed and founder-led direct selling resolved the causal question. This is founder/operator learning, not a venture-selection success. Software Chasms, 00:08:10–00:13:19
  2. Researcher-applied illustration — Tecton. [researcher inference · moderate] The 2020 announcement says a16z invested at seed in 2018 and again at Series A after engagement with ML teams and production feature-store demand. It is consistent with staged evidence release, but no public milestone memo or reserve gate exists. Tecton announcement, customer and seed/Series A paragraphs
  3. Researcher-applied illustration — Parsec. [researcher inference · moderate] Casado's multi-year observation from 2017 to the 2020 Series B fits learning before commitment, but the public record does not show product/buyer convergence measures. The acquisition nine months later leaves the standalone scaling thesis unresolved. Parsec announcement · Unity 10-K, Note 6
  4. Adverse illustration — Tackle. [researcher inference · strong] Strong early customer evidence and a preemptive Series C preceded admitted overbuilding. The case demonstrates why a new round must re-test repeatability, services content, and no-new-capital runway rather than simply extrapolate growth. Acquisition terms remain unknown, so this is an operating miss, not a proven investor loss. Tackle Series C, preemptive/scaling paragraphs · company update
• Evidence: The Nicira interview establishes the founder-led discovery origin; Market Annealing supplies the explicit sequence; the Tecton, Parsec, and Tackle records are labeled applications or adverse illustrations rather than proof of the model. Software Chasms, 00:08:10–00:13:19 · Market Annealing · Tackle update They do not prove causal selection skill or investor return. Exact evidence gaps: No named company cohort behind the 2023 thresholds; no failed-company denominator; no actual board staffing decisions tied to the model; no customer-level longitudinal data; no proof Tecton/Parsec used this process; no Tackle board vote or intervention record; and no outcome/return comparison for companies that scaled before versus after convergence.
• Limits, failure modes, and exception ledger: Limits and exceptions: Bottom-up products can reach repeatability without prolonged founder selling; regulated or complex categories may require domain sellers early; a rapid demand shock can make slow learning costly; and the essay's selected-company observations lack a failure denominator. Nicira is retrospective and success-selected. Tackle shows that early fit can coexist with later organizational error. Failure modes: treating bespoke services as discovery forever; using paid pilots as proof of demand; hiring route sellers for a nonexistent route; scaling marketing before category language works; founder charisma hiding nonrepeatability; survivalist underinvestment after the inflection; or applying category-creation patience to an ordinary market with clear comparables. Falsifiers: no convergence in product/buyer/deployment by the precommitted date; each customer needs a core fork; non-founder reps cannot reproduce wins; sales cycle and implementation do not improve; references do not create follow-on demand; or a mature incumbent category already supplies the buyer and budget, making “annealing” an excuse for weak execution.
• Evolution and contradictions: The early lesson was direct selling to discover demand. By 2019 it became a diagnostic separating execution from PMF. In 2023 it became an organizational sequence. In 2024 Casado added a boundary: use annealing for category creation, not as a universal company-building ritual. Software Chasms, 00:08:10–00:11:36 · Market Annealing · Uncharted, 00:42:28–00:45:45 The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: founder-led learning, convergence dashboards, and dated stop tests. Does not transfer: the essay's historical ARR/staffing ranges as universal gates. Changed assumptions: AI accelerates demos, code, and outbound activity while models and categories change faster. Future of Software Development, Takeaways Observable signals: reusable deployment, buyer recurrence, shorter cycles, declining review burden, and fully loaded rep productivity. Likely misuse: endless exploration disguised as speed. Current hypothetical: an agent startup should not scale field sales until three repeat deployments share workflow, buyer, and implementation; this is an analyst test, not a documented Casado deal.

Model 4: Underwrite the monetizable control point, not the technical layer or open-source adoption alone

• Rule: A technically important layer can be economically weak. Open source or a model API can create adoption while clouds, model suppliers, distributors, or customers capture the value. A venture-scale company needs a defensible point where operational necessity, customer relationship, switching cost, and the right to charge coincide.
• Evidence status: [investor-stated · strong] for the distinction among developer adoption, open-source insertion, operated service, and paid boundary; [researcher inference · strong] for converting architecture and GenAI stack maps into a bargaining-power test; [researcher inference · moderate] for case outcomes.
• Provenance and origin: At Nicira, Casado dealt directly with open/proprietary roadmap conflict and put strong engineering talent on the open project while separating teams under a senior arbiter. The 2016 developer/open-source and microservices discussions framed open source as distribution and technical trust rather than monetization. The 2020 data-stack essay looked for necessary unstable control points; the 2023 GenAI platform essay extended the map to compute, models, and applications; the 2024 interview said durable AI moats would migrate toward workflow, integration, community, distribution, and brand rather than model access alone. Selling to Developers, 11:35–24:58 · 2020 architecture essay · GenAI platform essay · Uncharted, 00:26:00–00:30:31
• Mechanism and reasoning: Technical insertion creates trust and lowers adoption friction. Workflow depth, operation, integration, security, proprietary control, community, distribution, and contracts can then create switching cost and pricing power. Supplier and platform concentration can reverse that power even when the product remains technically central. Selling to Developers, 11:35–24:58 · Data architecture, control-point framework · GenAI platform, stack economics Assumptions: The workload remains independently necessary; customers value the operated/integrated product; the paid boundary does not destroy community trust; contracts permit multi-supplier or migration options; and an incumbent cannot cheaply bundle the function or appropriate demand information.
• Historical and vehicle context: The operating origin is Nicira's open/proprietary product organization, not an investment. Selling to Developers, 20:42–24:58 Isovalent and Tecton are 2020 public thesis records; Cursor is a four-person 2024 a16z thesis and current board claim. Isovalent thesis · Tecton thesis · Cursor thesis, byline and model-choice paragraphs Cisco completed Isovalent's acquisition; Databricks had announced, but not completed at the cutoff, Tecton's transaction; and Cursor later entered an all-stock merger. Cisco completion, opening paragraphs · Databricks announcement, opening · SpaceX 8-K, Item 1.01 These are company events, not fund-return evidence; vehicle, ownership, and proceeds remain unknown.
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: production-critical operation; durable independent budget; broad integrations but low customer migration; user trust converting to paid administration/security/service; falling supplier concentration; contractual portability; expansion around workflow; and survival after adjacent bundling. Do not use it when the layer has no independent budget, remains a replaceable feature, cannot preserve a paid boundary, or is economically controlled by one supplier or bundling platform.
• How to apply it: (1) Draw technical and contractual stack maps. (2) For every layer, identify supplier, user, buyer, budget, operator, data access, distribution, and switching rights. (3) Distinguish free/open insertion from the thing customers pay to operate or control. (4) Test dependence on a dominant cloud/model/distributor and the platform's ability to observe and bundle. (5) Model replacement and portability. (6) Assign milestones for workflow depth, enterprise control, gross margin, and supplier diversification. (7) Kill the deal if the company remains a replaceable feature with no independent budget.
• Examples:
  1. Documented operating application — Nicira's open/proprietary organization. [investor-stated · strong] Casado described separate teams and a senior arbiter to protect the adoption and paid roadmaps. The source omits the project, economics, and comparative outcome. Selling to Developers, 20:42–24:58
  2. Documented thesis — Isovalent. [investor-stated · strong] The a16z post identifies eBPF/Cilium as an open, operational control point in networking, observability, and security; Cisco later completed an acquisition with terms undisclosed. That outcome is consistent with strategic control-point value, not proof of a16z return. Isovalent thesis · Cisco completion
  3. Documented thesis — Tecton. [investor-stated · strong] The post frames feature stores and production ML data as a necessary control point and records seed plus Series A support. Databricks later announced that Tecton would join it, but completion and terms were not public at the cutoff; strategic fit is not realized-return evidence. Tecton announcement · Databricks announcement
  4. Researcher-applied illustration — Cursor. [researcher inference · moderate] The 2024 thesis relied on high-frequency coding workflow and retained use despite externally supplied models. The later all-stock SpaceX transaction is an exceptional company event but does not reveal whether workflow power overcame supplier economics as a standalone business. Cursor thesis · SpaceX 8-K, Item 1.01
• Evidence: The developer/open-source discussion establishes insertion and organizational conflict; the architecture and GenAI essays establish the control-point and bargaining-power maps; the company and acquirer records establish only the named theses and events. Selling to Developers, 11:35–24:58 · Data architecture · GenAI platform · Cisco completion They do not prove causal selection skill or investor return. Exact evidence gaps: No project-level Nicira economics; no public supplier/customer contracts; no switching-test data; no independent-budget cohorts; no Isovalent/Tecton/Cursor entry ownership or proceeds; no comparison against mapped control points that were bundled or failed; and no evidence that Casado's abstract stack model caused those selections.
• Limits, failure modes, and exception ledger: Limits and exceptions: Some commodity layers produce excellent cash flows through scale; some open-source projects monetize through services without venture economics; strategic acquisitions can reward a weak standalone control point; and contractual power can move without any visible architecture change. The portfolio examples are success-selected, while Orbit's acquired-and-sunset product shows that acquisition does not preserve an independent control point. Postman, acquisition and 90-day sunset paragraphs Failure modes: equating GitHub stars with economics; open-source/proprietary roadmap conflict; services-heavy monetization; moving up-stack into fragmented demand; single-model or single-cloud dependence; platform copying; customer multi-homing; or confusing an acquisition by an incumbent with proof of standalone value capture. Falsifiers: the platform bundles the function and customers churn; no paid premium survives; switching tests are easy and common; one supplier can change price/terms enough to erase margin; integration work scales linearly with customers; or the independent budget disappears.
• Evolution and contradictions: The model advances from internal open-source governance to market insertion, then to architecture control points, and finally to a combined technical/contractual bargaining-power view in AI. Casado's 2024 moat migration claim narrows earlier enthusiasm for model-layer novelty. Selling to Developers, 11:35–24:58 · Data architecture · GenAI platform · Uncharted, moat migration at 00:26:00–00:30:31 The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: map technical and contractual control, commercial attach, and failover. Does not transfer: open-source adoption or present model scarcity as automatic capture. Changed assumptions: cloud credits, model-routing rights, exclusivity, revenue share, and platform vertical expansion. Observable signals: independent budget, paid attach, neutral governance, real failover, and durable workflow ownership. Likely misuse: calling GitHub stars or a bottleneck a moat. Current grounded example: Cursor's 2026 transaction is an exceptional workflow-level company event, but it still does not disclose standalone supplier economics or a16z return. SpaceX 8-K, Item 1.01

Model 5: Buy option value under uncertainty; specialize or internalize only when workload economics stabilize

• Rule: Flexible external infrastructure is valuable when demand, product, and hardware needs are uncertain. Once a workload is predictable, the provider's margin, egress, and control can outweigh flexibility. The right answer is not “cloud” or “owned”; it is a staged comparison for each workload after counting migration, idle capacity, staff, capital, facilities, and reliability.
• Evidence status: [investor-stated · strong] for the lifecycle and workload-specific rule; [investor-stated · provisional] for the original aggregate cloud-value estimates; [researcher inference · strong] for the 2026 hybrid decision procedure; [researcher inference · moderate] for portfolio applications.
• Provenance and origin: The 2021 Cost of Cloud essay argued that public cloud has high early option value but can become a material cost of revenue at scale. The 2022 infrastructure essay argued that hyperscaler abstractions also enable specialized vertical infrastructure. The 2022 interview described an unpublished cloud-versus-self-host counterfactual model. The 2023 compute essay expanded choices to hosted, reserved, specialized, and owned capacity, while the 2024 interview added the fiber-glut warning: secular demand can be real and supply investment badly timed. Cost of Cloud, “The paradox” through repatriation decision · Cloud Killed Infrastructure, conditions and conclusion · 2022 interview, 00:06:49–00:19:42 · High Cost of AI Compute · Uncharted, 00:03:33–00:04:43
• Mechanism and reasoning: Renting converts fixed capital and provisioning risk into variable expense and accelerates learning. Reservation/specialization/internalization can capture utilization and hardware advantages after variance falls. Portability and contracts preserve bargaining power while the firm learns. Timing error creates stranded capacity or permanent lock-in. Cost of Cloud, lifecycle and objections · Cloud Killed Infrastructure, specialization conditions · High Cost of AI Compute, decision framework Assumptions: Workloads can be measured independently; demand eventually becomes forecastable; the company can negotiate or migrate; specialized hardware/software creates real unit-cost advantage; reliability and staffing costs are observable; and financing duration matches the payback period.
• Historical and vehicle context: The written rule developed from the 2021 cloud-cost framework through the AI-capacity cycle. Cost of Cloud · High Cost of AI Compute Dropbox is an external public-company example rather than a Casado deal. Dropbox FY2025 10-K, MD&A pp. 57–59 Parsec, Cursor, and Astranis are researcher-applied portfolio illustrations with different workload and capital profiles. Parsec · Cursor · Astranis Their public records do not identify the responsible vehicle, negotiated infrastructure terms, ownership, reserves, or realized proceeds.
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: stable utilization; predictable workload mix; large provider margin; short modeled payback after all costs; successful portability test; multiple suppliers; improving reserved-price capture; high data gravity with long duration; and financing runway through deployment. Do not use it to justify internalization while demand is volatile, when all-in savings vanish after migration/reliability/capital costs, or when payback exceeds the product and financing horizon.
• How to apply it: (1) Segment workloads by latency, variability, data gravity, hardware, security, and sovereignty. (2) Calculate current all-in unit cost, including egress and labor. (3) Model hosted/on-demand, reserved, specialized-cloud, contracted, and owned cases with utilization ranges. (4) Include migration capital, reliability staff, depreciation, facilities, fallback capacity, and opportunity cost. (5) Run supplier-failure and demand-downside cases. (6) Preserve portability until savings exceed the migration hurdle with margin of safety. (7) Re-test before every large capacity commitment.
• Examples:
  1. Documented external case and counterexample — Dropbox. [contemporaneous record · strong] The 2021 essay used Dropbox as a repatriation illustration. Dropbox's FY2025 filing later reported gross margin declining from 83% to 80% amid infrastructure-refresh and facilities costs, falsifying any simple “owned is permanently cheaper” rule. It does not invalidate workload-specific internalization. 2021 essay, Dropbox case · Dropbox FY2025 10-K, MD&A pp. 57–59
  2. Researcher-applied illustration — Parsec. [researcher inference · moderate] Low-latency remote graphics makes egress, GPU utilization, reliability, and cloud concentration decision-critical. The public decision record does not disclose them, and Unity's later cash acquisition cannot validate the missing workload economics. Parsec thesis · Unity 10-K, Note 6
  3. Researcher-applied illustration — Cursor. [researcher inference · moderate] External models and compute gave the 2024 product speed and optionality, while creating supplier and per-task cost exposure. The public record does not establish whether infrastructure strategy later changed or whether workload economics improved. Cursor Series A thesis
  4. Researcher-applied illustration — Astranis. [researcher inference · moderate] Dedicated satellite manufacture, launch, supplier, and debt commitments are the physical extreme of an option-versus-specialization decision. A mission anomaly and replacement response provide operational evidence, not proof that capital was sized correctly. Astranis investment thesis · Astranis anomaly, failure and fix
• Evidence: The cloud-cost, specialization, and AI-compute essays establish the staged workload framework; Dropbox's filing and the claim-local deal records provide examples and counterevidence rather than a matched outcome cohort. Cost of Cloud · Cloud Killed Infrastructure · High Cost of AI Compute · Dropbox FY2025 10-K, MD&A pp. 57–59 They do not prove causal selection skill or investor return. Exact evidence gaps: The unpublished company-level counterfactual model; actual negotiated cloud/compute terms; named workload cost studies; engineering and reliability allocations; migration outcomes; specialist-provider failure data; deal-specific capacity covenants; and any evidence tying the model to venture returns.
• Limits, failure modes, and exception ledger: Limits and exceptions: Stable workloads may still favor cloud because negotiated discounts, reliability, and opportunity cost dominate. Owned systems can suffer refresh cycles, idle capacity, and facilities expense. Hardware-heavy companies cannot always preserve cheap options. The 2021 aggregate estimates and unpublished 2022 counterfactual model are not reproducible; Dropbox is one bundled operational case. Failure modes: ideological repatriation; company-wide decisions instead of workload segmentation; excluding labor/reliability/capital; forecasting peak demand; cloud commitments that destroy optionality; specialist providers with thin balance sheets; or financing a secular-demand thesis at a cyclical overbuild peak. Falsifiers: savings vanish under realistic utilization; migration payback exceeds product/financing horizon; reliability falls; supplier concentration rises; the workload changes faster than dedicated capacity can adapt; or customer/sovereignty constraints prevent the proposed architecture.
• Evolution and contradictions: The argument starts as cloud lifecycle economics, expands to specialized infrastructure opportunity, then to AI accelerator procurement and macro overbuild timing. The more mature version is a portfolio of workload options, not a repatriation slogan. Cost of Cloud · Cloud Killed Infrastructure · High Cost of AI Compute · Uncharted, overbuild at 00:03:33–00:04:43 The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: price flexibility as an option and stage specialization behind measured workload economics. Does not transfer: “cloud bad,” permanent repatriation savings, or average utilization. Changed assumptions: accelerator obsolescence, power/interconnect, sovereignty, model routing, and provider solvency. Observable signals: utilization distributions, portability tests, short all-in payback, supplier alternatives, and funded duration. Likely misuse: financing peak capacity from a secular-demand story. Current grounded example: Astranis' 2026 equity/debt financing funds a physical capacity system but does not prove capital sizing or return. Astranis Series E/debt, opening capital paragraphs

Model 6: Underwrite quality-adjusted contribution margin per completed task

• Rule: Revenue per attempt, seat, or token can obscure whether the product completes economically valuable work. Retries, model cascades, human review, support, implementation, and failures belong in the denominator. Gross margin can improve, but expected deflation is not a substitute for a measurable path.
• Evidence status: [investor-stated · strong] for the AI cost taxonomy and margin-bridge mechanisms; [investor-stated · provisional] for the 2020 private-company percentage ranges; [researcher inference · moderate] for the exact metric and named deal applications because no public portfolio dashboard exists.
• Provenance and origin: The 2020 New Business of AI essay argued that inference, retraining, human review, implementation, and customer-specific edge cases make many AI companies economically unlike classic software. The 2023 economic essay added a task-tolerance test: capability is not useful economics when failure, evaluation, and workflow adaptation remain costly. The 2023 compute essay made procurement part of cost of revenue. The 2025 margin essay argued that low early margins can improve through routing, tiering, repricing, and enterprise mix—but only with an explicit bridge. New Business of AI, gross-margin and scaling sections · Economic Case for GenAI · High Cost of AI Compute · Questioning Margins
• Mechanism and reasoning: Narrower tasks and better evaluation reduce errors and review; routing sends simple work to cheaper models; caching/batching and negotiated compute reduce inference cost; reusable deployment reduces services; pricing/tiering aligns revenue with delivered value. Each change should improve contribution per accepted completed task, not merely reported utilization. New Business of AI, cost taxonomy and scaling · Economic Case for GenAI, task-tolerance framework · Questioning Margins, operating levers Assumptions: Completed-task quality can be defined; failed attempts and review can be measured; customers value the output enough to pay; cost declines are not fully passed through; workflow integration is reusable; and the company retains some pricing or routing control.
• Historical and vehicle context: The 2020 framework came from an unnamed private-company interview sample before foundation-model economics changed the workload; the 2023 and 2025 updates respond to those changes rather than erase the earlier accounting rule. New Business of AI, “Gross Margins” and “Scaling AI systems” · Economic Case for GenAI · Questioning Margins Cursor, Tackle, and Parsec are deal illustrations; Snowflake is an external counterexample. Cursor · Tackle · Parsec · Snowflake FY2025 10-K, MD&A pp. 70–71 None supplies a Casado/a16z cost basis, portfolio margin dashboard, vehicle allocation, or realized return.
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: lower retry and intervention rates; stable accepted-output quality; decreasing compute and support per success; reusable deployment; positive cohort contribution; value-aligned price; better enterprise mix without hidden services; and observed, not assumed, routing savings. Do not use it as a single margin veto where the task inherently requires a coherent services component, or when accepted output cannot be defined; in those cases use a broader customer-value and cash-contribution test.
• How to apply it: (1) Define the accepted output and customer value. (2) Count attempts, retries, latency, review, escalation, support, implementation, inference/training, storage, and third-party fees. (3) Compute revenue and contribution per accepted completed task by use case/customer. (4) stress model price, error rate, and usage mix. (5) Specify a dated margin bridge with owner and milestone for routing, tiering, automation, pricing, or mix. (6) re-underwrite negative-margin customers and customer-specific forks. (7) Stop if scale worsens quality-adjusted economics without a validated repair.
• Examples:
  1. Researcher-applied illustration — Cursor. [researcher inference · moderate] Paid developer pull is strong evidence, but the 2024 freeze omitted inference-adjusted margin, retries, useful-task completion, and supplier concentration. The later company outcome cannot retroactively fill those unit economics. Cursor thesis · SpaceX 8-K, Item 1.01
  2. Adverse illustration — Tackle. [researcher inference · moderate] Marketplace volume and customer count did not disclose software/services mix, take rate, gross margin, implementation labor, or CAC payback. The later overbuilding admission makes those omissions decision-relevant, while completed-acquisition terms leave investor outcome unknown. Tackle thesis · Tackle company update · Tackle issuer status
  3. Researcher-applied illustration — Parsec. [researcher inference · moderate] Low-latency graphics creates an obvious quality/cost denominator: successful session quality after GPU, bandwidth, egress, support, and enterprise security. None of those figures is public in the deal record. Parsec announcement
  4. Grounded external counterexample — Snowflake. [contemporaneous record · strong] Snowflake's FY2025 filing documented gross-margin compression; it is a useful warning against automatic margin expansion, not a matched AI-startup application or a Casado investment outcome. Snowflake FY2025 10-K, MD&A pp. 70–71
• Evidence: The 2020 AI-business essay establishes the cost taxonomy; the 2023 economic and compute work updates task economics; the 2025 margin essay supplies possible repair levers; the examples remain illustrations without a portfolio cohort. New Business of AI · Economic Case for GenAI · High Cost of AI Compute · Questioning Margins They do not prove causal selection skill or investor return. Exact evidence gaps: Names and denominator of the 2020 sample; company-level cost allocation; accepted-task definitions; retry/review/implementation cohorts; actual margin bridges and board milestones; price elasticity; failed AI investments behind Casado's “bloodbath” remark; and matched returns by quality-adjusted margin.
• Limits, failure modes, and exception ledger: Limits and exceptions: Low-margin products can be valuable when the market is enormous, capital is patient, or services create a defensible compound system. High gross margin can hide weak retention or customer value. “Completed task” may be ambiguous in creative/interactive products. The original 50–60%, 25%+, and 10–15% observations come from an unnamed early-company sample and must not be treated as 2026 base rates. Failure modes: counting tokens/tasks as value; excluding review or implementation from cost of revenue; assuming model prices always fall; customer-specific ML backlogs; using revenue growth to hide gross loss; per-seat pricing when agents do the work; outcome pricing without reliable attribution; or preserving usage by subsidizing uneconomic heavy users. Falsifiers: cost per accepted result does not fall with scale; review/escalation remains structurally high; customers reject output or will not pay; price competition captures all cost deflation; deployment/support grows linearly; or margin improvement depends only on an uncontracted future supplier price.
• Evolution and contradictions: Casado's early view emphasized structural AI cost and services drift. The foundation-model update allowed more tasks to become economical. The 2025 view became less categorical about current margin but more demanding about the bridge. The durable element is accounting discipline, not a fixed software-margin threshold. New Business of AI · Economic Case for GenAI · Questioning Margins The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: measure revenue and every variable/semi-variable cost per accepted customer outcome. Does not transfer: fixed classic-SaaS margin thresholds or assumed model-cost deflation. Changed assumptions: agents add retries, tool calls, runtime, human escalation, and warranty/error cost. Observable signals: stable accepted quality, falling intervention and cost per success, reusable deployment, and contracted pricing. Likely misuse: counting tokens or attempts as value. Current grounded example: the 2025 margins essay offers routing/tiering/repricing mechanisms, but no named company bridge proves them. Questioning Margins, “The margin question” through operating levers

Model 7: Treat financing timing, ownership, and downside scenarios as part of the product thesis

• Rule: Venture outcomes depend not only on product truth but also on how much capital is raised, at what price and seniority, what ownership is preserved, which milestones it buys, and whether reserves survive the downside. Too little capital can kill a correct thesis before the market forms; too much or too easy capital can accelerate hiring/capacity ahead of repeatability and reduce future flexibility.
• Evidence status: [investor-stated · strong] for ownership as a binding venture variable, scenario debate, and the danger of both under- and over-capitalization; [researcher inference · moderate] for the integrated model; [researcher inference · insufficient public record] for Casado's actual deal ownership, reserve, and governance practice.
• Provenance and origin: Nicira's 2008 survival at roughly twelve people and later difficulty reversing austerity supplied the two-sided operating lesson. The 2022 venture-model interview described stage-specific teams, scenario planning, and vigorous debate rather than precision forecasting. The 2024 overbuild warning separated secular demand from investment timing. The 2025 monopoly/oligopoly interview made ownership and fund math explicit. The 2026 AI capital-flywheel discussion made financing, compute, capability, demand, and revenue interdependent while warning that deployed GPUs alone do not prove durable demand. Software Chasms, 00:59:12–01:01:27 · 2022 venture-model interview, 14:15–15:33 and 23:00–29:23 · Uncharted, 00:03:33–00:04:43 and 00:50:34–00:52:34 · 2025 ownership/fund math, 53:38–57:09 · 2026 Latent Space capital flywheel, 03:19–11:24
• Mechanism and reasoning: Scenario-based financing connects a tranche to technical, customer, regulatory, or economic uncertainty reduction. Ownership determines whether even a successful company matters to a fund. Reserves provide option value when evidence improves. Downside runway and covenant/structure analysis prevent a financing mark from being confused with value creation. Software Chasms, 00:59:12–01:01:27 · Venture-model interview, 14:15–15:33 and 23:00–29:23 · Ownership/fund math, 53:38–57:09 Assumptions: Milestones can be defined and verified; capital actually retires named risks; the investor can obtain terms commensurate with uncertainty; financing markets may close; follow-on rights are usable; and governance can change spend when evidence disappoints.
• Historical and vehicle context: Nicira supplies operator evidence about survival and reinvestment. Software Chasms, 00:59:12–01:01:27 The 2022 venture-model interview describes stage-specific teams and scenario planning; the 2025 monopoly/oligopoly interview ties ownership to fund math and refers to an approximately $1.2 billion infrastructure fund, versus the announced $1.25 billion Fund IX Infrastructure vehicle. Venture-model interview, 14:15–15:33 and 23:00–29:23 · Ownership/fund math, 53:38–57:09 · Fund IX, allocation paragraph Tackle, Astranis, Tecton, and World Labs supply financing or operating illustrations, but no source maps them to that vehicle or discloses check, security, ownership, reserves, preference stack, or proceeds. Tackle Series C · Astranis financing · Tecton · World Labs financing
• When to use: Use the model when its core uncertainty is decision-relevant and these signals can be observed: milestone delivery per dollar; longer runway without slower learning; stable/improving burn multiple; validated technical/regulatory gates; repeatable customer cohorts; financing alternatives; preserved pro rata/ownership; management willing to cut before liquidity forces it. Do not use it as a rigid ownership or tranching formula when milestones cannot be measured, when strict staging would destroy exploration value, or when the ownership target would force a pass on an exceptional company without improving fund downside.
• How to apply it: (1) Write base, upside, and survival cases using only decision-date information. (2) Define what the round buys and the risk retired per dollar. (3) model post-money ownership, dilution, preference/seniority, pro rata, reserves, time to next evidence, and no-new-capital runway. (4) Tie hiring/capacity to repeatability or technical milestones. (5) Define board-level spend triggers, pauses, and stop criteria. (6) Re-underwrite a preemptive round as a new deal. (7) Keep financing amount, valuation, company event, fund proceeds, and realized distributions separate.
• Examples:
  1. Documented origin tension — Nicira. [retrospective witness · moderate] Casado credits austerity with survival through 2008 and says it later made reinvestment difficult. The case supports two-sided capital timing, not a venture return calculation. Software Chasms, 00:59:12–01:01:27
  2. Adverse application — Tackle. [documented behavior · strong] The issuer described its 2021 Series C as preemptive and planned aggressive scaling; it later admitted overbuilding and separated 75 employees. This is direct operating evidence for re-underwriting easy capital. The completed acquisition has undisclosed terms, so loss, return, and recovery cannot be inferred. Tackle Series C · Tackle update · Tackle issuer status
  3. Researcher-applied illustration — Astranis. [researcher inference · moderate] Satellite manufacture, launch, regulation, suppliers, customers, debt, and replacement missions create explicit capital milestones. Public financing and mission records show activity, not ownership quality or return. Astranis thesis · Astranis anomaly · Astranis Series E/debt
  4. Documented repeat-financing record — Tecton. [investor-stated · strong] The announcement records a16z seed support in 2018 and a Series A after further customer/production evidence. It does not disclose checks, valuations, reserve policy, or whether evidence formally unlocked the second check. Tecton announcement, final two paragraphs
  5. Researcher-applied illustration — World Labs. [researcher inference · provisional] The company later announced $1 billion of new financing. That is capital raised, not valuation, company value, or a16z return, and the public record does not expose capacity milestones or ownership. World Labs financing announcement, opening paragraph
• Evidence: The Nicira account establishes the capital-timing tension; the venture-model and monopoly/oligopoly interviews establish scenario, vehicle, ownership, and fund-math claims; the company records establish only named financing or operating facts. Software Chasms, 00:59:12–01:01:27 · Venture-model interview, 14:15–15:33 and 23:00–29:23 · Ownership/fund math, 53:38–57:09 They do not prove causal selection skill or investor return. Exact evidence gaps: No fund/check/security/price/ownership/reserve ledger for any Casado deal; no IC scenario sheets or votes; no board budgets or stop decisions; no Tackle preference stack or acquisition consideration; no Astranis debt/covenant allocation; no World Labs ownership; and no realized cash-flow series linking capital discipline to fund performance.
• Limits, failure modes, and exception ledger: Limits and exceptions: Some category winners require aggressive preemption before metrics mature; strict milestone tranching can starve exploration; ownership targets may cause an investor to miss exceptional companies; and cutting early can destroy the very option the capital was meant to buy. Nicira and Tackle are two anecdotes in different regimes. The capital-flywheel thesis can become circular when financing demand funds suppliers and masks end demand. Failure modes: treating a large round or valuation as an outcome; financing secular demand at a cyclical peak; preemptive capital without a revised downside case; overbuilding teams or capacity; austerity becoming identity after the inflection; hidden debt/covenants; or good company outcomes producing weak fund returns because ownership/dilution/structure were poor. Falsifiers: the next tranche retires no named risk; burn grows faster than repeatability; no-new-capital runway falls; milestone dates slip without learning; downside proceeds would not clear senior claims; follow-on ownership cannot be defended; or the expected company value does not produce an acceptable fund outcome under realistic dilution.
• Evolution and contradictions: The founder lesson was survival versus timely reinvestment. The investor articulation adds ownership, stage teams, and debate. The AI-era update adds capital-intensity, supplier finance, and overbuild timing. The invariant is to price uncertainty through capital and terms rather than narrate it away. Software Chasms, 00:59:12–01:01:27 · Venture-model interview, 14:15–15:33 and 23:00–29:23 · Ownership/fund math, 53:38–57:09 · Capital flywheel, 03:19–11:24 The live contradictions are preserved by the limits, adverse examples, and falsifiers above; later evidence does not retroactively enter an earlier decision.
• Modern VC translation (as of 2026-08-20): [researcher inference · strong] Transfers: tie each financing to changed evidence, milestones, ownership, structure, reserves, governance, and a no-new-capital case. Does not transfer: capital raised, valuation, or deployed GPUs as product/return evidence. Changed assumptions: concentrated AI suppliers, cloud credits, strategic/circular financing, power commitments, and long private duration. Observable signals: risk retired per dollar, independent customer cash, stable burn multiple, defensible ownership, and milestone delivery. Likely misuse: financing a secular truth at a cyclical peak. Current grounded example: World Labs' 2026 $1B financing is capital raised only; capacity milestones, ownership, and return remain unknown. World Labs, opening funding paragraph

Evidence gaps and candidate models

Evidence contract

Information cutoff: 2026-08-20. This is a reconstruction from the frozen Casado package, not an account of private a16z investment-committee practice. The public record supports a set of testable decision tools; it does not show that any model caused a named investment, establishes no personal hit rate, and exposes no complete check/price/ownership/reserve/return history. A deal post proves only the named author's participation in a public thesis; a board roster proves service, not sourcing or economics. Applications below are therefore called documented applications only when a source connects the rule to the action. Everything else is explicitly a researcher-applied illustration.

Inventory reconciliation and retention rule

The corpus produced fifteen plausible model families across the complete writing, talk, biography, and deal-note inventories. Seven clear the working bar below; eight are excluded or absorbed rather than promoted as standalone models. Retention required: a repeatable procedure, observable signals, a falsifier, at least three distinct grounded applications, and useful support from at least two independent underlying sources. Selected-winner examples establish plausibility, not efficacy. The retained set is not an arbitrary “top seven”; it is every candidate that survived those tests.

Across all seven models, the decisive missing layer is private process and economics: original IC memoranda, dated alternatives and passes, partner votes, check/security/price, ownership and dilution, reserve decisions, board interventions, realized proceeds, and complete opportunity/outcome denominators. Model-specific gaps remain in each model's Evidence component; absence is unknown, never clean or zero.

Excluded or absorbed candidates — attractive, but not standalone evidence-backed models

The exclusion test was applied after reviewing the complete priority writing/talk notes, the 43 announcement records, the current/former board roster, seven decision labs, and the primary adverse/outcome sweep. “No result” is an evidence state, not proof of a clean outcome.

1. Marginal-data-value / “data moat” test — absorbed into Models 4 and 6; standalone support insufficient. The 2019 essay supplies an excellent rule—measure marginal observation value, acquisition/labeling cost, decay, rights, exclusivity, and competitor catch-up—and preserves scarce, regulated, or trust-mediated exceptions. But the frozen record does not connect it to three distinct named investment decisions or outcomes. Eloquent Labs is one domain illustration; robotics deployment data and Cursor context are later hypotheses, not documented applications. Attempted routes: essay, later margin/robotics writings, all announcements, board roster, and outcome sweep. Missing: deal memos, data-rights diligence, longitudinal model-lift cohorts, and passes. Data-moats essay
2. Evidence-and-monitoring governance for open models — excluded from the investment-model set. The 2024 policy essay and NTIA record support a policy procedure—define the artifact, compare marginal risk, identify legal gaps, monitor, and precommit escalation—but not three investment applications, portfolio-governance cases, or conflict-management outcomes. Attempted routes: full policy note, partial open-code essay, announcements, profile, and adverse sweep. Missing: investment decisions changed by evaluation evidence, incident thresholds, lobbying/conflict records, and monitoring actions. Policy essay · NTIA report
3. Physical AI as an industrial system — excluded as too new and financially interested. The 2025 robotics essay offers a useful component→integration→deployment-data chain and can inform Astranis/DeepMap diligence, but it is one current sector thesis with no three-case production/outcome record. World Labs is a financially interested illustration and not manufacturing evidence. Attempted routes: robotics essay, Uncharted board-learning discussion, Astranis/DeepMap/World Labs records, and adverse sweep. Missing: bill-of-materials, yield, field failures, utilization, repair, supplier concentration, data rights, working-capital, and realized outcomes. Robotics essay
4. Bounded autonomy / tail-and-correction-loop test — excluded as a prospective technical thesis. Casado's agent skepticism is direct and falsifiable—long-horizon correctness, state, memory, planning, and correction must improve—but the corpus supplies no three-company decision/outcome set. Cursor coding, autonomous vehicles/DeepMap, and support agents are cross-domain illustrations only; Casado explicitly notes internal dissent. Attempted routes: 2024 hard-limits and innovation talks, 2025 software-development talk, Cursor/DeepMap records, and adverse sweep. Missing: task-horizon cohorts, intervention rates, state-consistency tests, economic reliability, named passes, and investment changes. Uncharted, 00:56:32–00:58:41
5. Technical-founder prior / match leadership to the bottleneck — excluded as an admitted bias, not a validated model. Nicira's founder selling, sales-team retention, open/proprietary team structure, and a16z's later physical-domain specialization show useful behaviors, but only the first company has detailed operating testimony and no comparison demonstrates technical founders outperform commercial/domain founders. Attempted routes: pricing/developer/chasms talks, Uncharted, profile, and all deal-team descriptions. Missing: selection denominator, false positives/negatives, CEO-change cases, role scorecards, board interventions, and outcomes. Uncharted, 00:42:28–00:43:41
6. “One mortal enemy” conflict rule — excluded for lack of cases. The 2025 monopoly/oligopoly interview supplies a memorable direct statement about portfolio conflict, but no public conflict decision, recusal, consent, information barrier, or harmed company is disclosed. Attempted routes: interview, current portfolio, board roster, vehicle analysis, and overlapping category map. Missing: policy text, actual exceptions, company consent, enforcement, and outcome evidence. Monopolies vs. Oligopolies, 57:09–59:35
7. AI capital flywheel / 1996 supercycle / model fragmentation — retained as scenarios inside Models 1, 5, and 7, not standalone models. These are competing forecasts: financing can create compute/capability/demand feedback; genuine demand can coexist with overbuild; model markets may fragment or consolidate. They generate scenarios and falsifiers but do not yet have three independent decision/outcome applications. Attempted routes: 2024 Uncharted; 2025 generalist and monopoly/oligopoly talks; the 2026 capital-flywheel talk; current AI announcements, financings, and adverse sweep. Missing: audited end-demand versus circular financing, workload utilization, supplier/customer contracts, market-share transitions, passes, and returns. Uncharted · Latent Space capital-flywheel discussion
8. Update-through-failure / navigation under uncertainty — absorbed, not retained alone. It is a valuable epistemic stance but is not sufficiently discriminating to select or govern an investment without the concrete map, annealing, cost, or capital procedure above. The record gives Hugging Face, Nicira, and Tackle as different kinds of learning, yet no common ex-ante rule or process-change ledger binds them. Attempted routes: uncertainty talk, labs, biographies, and adverse sweep. Missing: precommitted update triggers, recorded belief changes, decisions reversed, and comparative results.

Critical tensions the reader should keep live

1. Maps versus graphs: architecture reasoning is Casado's comparative advantage, but his clearest process correction says observed use can beat theory. The procedure must specify when the map yields.
2. Adoption versus capture: developer/open-source pull is a sourcing and insertion signal; it can coexist with no buyer, no paid control point, supplier dependence, or services-heavy economics.
3. Discovery versus discipline: annealing justifies patient learning only while product, buyer, and deployment converge. Tackle shows that strong customer truth does not excuse re-underwriting capital and organizational scale.
4. Flexibility versus cost: rented infrastructure preserves option value, while stable scale can justify specialization. Both cloud lock-in and premature internalization are live failure modes.
5. Capability versus completed-work economics: better models and falling nominal prices do not automatically lower cost per accepted outcome once retries, review, support, and integration are counted.
6. Secular truth versus entry timing: AI or infrastructure demand can be enormous while capacity, valuation, or financing is mistimed. A correct technical thesis is not a sufficient investment thesis.
7. Company outcome versus fund outcome: Cursor, Yubico, Parsec, and other events can be economically important company records, but without cost, ownership, dilution, proceeds, and dates they cannot validate these models as return-generating skill.

What the evidence supports

[researcher inference · moderate] The defensible thesis is narrower than “great investor.” From 2016 through 2026, Casado repeatedly translated changes in infrastructure architecture into testable questions about workflow, user, buyer, budget, control point, contribution margin, and financing. The dated writing and spoken records make that method unusually legible; the public deal record supplies several ex-ante illustrations and one clear adverse capital-scaling case. The record does not establish that Casado personally sourced or approved every attributed deal, that the method caused any outcome, or that it produced superior ownership-adjusted or net returns. _profile.md, “A falsifiable account of the edge” · writings.md, “Core written philosophy” · talks.md, “Evolution, contradictions, and present translation”

The formation sequence is real but collaborative. [retrospective witness · moderate] LLNL's 2015 institutional profile records Casado's 1998–2003 systems/security work and post-9/11 network-security assignment. LLNL, career and post-9/11 paragraphs [contemporaneous record · strong] Stanford and the technical papers record SANE/Ethane-era research with coauthors, while VMware records the Nicira corporate transaction. technical-record map, SANE/Ethane entries · VMware 2012 Form 10-K, Note B, “Acquisition of Nicira, Inc.” [researcher inference · moderate] That sequence supports a plausible systems lens, not a causal proof that operator experience created later investing performance.

[researcher inference · strong] The package's most useful learning result is therefore procedural: freeze the market map and alternatives; demand a user-to-buyer-to-budget chain; normalize economics per accepted outcome; tie capital to risk retirement; preserve team, vehicle, company, investor-proceeds, and LP-distribution boundaries; then reveal the outcome. The seven decision labs and seven retained models operationalize those rules without claiming private IC provenance. investments.md, “Ex-ante decision labs” · mental-models.md, inventory reconciliation

Central-claim audit trail

The audit below re-opened the highest-risk claims against the linked underlying record and recalculated the arithmetic. Pass means the bounded claim is supported; it does not upgrade company evidence into personal performance evidence.

What could falsify the thesis

[researcher inference · strong] The favorable account is falsifiable only if future private evidence can separate method from access, team, vehicle, era, and luck. The following tests are decisive rather than rhetorical:

1. Selection test. A complete sourced/pass/loss ledger, frozen at each decision, would weaken the thesis if architecture/customer evidence did not predict selection or ordinary-case outcomes better than stage-, sector-, and vintage-matched a16z opportunities. denominator and continuation routes
2. Behavior test. Original market maps, memos, dissent, votes, and follow-on records would weaken it if the published rules were retrospective explanations, if milestones did not change capital decisions, or if negative evidence was ignored. profile falsifiers
3. Attribution test. Signed role ledgers would weaken a personal-edge claim if sourcing, diligence, approval, governance, and exit work were principally performed by coauthors, other partners, founders, or platform teams. people and influence map
4. Economics test. Security-level cash flows would weaken the thesis if ownership-adjusted, net, and ex-Cursor results were no better than matched funds or deals despite strong company outcomes. non-computability proof
5. Failure test. Board and reserve records would weaken it if Tackle-like overbuilding recurred without precommitted stop rules, revised downside cases, governance response, or a documented process change. failure matrix
6. Conflict test. Company-by-company records would weaken the institutional explanation if competitor lanes, cross-vehicle allocation, portfolio references, related valuation, or policy advocacy were not governed by disclosed consent, recusal, information barriers, and independent checks. allocation and conflict incentives

[researcher inference · strong] The strongest skeptical explanation currently survives: Casado may be an exceptional technical explainer with privileged operator and a16z access, participating in team decisions during favorable infrastructure and AI regimes, while one extreme company event dominates visible value. The public record cannot distinguish that explanation from repeatable personal selection skill. The ex-Cursor test prevents the skeptical case from collapsing into “only one outcome”—Yubico and Parsec remain visible—but neither supplies fund economics. profile, “The strongest case against greatness” · philosophy, decision evidence that would discriminate

Attribution and denominator audit

[researcher inference · strong] The mechanical ledger reconciles, but its unit must remain explicit. It is a public claim-record/company-lineage universe, not a complete portfolio: 43 announcement claims plus 25 board claims produce 68 records; removing 22 company overlaps produces 46 included lineages. Adding two excluded founder/operator candidates produces 70 candidate records across 48 companies. Of the 46 included lineages, 43 are sufficiently reconstructed and three—DeepMap, Netlify, and Pindrop—remain board-only incomplete. Thus 43/46 = 93.5% public-record reconstruction coverage, while career-wide coverage, hit rate, loss rate, and returns are not computable. denominator table · source reconciliation

[researcher inference · strong] The economic vocabulary also reconciles: corporate consideration, listing, private mark, tender/secondary, investor proceeds, and LP distributions are distinct. Yubico is the one visible case in which the issuer disclosed an expected firm-level post-combination stake and lockup; even there the record does not identify entry cost, exact close/dilution path, sales, proceeds, fund mapping, or distributions. An unknown is not zero. Yubico, ownership table and lockup paragraph · glossary

Material deal and team cross-checks

[researcher inference · strong] A byline proves participation in a public thesis; a board roster proves service. Neither proves sole sourcing, sponsorship, diligence, negotiation, IC approval, check allocation, reserves, follow-on authority, exit vote, ownership, or proceeds. The filing record establishes Andreessen/Horowitz legal control, not deal-by-deal votes; Casado's practice-lead title establishes operational prominence, not legal authority or carry. a16z portfolio disclosure, disclosure below table · A16Z ADV, Schedules A/B

Transfer limits for modern VCs

Information cutoff: 2026-08-20. [researcher inference · strong] The transferable object is the question set, not Casado's reputation, network, fund size, or visible marks.

[researcher inference · strong] Important assumptions have changed: AI capital and supplier concentration are exceptional; strategic/circular financing can mimic end demand; model capability and price move quickly; long private duration raises liquidity and valuation-process risk; policy positions can be economically interested. The current package handles those changes with dated transfer tests, the 2026 AI base rate, fund-vehicle cautions, and independent policy controls. It does not show that an emerging manager without a16z's brand, portfolio references, check size, stage teams, or access can reproduce the sourcing advantage. historical environment · modern translation

Unresolved evidence

[researcher inference · strong] Acceptance is for a bounded public-record curriculum, not for a completed economic proof. The highest-value missing records remain: a complete opportunity/pass/loss ledger; historical a16z roster snapshots; original market maps and IC memoranda; partner votes and dissent; company-by-company source/sponsor/diligence/negotiation roles; conflict consent, recusal, and information barriers; security, price, ownership, dilution and reserve schedules; board interventions and stop decisions; tenders/secondaries; seller waterfalls; residual NAV; fees/carry; and dated LP calls and distributions. source-map access boundaries · investment continuation routes

[researcher inference · strong] Three included companies—DeepMap, Netlify, and Pindrop—still lack adequate entry/thesis reconstruction; Illuminics lacks a reconstructed a16z investment record and remains excluded; no original Illuminics–Quova transaction record or complete Nicira cap table was found. Two gated essays remain analyzed only to authenticated visible boundaries. Audio-only talks are bound to exact media/time locators but lack creator transcripts. These are disclosed limitations after mapped legitimate routes, not permission to reconstruct hidden bodies or infer a clean record. writings coverage · talks coverage · source-map archival gaps

Ordinary-claim reopenings

Selection rule: two ordinary claims were chosen from each of the seven pre-REVIEW core documents, alternating mechanical/inventory claims with factual or interpretive claims. This yields 14 reopenings across every root rather than a convenient cluster.

Twenty-point acceptance matrix

Acceptance judgment. [researcher inference · strong] All twenty requirements pass as a public-record learning package after correcting the Yubico firm-ownership/lockup boundary. The curriculum supports study of a falsifiable method and its limits; it does not support a personal return, alpha, sole-credit, or clean-conflict conclusion. Any later completion bookkeeping should preserve that distinction.

Source-map boundary and retrieval priorities

This source map treats Casado as a researcher, founder/operator, venture investor, board member, nonprofit cofounder, and policy advocate. It separates Casado's direct voice from coauthored work, team research, a16z firm claims, company outcomes, regulatory records, and independent criticism. Stars rank evidentiary priority, not truth. Every current roster, private valuation, run-rate figure, acquisition amount, and policy claim retains its own measurement and attribution boundary.

Discovery archives and current institutional identity

• ★★★ | Andreessen Horowitz — Martin Casado author and team archive | https://a16z.com/author/martin-casado/ | live; accessed 2026-08-20 | official mutable profile and content index | Current role, 17 current/eight former boards, biography, writings, and spoken-corpus discovery route; not proof of complete deal attribution or economics. (reliability: A for current firm record; access: full; accessed: 2026-08-20; locator: “About,” current/former board paragraphs, chronological content index)
• ★★★ | Martin Casado — Stanford personal archive | https://yuba.stanford.edu/~casado/ | live/undated; accessed 2026-08-19 | first-person academic homepage | Education, affiliations, projects, teaching, publications, awards, and archived research routes; mutable and partly stale. (reliability: A/B; access: full; accessed: 2026-08-19; locator: “Selected Projects and Research,” “Teaching,” “Education”)
• ★★★ | Martin Casado — Stanford publication index | https://yuba.stanford.edu/~casado/publications.html | live/undated; accessed 2026-08-19 | first-person publication bibliography | Primary discovery denominator for the technical corpus; individual author lists and dates still control. (reliability: A/B; access: full; accessed: 2026-08-19; locator: chronological publication list)
• ★★ | Open vSwitch — research-paper index | https://www.openvswitch.org/support/papers/ | live; accessed 2026-08-19 | project bibliography | Routes to primary technical papers and preserves team authorship; not a Casado-only archive. (reliability: A; access: full; accessed: 2026-08-19; locator: paper list)

Biography, education, operating chronology, public service, and awards

• ★★★ | Lawrence Livermore National Laboratory — The accidental entrepreneur | https://www.llnl.gov/article/41011/accidental-entrepreneur | 2015-04-28 | official direct-interview profile | Richest record of family moves, LLNL internship/employment, simulation work, post-9/11 security assignment, Stanford transition, and later origin narrative; retrospective and celebratory. (reliability: B; access: full; accessed: 2026-08-19; locator: opening biography and LLNL/post-9/11 passages)
• ★ | LLNL — Martin Casado Entrepreneurs' Hall of Fame profile, retired route | https://ipo.llnl.gov/success/hall-of-fame/martin-casado | 2015/current archive record | institutional biography lead | The canonical route returned HTTP 404 on 2026-08-19; no claim extraction relies on it, and the two working LLNL pages carry the chronology. (reliability: A/B for institutional provenance; access: inaccessible/404; accessed: 2026-08-19; locator: retired canonical route, no live body)
• ★★ | LLNL — Four researchers to be inducted into Entrepreneurs' Hall of Fame | https://www.llnl.gov/article/40721/four-researchers-be-inducted-lawrence-livermore-entrepreneurs-hall-fame | 2015-01-15 | official award announcement | Confirms induction date, LLNL tenure, and institutional framing; use primary papers for invention credit. (reliability: A/B; access: full; accessed: 2026-08-19; locator: Casado paragraphs)
• ★★★ | Martin Casado — Stanford PhD dissertation, Architectural Support for Security Management in Enterprise Networks | https://yuba.stanford.edu/~nickm/papers/martin-casado-thesis.pdf | 2007-12 | primary academic PDF | Decision-date account of SANE/Ethane, centralized control, advisers, and research problem; decisive for technical chronology. (reliability: A; access: full; accessed: 2026-08-19; locator: title/approval pp. 0–2, abstract pp. 4–5)
• ★★★ | Stanford EE204 — Nicira Networks company case | https://web.stanford.edu/class/ee204/2012/Nicira%20Networks.pdf | 2012 | university teaching case PDF | Near-contemporaneous management, founder, education, board, and financing scaffold; not an audited filing. (reliability: B; access: full; accessed: 2026-08-19; locator: Exhibit 7, pp. 29–31)
• ★★ | Northern Arizona University — honorary doctorate recipients | https://in.nau.edu/commencement/new-page/honorary-doctorate-recipients/ | live; accessed 2026-08-19 | institutional roster | Confirms 2016–17 honorary doctorate; supplies no operating or investment evidence. (reliability: A; access: full; accessed: 2026-08-19; locator: 2016–2017 roster)
• ★★ | Northern Arizona University — commencement program | https://archive.library.nau.edu/digital/api/collection/nauarch/id/46661/download | 2017 | institutional PDF | Retrospective biography and Illuminics/Quova lead; verify transaction mechanics elsewhere. (reliability: B; access: full; accessed: 2026-08-19; locator: p. 6)
• ★★★ | Martin Casado — And So… | https://blogs.vmware.com/networkvirtualization/2016/02/martin-casado.html/ | 2016-02-24 | first-person contemporaneous transition post | Exact April 1 a16z start, VMware adviser transition, and succession framing. (reliability: A; access: full; accessed: 2026-08-19; locator: closing transition paragraphs)
• ★★★ | Marc Andreessen — Martin Casado joins a16z | https://a16z.com/martin-casado/ | 2016-02-24 | official sponsor/firm announcement | a16z appointment, 2009 Nicira relationship, named people, and firm-reported NSX growth; retrospective sponsor account, not independent financing or audited revenue. (reliability: A/B; access: full; accessed: 2026-08-19; locator: opening, investment, and VMware paragraphs)
• ★★★ | U.S. House Science Committee — Cyber Security: What the Federal Government Can Learn from the Private Sector | https://science.house.gov/2016/1/research-and-technology-subcommittee-and-oversight-subcommittee-hearing-cyber | 2016-01-08 | congressional hearing record | Establishes pre-a16z public testimony, witness role, and contemporaneous VMware title; link through to Casado's written statement. (reliability: A; access: full; accessed: 2026-08-19; locator: witness list and statement link)
• ★★ | Martin Casado — House cybersecurity written statement | https://science.house.gov/sites/republicans.science.house.gov/files/documents/HHRG-114-SY15-WState-MCasado-20160108.pdf | 2016-01-08 | first-person congressional statement PDF | Direct policy and operating record before venture tenure; distinguish recommendations from measured outcomes. (reliability: A; access: full; accessed: 2026-08-19; locator: pp. 1–8)
• ★★★ | MuralNet — About | https://www.muralnet.org/about/ | live; accessed 2026-08-19 | official nonprofit profile | Records 2017 cofounding with Brian Shih and current Casado/Mariel Triggs roles; mutable and insufficient alone for impact totals. (reliability: A/B; access: full; accessed: 2026-08-19; locator: history and team sections)
• ★★★ | Mariel Triggs — House testimony on tribal broadband | https://docs.house.gov/meetings/IF/IF16/20190716/109797/HHRG-116-IF16-Wstate-TriggsM-20190716.pdf | 2019-07-16 | primary congressional testimony PDF | Gives February 2017 MuralNet founding, Casado/Shih roles, motivation, and operating model through the CEO's account. (reliability: A; access: full; accessed: 2026-08-19; locator: pp. 1–3)
• ★★ | Stanford Magazine — Tackling the Tribal Digital Divide | https://stanfordmag.org/contents/tackling-the-tribal-digital-divide | 2020-03 | reported university feature | Named-participant history of MuralNet and Flagstaff formation; conflicts with stronger sources on the MS year. (reliability: B; access: full; accessed: 2026-08-19; locator: biography and MuralNet operating paragraphs)
• ★★ | Multiplying Good — 49th Annual Jefferson Awards | https://www.globenewswire.com/news-release/2021/10/01/2307177/0/en/JOAN-GANZ-COONEY-DR-ANTHONY-FAUCI-AND-JUST-GOODS-AMONG-THIS-YEAR-S-JEFFERSON-AWARD-RECIPIENTS.html | 2021-10-01 | award-giver release | Confirms Casado and Triggs received the 2021 Onassis public-service award; does not independently audit impact. (reliability: A/B; access: full; accessed: 2026-08-19; locator: recipient paragraphs)
• ★★★ | U.S. Senate — eighth bipartisan AI Insight Forum | https://www.schumer.senate.gov/newsroom/press-releases/statements-from-the-eighth-bipartisan-senate-forum-on-artificial-intelligence | 2023-12-06 | Senate event record | Establishes participant roster and routes to Casado's statement; policy participation is not neutral authority. (reliability: A; access: full; accessed: 2026-08-19; locator: Casado entry and statement link)
• ★★★ | Martin Casado — Senate AI Insight Forum statement | https://www.schumer.senate.gov/imo/media/doc/Martin%20Casado%20-%20Statement.pdf | 2023-12-06 | first-person policy PDF | Direct dated position on AI risk, open source, cybersecurity, data, and regulatory capture; disclose AI-investor interest. (reliability: A; access: full; accessed: 2026-08-19; locator: pp. 1–4)
• ★★ | ACM — 2012 technical awards | https://awards.acm.org/binaries/content/assets/press_releases/tech_awards_2012.pdf | 2013 award cycle | award-giver PDF | Officially labels the Grace Murray Hopper Award as 2012 and records then-current roles. (reliability: A; access: full; accessed: 2026-08-19; locator: Casado award paragraph)
• ★★ | NEC C&C Foundation — 2015 C&C Prize | https://sg.nec.com/en_AP/press/201510/ap_20151016_02.html | 2015-10-16 | award-giver announcement | Confirms 2015 team prize with McKeown and Shenker, contradicting current personal/firm 2014 labels. (reliability: A; access: full; accessed: 2026-08-19; locator: Group B recipient section)
• ★★ | Stanford SystemX — SDN and OpenFlow originators receive NEC C&C award | https://systemx.stanford.edu/news/sdn-and-openflow-originators-receive-nec-cc-award | 2015 | institutional award announcement | Corroborates the 2015 award and team attribution; not a source for sole invention. (reliability: A/B; access: full; accessed: 2026-08-19; locator: award and recipient paragraphs)
• ★★ | NEC C&C Foundation — recipient roster | https://www.candc.or.jp/kensyo/recipient_cc.html | live; accessed 2026-08-19 | award-giver index | Independent official year check placing Casado, McKeown, and Shenker in the 2015 roster. (reliability: A; access: full; accessed: 2026-08-19; locator: 2015 Group B entry)
• ★★ | USENIX — Test of Time Awards | https://www.usenix.org/conferences/test-of-time-awards | live through 2026 | award-giver index | Confirms 2025 Open vSwitch and 2026 network-virtualization team awards; durable technical influence is not venture performance. (reliability: A; access: full; accessed: 2026-08-19; locator: 2025 and 2026 NSDI entries)

Primary technical record and independent intellectual history

• ★★★ | Casado et al. — SANE: A Protection Architecture for Enterprise Networks | https://yuba.stanford.edu/~casado/sane/paper.html | 2006 | primary peer-reviewed paper | Earliest priority architecture in the direct technical sequence; preserve full author team and enterprise-security problem. (reliability: A; access: full; accessed: 2026-08-19; locator: abstract and author list)
• ★★★ | Casado et al. — Ethane: Taking Control of the Enterprise | https://yuba.stanford.edu/~casado/ethane_ton_08.pdf | 2007/2009 | primary peer-reviewed PDF | Central logical-control architecture, deployment evidence, and contemporaneous author biography including LLNL, Illuminics, and Nicira. (reliability: A; access: full; accessed: 2026-08-19; locator: abstract; author biography p. 13)
• ★★★ | McKeown et al. — OpenFlow: Enabling Innovation in Campus Networks | https://web.stanford.edu/class/archive/cs/cs244/cs244.1156/papers/openflow.pdf | 2008-04 | primary peer-reviewed PDF | Canonical OpenFlow artifact and essential authorship boundary: Casado is not in the author list. (reliability: A; access: full; accessed: 2026-08-19; locator: title/authors and abstract, pp. 1–2)
• ★★★ | Gude et al. — NOX: Towards an Operating System for Networks | https://yuba.stanford.edu/~casado/nox-ccr-final.pdf | 2008-07 | primary peer-reviewed PDF | Documents Casado's coauthorship inside a broad Stanford/Nicira team and the network-operating-system concept. (reliability: A; access: full; accessed: 2026-08-19; locator: title, author block, abstract)
• ★★★ | Koponen et al. — Onix: A Distributed Control Platform for Large-scale Production Networks | https://www.usenix.org/events/osdi10/tech/full_papers/Koponen.pdf | 2010-10-06 | primary conference PDF | Production-scale control-platform evidence with Nicira, Google, NEC, ICSI, and Berkeley collaborators; no individual engineering allocation. (reliability: A; access: full; accessed: 2026-08-19; locator: title, authors, abstract)
• ★★ | Pettit et al. — Virtual Switching in an Era of Advanced Edges | https://www.openvswitch.org/support/papers/dccaves2010.pdf | 2010 | primary technical PDF | Places Casado in the Nicira/Open vSwitch technical team before acquisition. (reliability: A; access: full; accessed: 2026-08-19; locator: title and author block, p. 1)
• ★★★ | Pfaff et al. — The Design and Implementation of Open vSwitch | https://www.usenix.org/system/files/conference/nsdi15/nsdi15-paper-pfaff.pdf | 2015 | primary NSDI paper PDF | Mature implementation record and later test-of-time source; team outcome rather than proof of Casado primacy. (reliability: A; access: full; accessed: 2026-08-19; locator: title, authors, abstract)
• ★★★ | Koponen et al. — Network Virtualization in Multi-tenant Datacenters | https://www.usenix.org/conference/nsdi14/technical-sessions/presentation/koponen | 2014-04 | primary NSDI paper, slides, video, and audio | Full 25-author production-network-virtualization record and later test-of-time source; collective authorship remains explicit. (reliability: A; access: full; accessed: 2026-08-19; locator: proceedings pp. 203–216; canonical session page)
• ★★★ | Martín Casado, Nate Foster, and Arjun Guha — Abstractions for Software-Defined Networks | https://doi.org/10.1145/2661061.2661063 | 2014-09 online; 2014-10 issue | CACM journal article | Conceptual synthesis of SDN abstractions; canonical DOI may be gated, while the full author PDF is public. (reliability: A; access: full via author PDF/partial canonical; accessed: 2026-08-19; locator: CACM 57(10), pp. 86–95)
• ★★★ | Martín Casado, Nick McKeown, and Scott Shenker — From Ethane to SDN and Beyond | https://web.stanford.edu/class/cs244/papers/ethane_to_sdn_and_beyond.pdf | 2019-10 | ACM SIGCOMM CCR editorial note PDF | Three-author retrospective on the technical lineage; explicitly non-peer-reviewed and unsuitable as sole decision-date evidence. (reliability: A for author recollection; access: full; accessed: 2026-08-19; locator: CCR 49(5), four-page PDF)
• ★★★ | Patil et al. — GoEX: Perspectives and Designs Towards a Runtime for Autonomous LLM Applications | https://arxiv.org/abs/2404.06921 | 2024-04-10 | arXiv preprint | Ten-author agent-runtime paper on delegation, undo, and damage confinement; primary research record, not investment-performance evidence. (reliability: A; access: full abstract/PDF/source; accessed: 2026-08-19; locator: v1 title/authors, abstract, system sections)
• ★★★ | Yang et al. — Cloud Infrastructure Management in the Age of AI Agents | https://arxiv.org/abs/2506.12270 | 2025-06-13 preprint; 2025-08 journal | ACM SIGOPS OSR article/preprint | Nine-author current technical work on AI-agent cloud management; useful for doctrine evolution while preserving research-team credit. (reliability: A; access: full; accessed: 2026-08-19; locator: eight-page article; OSR 59(1))
• ★★★ | Feamster, Rexford, and Zegura — The Road to SDN: An Intellectual History of Programmable Networks | https://www.cs.princeton.edu/courses/archive/fall13/cos597E/papers/sdnhistory.pdf | 2014-04 | independent peer-reviewed history PDF | Strongest check on lone-inventor narratives and map of active networking, control/data-plane separation, OpenFlow, operating systems, and virtualization. (reliability: B; access: full; accessed: 2026-08-19; locator: historical synthesis and conclusion, pp. 9–10)
• ★★ | Stanford Engineering — Open networking summit to explore SDN | https://engineering.stanford.edu/news/open-networking-summit-explore-software-defined-networking | 2011 | near-contemporaneous institutional history | NSF/Clean Slate, Stanford/Berkeley collaboration, OpenFlow, GENI, and industry-adoption context. (reliability: B; access: full; accessed: 2026-08-19; locator: research-origin paragraphs)
• ★★ | ICSI — A Revolution with ICSI Roots | https://icsi.berkeley.edu/stories/a-revolution-with-icsi-roots/ | retrospective institutional history | named-collaborator account of Nicira formation and the shift from vendor indifference toward network virtualization; requires contemporaneous control. (reliability: B; access: full; accessed: 2026-08-19; locator: Nicira formation and research-to-company passages)
• ★★ | ACM SIGCOMM — Test of Time Paper Award roster | https://www.sigcomm.org/awards/acm-sigcomm-test-of-time-paper-award | live | award-giver index | Confirms Ethane's 2017 recognition; use only for award fact and collective technical longevity. (reliability: A; access: full; accessed: 2026-08-19; locator: 2017 entry)
• ★★ | ACM SIGOPS — Hall of Fame history | https://sigops.hosting.acm.org/about/history/ | live | award-giver history | Confirms Onix's 2021 Hall of Fame entry; not venture-performance evidence. (reliability: A; access: full; accessed: 2026-08-19; locator: 2021 entries)

Nicira, VMware, financing, transaction, and operating records

• ★★★ | Virtualization Review — Nicira interview with Martin Casado | https://virtualizationreview.com/articles/2012/07/09/interview-nicira-martin-casado-nicira-network-virtualization.aspx | 2012-07-09 | contemporaneous founder interview | Direct pre-exit account of product, customers, competitors, pricing, and adoption; valuable decision-date record. (reliability: A/B; access: full; accessed: 2026-08-19; locator: background, customers, and competitors questions)
• ★★★ | Fortune — Nicira wants to reinvent routing | https://fortune.com/2012/05/01/nicira-wants-to-reinvent-routing/ | 2012-05-01 | contemporaneous reported profile | Financing, founders, customers, competitive risks, and biographical claims before the sale; contains Cartagena birthplace conflict. (reliability: B; access: full; accessed: 2026-08-19; locator: financing, customers, risks, and biography paragraphs)
• ★★★ | VMware Form 8-K — Nicira acquisition completion | https://www.sec.gov/Archives/edgar/data/1124610/000112461012000026/form8-k.htm | 2012-08-24 | SEC-filed transaction record | Close date, wholly owned status, approximately $1.095B cash, $168M assumed unvested awards, and escrow; do not equate with personal proceeds. (reliability: A; access: full; accessed: 2026-08-19; locator: Item 2.01)
• ★★★ | VMware Form 10-Q — Nicira acquisition accounting | https://www.sec.gov/Archives/edgar/data/1124610/000112461012000030/vmw-9302012x10q.htm | 2012 | SEC-filed financial record | $1.0996B consideration net of acquired cash and the pre-/post-combination equity distinction; accounting measure differs from headline price. (reliability: A; access: full; accessed: 2026-08-19; locator: Note B, “Acquisition of Nicira, Inc.”)
• ★★★ | VMware Form 10-K — Nicira acquisition note | https://www.sec.gov/Archives/edgar/data/1124610/000112461013000009/vmw-12312012x10k.htm | 2013 | audited annual filing | Final annual acquisition accounting and business rationale; no founder-level ownership or proceeds. (reliability: A; access: full; accessed: 2026-08-19; locator: Note B, “Acquisition of Nicira, Inc.”)
• ★★ | VMware — Introducing VMware NSX | https://blogs.vmware.com/networkvirtualization/2013/08/vmware-nsx.html/ | 2013-08-26 | official product announcement | Shows the platform combined Nicira and VMware technologies; prevents attribution of the entire outcome to one founder. (reliability: A; access: full; accessed: 2026-08-19; locator: opening and architecture passages)
• ★★ | Network World — VMware's Casado talks about evolving SDN use cases | https://www.networkworld.com/article/930386/vmwares-casado-talks-about-evolving-sdn-use-cases-including-a-prominent-role-for-security.html | 2014-10-06 | operating interview | Direct role, security thesis, and $100M run-rate claim; annualized metric needs audited-company context. (reliability: A/B; access: full; accessed: 2026-08-19; locator: opening role and run-rate questions)
• ★★ | VMware — Rajiv Ramaswami appointed to lead Networking and Security | https://www.globenewswire.com/news-release/2016/02/24/1325965/0/en/VMware-Appoints-Former-Broadcom-and-Cisco-Executive-Dr-Rajiv-Ramaswami-to-Lead-Networking-and-Security-Business.html | 2016-02-24 | official company release | Succession and role boundary as Casado moved to a16z; company-authored. (reliability: A; access: full; accessed: 2026-08-19; locator: appointment and Casado-transition paragraphs)
• ★★★ | Martin Casado — Bringing in an External CEO | https://a16z.com/bringing-in-an-external-ceo/ | 2020-03-06 | first-person retrospective essay | Direct account of Nicira's three-CEO sequence, founder authority, and formal-title choices; demands contemporaneous governance checks. (reliability: A; access: full; accessed: 2026-08-19; locator: opening and three-CEO/Steve Mullaney passages)

Independent market, policy, conflict, and skeptical context

• ★★★ | FTC — staff report on AI partnerships and investments | https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study | 2025-01 | regulator report release | Independent concentration, cloud-partnership, lock-in, access, and governance questions for an AI-infrastructure investor's claims. (reliability: A; access: full; accessed: 2026-08-19; locator: findings and report link)
• ★★★ | FTC Office of Technology — Partnerships Between Cloud Service Providers and AI Developers | https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted_0.pdf | 2025-01 | regulator staff report PDF | Primary 6(b)-study counterevidence on cloud commitments, equity and revenue-sharing, control and exclusivity, information transfer, compute access, and switching costs; the report disclaims a complete legal or economic market analysis. (reliability: A; access: full; accessed: 2026-08-19; locator: executive summary pp. 1–3; §§4–5, printed pp. 16–37)
• ★★★ | NTIA — Dual-Use Foundation Models with Widely Available Model Weights | https://www.ntia.gov/sites/default/files/publications/ntia-ai-open-model-report.pdf | 2024-07 | U.S. Commerce Department report PDF | Independent policy baseline recognizing competition, research, privacy, safety, misuse, and accountability tradeoffs; found insufficient evidence for immediate restrictions while recommending monitoring, audits, transparency, and future action if risks change. (reliability: A; access: full; accessed: 2026-08-19; locator: executive summary and policy recommendations)
• ★★★ | Snowflake — FY2025 Form 10-K | https://www.sec.gov/Archives/edgar/data/1640147/000164014725000052/snow-20250131.htm | filed 2025-03-26 | SEC-filed issuer financial record | Current test of compute and margin claims: product gross margin fell from 74% to 71%, third-party cloud infrastructure cost rose, and the filing identifies workload mix, pricing, new products, AI investment, efficiency, and stock compensation as drivers. (reliability: A; access: full; accessed: 2026-08-19; locator: MD&A, printed pp. 70–71)
• ★★★ | Dropbox — FY2025 Form 10-K | https://www.sec.gov/Archives/edgar/data/1467623/000146762326000008/dbx-20251231.htm | filed 2026-02-20 | SEC-filed issuer financial record | Current boundary on the 2021 repatriation case: Dropbox still operates owned/colocated infrastructure, while 2025 gross margin fell from 83% to 80% as infrastructure refresh and facilities costs rose; architecture changes cost levels but does not freeze them. (reliability: A; access: full; accessed: 2026-08-19; locator: MD&A, printed pp. 57–59)
• ★★★ | California Governor — SB 1047 veto message | https://www.gov.ca.gov/wp-content/uploads/2024/09/SB-1047-Veto-Message.pdf | 2024-09-29 | official executive-action PDF | Primary counterevidence to a simple regulation/no-regulation frame: rejects a compute proxy while calling for empirical analysis, proactive safeguards, and risk-sensitive guardrails. (reliability: A; access: full; accessed: 2026-08-19; locator: pp. 1–2)
• ★★★ | International AI Safety Report — interim report | https://www.gov.uk/government/publications/international-scientific-report-on-the-safety-of-advanced-ai/international-scientific-report-on-the-safety-of-advanced-ai-interim-report | 2024 | international scientific synthesis | Independent boundary on the “safe” claim: records present harms, disputed severe risks, evidence gaps, and limited assurance methods without collapsing uncertainty into either catastrophe or no-risk advocacy. (reliability: B; access: full; accessed: 2026-08-19; locator: executive summary, malicious-use sections, and §5)
• ★★ | A Path for Science- and Evidence-based AI Policy — expert statement | https://understanding-ai-safety.org/ | 2024 | multi-signatory policy statement cited by Casado | Supports marginal-risk study while calling for research across discrimination, scams, misinformation, sexual-abuse material, cybersecurity, environment, biosecurity, and extreme risks; endorsement is not outcome evidence. (reliability: B; access: full; accessed: 2026-08-19; locator: recommendation and risk-domain list)
• ★★ | a16z — Katherine Boyle profile | https://a16z.com/author/katherine-boyle/ | live; accessed 2026-08-19 | official mutable contributor profile | Establishes Boyle's a16z role and American Dynamism affiliation for the joint WSJ policy-essay interest boundary; not evidence of the gated article body. (reliability: A for current firm identity; access: full; accessed: 2026-08-19; locator: profile description)
• ★★ | a16z — Emerging Architectures for Modern Data Infrastructure, updated landing page | https://a16z.com/emerging-architectures-for-modern-data-infrastructure/ | live update route; accessed 2026-08-19 | official version-navigation page | Points to the original 2020 article and later architecture updates; used only to preserve version boundaries, not as evidence known in October 2020. (reliability: A for version record; access: full; accessed: 2026-08-19; locator: updated-reference-architecture notice and original-version link)
• ★★★ | California Legislature — SB 1047 status and bill history | https://leginfo.legislature.ca.gov/faces/billStatusClient.xhtml?bill_id=202320240SB1047 | 2024 | enacted-process record | Primary legislative baseline for Casado/a16z AI-safety advocacy; bill text, votes, veto, and commentary must remain distinct. (reliability: A; access: full; accessed: 2026-08-19; locator: status, votes, and bill-text tabs)
• ★★★ | Washington Post — AI spending and revenue-gap analysis | https://www.washingtonpost.com/technology/2024/07/24/ai-bubble-big-tech-stocks-goldman-sachs/ | 2024-07-24 | reported market analysis | Capital-intensity and monetization counterweight to falling-cost and expanding-demand narratives; access may vary. (reliability: B; access: full on access date; accessed: 2026-08-19; locator: spending, revenue-gap, and analyst-argument paragraphs)
• ★★ | Fortune — Base AI Policy on Evidence, Not Existential Angst | https://fortune.com/2024/12/14/ai-regulation-policy/ | 2024-12-14 | first-person guest commentary | Original publication route for Casado's policy argument; compare with the a16z republication and preserve access/version boundaries. (reliability: A for authored view; access: full at mapping date; accessed: 2026-08-19; locator: full commentary paragraphs)
• ★★ | TechCrunch — a16z lobbying spending | https://techcrunch.com/2025/08/25/a16z-spends-1-49m-in-washington-lobbying-while-rivals-mostly-sit-out/ | 2025-08-25 | reported policy/conflict record | Quantifies firm policy engagement and supplies economic-interest context; does not establish Casado's individual direction or benefit. (reliability: B; access: full; accessed: 2026-08-19; locator: spending and firm-policy paragraphs)
• ★★ | PitchBook-NVCA — Venture Monitor hub | https://nvca.org/pitchbook-nvca-venture-monitor/ | current archive; accessed 2026-08-19 | industry dataset/report hub | Routes to vintage-correct financing, fundraising, concentration, and exit context; methodology and later revision must be disclosed. (reliability: B; access: full; accessed: 2026-08-19; locator: quarterly report archive)
• ★★ | Cambridge Associates — US PE/VC Benchmark Commentary, Calendar Year 2024 | https://www.cambridgeassociates.com/insight/us-pe-vc-benchmark-commentary-calendar-year-2024/ | 2025 | benchmark commentary | Net-return and vintage comparison boundary; private marks, company exit values, and LP performance remain different measures. (reliability: B; access: full; accessed: 2026-08-19; locator: methodology and US VC results sections)

Priority direct writings and presentations

• ★★★ | Martin Casado and Matt Bornstein — The New Business of AI (and How It's Different From Traditional Software) | https://a16z.com/the-new-business-of-ai-and-how-its-different-from-traditional-software/ | 2020-02-16 | coauthored essay | Early falsifiable account of AI cost structures, services, defensibility, data, and margins; preserve joint authorship and publication-date information set. (reliability: A; access: full; accessed: 2026-08-19; locator: section headings and charts)
• ★★★ | Sarah Wang and Martin Casado — The Cost of Cloud, a Trillion Dollar Paradox | https://a16z.com/the-cost-of-cloud-a-trillion-dollar-paradox/ | 2021-05-27 | coauthored essay | Central cloud-economics and repatriation argument with public-company examples; forecast and accounting assumptions require dated outcome checks. (reliability: A; access: full; accessed: 2026-08-19; locator: cost/paradox framework and case tables)
• ★★★ | Martin Casado — The Cloud Killed Infrastructure. Long Live Infrastructure! | https://a16z.com/the-cloud-killed-infrastructure-long-live-infrastructure/ | 2022-06-09 | first-person essay | Investment-era thesis on cloud abstraction, developer control, and a renewed infrastructure layer; connect to deals without assuming performance. (reliability: A; access: full; accessed: 2026-08-19; locator: section headings)
• ★★★ | Martin Casado and Peter Levine — Market Annealing: Getting to $10M ARR in Very Early Markets | https://a16z.com/market-annealing-getting-to-10m-arr-in-very-early-markets/ | 2023-01-11 | coauthored operating essay | Explicit early-market creation mechanism and $10M-ARR milestone; later work must separate advice, observed examples, and causal efficacy. (reliability: A; access: full; accessed: 2026-08-19; locator: “Market annealing” framework and stages)
• ★★★ | Matt Bornstein, Guido Appenzeller, and Martin Casado — Who Owns the Generative AI Platform? | https://a16z.com/who-owns-the-generative-ai-platform/ | 2023-01-19 | coauthored market map | Dated stack, value-accrual, moat, and market-structure predictions; three-author evidence, not a sole Casado doctrine. (reliability: A; access: full; accessed: 2026-08-19; locator: stack diagram and conclusion)
• ★★★ | Guido Appenzeller, Matt Bornstein, and Martin Casado — Navigating the High Cost of AI Compute | https://a16z.com/navigating-the-high-cost-of-ai-compute/ | 2023-04-27 | coauthored essay | Concrete cost, capacity, procurement, and infrastructure constraints for AI companies; assumptions should be checked against later prices and architectures. (reliability: A; access: full; accessed: 2026-08-19; locator: cost drivers and mitigation sections)
• ★★★ | Martin Casado and Sarah Wang — The Economic Case for Generative AI and Foundation Models | https://a16z.com/the-economic-case-for-generative-ai-and-foundation-models/ | 2023-08-03 | coauthored essay/presentation | Core demand, cost-curve, surplus, and business-model argument; publication-date forecasts permit later falsification. (reliability: A; access: full; accessed: 2026-08-19; locator: presentation sections and figures)
• ★★★ | Martin Casado — Base AI Policy on Evidence, Not Existential Angst | https://a16z.com/base-ai-policy-on-evidence-not-existential-angst/ | 2024-12-16 | first-person policy essay | Current deregulatory/open-source thesis from a financially interested investor; pair with bill text, regulator evidence, and credible opposing cases. (reliability: A; access: full; accessed: 2026-08-19; locator: evidence, risk, and policy-recommendation sections)
• ★★ | Sarah Wang and Martin Casado — Questioning Margins Is a Boring Cliché | https://a16z.com/questioning-margins-is-a-boring-cliche/ | 2025-08-21 | coauthored essay | Challenges static margin screens and reframes value/cost dynamics; needs matched company outcomes and cash economics. (reliability: A; access: full; accessed: 2026-08-19; locator: opening thesis and examples)
• ★★★ | Martin Casado and Peter Lauten — The Empty Promise of Data Moats | https://a16z.com/the-empty-promise-of-data-moats/ | 2019-05-09 | coauthored essay | Early direct statement on data-network effects and defensibility; valuable precursor to later generative-AI moat claims. (reliability: A; access: full; accessed: 2026-08-19; locator: data-moat argument and conditions)
• ★★ | Peter Lauten and Martin Casado — Growth+Sales: The New Era of Enterprise Go-to-Market | https://a16z.com/growthsales-the-new-era-of-enterprise-go-to-market/ | 2020-07-29 | coauthored essay | Product-led growth, enterprise sales, and route-to-market evidence; verify each named example and preserve joint authorship. (reliability: A; access: full; accessed: 2026-08-19; locator: section headings)
• ★★ | Matt Bornstein, Martin Casado, and Jennifer Li — Emerging Architectures for Modern Data Infrastructure | https://a16z.com/emerging-architectures-for-modern-data-infrastructure-2020/ | 2020-10-15 | coauthored technical essay | Original dated data-stack map and market forecast connecting technical architecture to investment search; later landing-page updates are excluded from this information set. (reliability: A; access: full; accessed: 2026-08-19; locator: architecture diagrams, blueprint sections, and version boundary)
• ★★ | Jennifer Li and Martin Casado — Bottom Up Pricing & Packaging: Let the User Journey Be Your Guide | https://a16z.com/bottom-up-pricing-packaging-let-the-user-journey-be-your-guide/ | 2021-03-11 | coauthored essay | Pricing and adoption doctrine for developer-led companies; analysis must separate examples from demonstrated causal efficacy. (reliability: A; access: full; accessed: 2026-08-19; locator: user-journey and pricing sections)
• ★★ | Martin Casado and Katherine Boyle — AI Talks Leave “Little Tech” Out | https://www.wsj.com/articles/ai-talks-leave-little-tech-out-homeland-security-adversaries-open-source-board-46e3232d | 2024-05-15 | coauthored opinion essay | Headline-level policy signal only; no hidden reasoning, examples, or prescriptions are reconstructed without legitimate full access. (reliability: A for authored view; access: partial/paywalled; accessed: 2026-08-19; locator: headline-level record and a16z archive attribution; no full body)
• ★★ | Martin Casado and Ion Stoica — Keep the Code Behind AI Open, Say Two Entrepreneurs | https://www.economist.com/by-invitation/2024/07/29/keep-the-code-behind-ai-open-say-two-entrepreneurs | 2024-07-29 | coauthored guest essay | The Economist's public post identifies both writers and preserves one bounded joint excerpt; the article body, evidence, exceptions, and internal contribution allocation remain unresolved. (reliability: A for publisher attribution and excerpt; access: partial/paywalled; accessed: 2026-08-19; locator: canonical headline plus The Economist public-post attribution; no full body)
  • Supporting publisher route | The Economist — public attribution for Keep the Code Behind AI Open | https://www.linkedin.com/posts/the-economist_keep-the-code-behind-ai-open-say-two-entrepreneurs-activity-7224300277560946688-jvPw | 2024-07 | primary publisher social post | Names Martin Casado and Ion Stoica as the guest-essay writers and supplies one short attributed excerpt; does not substitute for the gated body. (reliability: A for publisher attribution; access: full; accessed: 2026-08-19; locator: post body, sentence immediately above linked By Invitation card)
• ★★ | Martin Casado and Anne Neuberger — America Cannot Lose the Robotics Race | https://www.a16z.news/p/america-cannot-lose-the-robotics | 2025-09-26 | coauthored newsletter/op-ed reproduction | Industrial-policy and robotics-capital thesis; use the full publisher copy while preserving joint authorship and advocacy interest. (reliability: A; access: full; accessed: 2026-08-19; locator: section headings)

Priority talks, podcasts, and transcript routes

• ★★★ | Invest Like the Best — Martin Casado: The Past, Present, and Future of Digital Infrastructure | https://joincolossus.com/episode/casado-the-past-present-and-future-of-digital-infrastructure/ | 2022-06-07 | long-form podcast | Full 56:44 publisher audio; current publisher text is only partial, with a secondary archival excerpt used as a locator and selected wording checked against audio. (reliability: A for direct voice, B for secondary text; access: full audio / partial current publisher transcript; accessed: 2026-08-20; locator: official enclosure timestamps, with approximately four-second page-clock drift)
• ★★★ | Invest Like the Best — Martin Casado: Entering Uncharted AI Territory | https://joincolossus.com/episode/casado-entering-uncharted-ai-territory/ | 2024-07-09 | long-form podcast | Full 1:08:03 publisher audio; current publisher text is partial and third-party ASR is navigation-only. (reliability: A for direct voice, B for navigation text; access: full audio / partial current publisher transcript; accessed: 2026-08-20; locator: official enclosure timestamps, with approximately one-minute page-clock drift)
• ★★★ | From the New World — Martin Casado: AI's Hard Limits | https://www.fromthenew.world/p/martin-casado-ai-hard-limits | 2024-09-02 | long-form podcast/post | Direct AI capability, constraint, policy, and market argument; preserve host framing and exact audio/transcript version. (reliability: A for direct voice, B for publisher framing; access: full; accessed: 2026-08-20; locator: chapter/timestamp scheme on publisher page)
• ★★ | Turpentine VC E70 — Martin Casado on AI Innovation and AGI | https://metacast.app/podcast/turpentine-vc-venture-capital-and-investing/LUJoKypp/e70-martin-casado-of-a16z-on-ai-innovation-and-agi/zb2A3RvL | 2024-12-24 | replay podcast | E70 is a 1:43:40 replay of the 2024-06-28 Cognitive Revolution original, whose 1:48:20 publisher audio/video and speaker-labelled transcript supply the evidence; count once. (reliability: A for original direct record, B for replay metadata; access: full original transcript/audio and full replay audio; accessed: 2026-08-20; locator: original publisher timestamps)
• ★★★ | a16z Podcast — Where Value Will Accrue in AI, with Martin Casado and Sarah Wang | https://a16z.com/podcast/where-value-will-accrue-in-ai-martin-casado-sarah-wang/ | 2025-05-27 | firm LP Summit panel | Full 21:41 official audio, no publisher transcript; third-party ASR was navigation-only and Casado/Wang turns remain separate. (reliability: A for direct voice, B for navigation text; access: full audio / partial text; accessed: 2026-08-20; locator: official-audio timestamps)
• ★★ | a16z Podcast — AI Is Upending SaaS Pricing | https://a16z.com/podcast/ai-is-upending-saas-pricing/ | 2025-07-18 | firm podcast | Full 42:52 official audio with partial public transcript and bounded speaker-checked navigation ASR; preserve Scott Woody's separate operating claims and Metronome interest. (reliability: A for direct voice, B for navigation text; access: full audio / partial text; accessed: 2026-08-20; locator: official-audio timestamps)
• ★★★ | a16z Podcast — Why a16z's Martin Casado Believes the AI Boom Still Has Years to Run | https://a16z.com/podcast/why-a16zs-martin-casado-believes-the-ai-boom-still-has-years-to-run/ | 2025-12-30 | replay of Generalist interview | Underlying Generalist interview published 2025-10-07 (1:24:44); a16z replay is 1:22:20. Count once and date claims to the original. (reliability: A for original audio, B for chapter map/third-party transcript; access: full original and replay audio / partial text; accessed: 2026-08-20; locator: Generalist original chapter times)
• ★★★ | a16z Podcast — AI's Capital Flywheel: Models, Money, and the Future of Power | https://a16z.com/podcast/ais-capital-flywheel-models-money-and-the-future-of-power/ | 2026-02-19 | replay of Latent Space interview | Original Latent Space audio/transcript is 55:18; a16z replay is 57:52 with inserted material. Count once and preserve Casado/Wang attribution. (reliability: A for original direct record; access: full original transcript/audio and replay audio; accessed: 2026-08-20; locator: original Latent Space timestamps)
• ★★★ | 20VC — Martin Casado on whether the venture model is broken | https://www.tapesearch.com/episode/20vc-a16z-s-martin-casado-on-how-the-venture-model-is-broken-why-vcs-should-be-running-wall-st-who-wins-and-who-loses-in-the-next-generation-of-venture-investing-lessons-from-marc-andreesen-ben-horowitz-and-chris-dixon/BbFqixXybsSEEV8iJT3CuE | 2022-12-05 | long-form podcast | Complete 51:52 official RSS audio; public transcript routes are incomplete/gated, so bounded ASR supported timestamped paraphrase and no verbatim quote. (reliability: A for direct official audio, B for navigation text; access: full audio / partial text; accessed: 2026-08-20; locator: official Libsyn audio timestamps)
• ★★★ | a16z Podcast — Is Non-Consensus Investing Overrated? | https://a16z.com/podcast/is-non-consensus-investing-overrated/ | 2025-09-04 | firm podcast | Direct current debate on contrarianism, consensus, selection, and venture process; preserve participant attribution and firm-production bias. (reliability: A for direct voice; access: full transcript/audio; accessed: 2026-08-20; locator: publisher transcript timecodes)
• ★★ | a16z Podcast — Infrastructure… Is Everything | https://a16z.com/podcast/a16z-podcast-infrastructure-is-everything/ | 2016-02-25 | firm podcast | Earliest transition-era direct discussion of infrastructure and operating formation; complete official audio was reviewed with bounded navigation text and no machine-only quote. (reliability: A for direct voice; access: partial text/full audio; accessed: 2026-08-20; locator: audio timestamps; show notes)
• ★★ | a16z Podcast — Selling to Developers & Open Source Business Models | https://a16z.com/podcast/a16z-podcast-selling-to-developers/ | 2016-04-13 | firm podcast | Early developer-first go-to-market and open-source business-model evidence. (reliability: A for direct voice; access: partial text/full audio; accessed: 2026-08-20; locator: audio timestamps; show notes)
• ★★ | a16z — Trends in—and the Future of—Infrastructure | https://a16z.com/trends-in-and-the-future-of-infrastructure/ | 2016-08-21 | 23-minute talk/video | Early a16z infrastructure map close to investment-career entry; full video permits timestamped analysis. (reliability: A for direct voice; access: full video; accessed: 2026-08-20; locator: video timestamps)
• ★★ | a16z Podcast — Pricing, Pricing, Pricing | https://a16z.com/podcast/a16z-podcast-pricing-pricing-pricing/ | 2016-08-13 | firm podcast with transcript | Early pricing doctrine and examples; preserve speaker turns. (reliability: A for direct voice; access: full transcript/audio; accessed: 2026-08-20; locator: timestamped transcript)
• ★★ | a16z Podcast — All About Microservices | https://a16z.com/podcast/a16z-podcast-all-about-microservices/ | 2016-09-01 | firm podcast with transcript | Technical and market-architecture evidence from the first investing year. (reliability: A for direct voice; access: full transcript/audio; accessed: 2026-08-20; locator: timestamped transcript)
• ★★ | a16z — Software Secures the World | https://a16z.com/software-secures-the-world/ | 2017-02-12 | 20-minute talk/video | Security thesis linking technical formation to market search; claim-level outcome tests remain later work. (reliability: A for direct voice; access: full video; accessed: 2026-08-20; locator: video timestamps)
• ★★ | a16z — Going to Market When No Market Exists | https://a16z.com/going-to-market-when-no-market-exists/ | 2017-06-09 | talk/video and deck | Early precursor to market annealing and category formation, with a full audiovisual route. (reliability: A for direct voice; access: full video/page; accessed: 2026-08-20; locator: video timestamps and deck sections)
• ★ | a16z — Navigating Uncertainty: Advice for Graduates (and Others Changing Course) | https://a16z.com/navigating-uncertainty-advice-for-graduates-and-others-changing-course/ | 2017-07-02 | eight-minute talk/video | Biographical judgment and career advice; lower direct investment value but useful for formation/evolution. (reliability: A for direct voice; access: full video; accessed: 2026-08-20; locator: video timestamps)
• ★★ | a16z Podcast — Creating a Category, from Pricing to Positioning | https://a16z.com/podcast/a16z-podcast-creating-a-category-from-pricing-to-positioning/ | 2018-02-24 | three-speaker firm podcast | Complete 31:50 official audio was reviewed with strict Casado/Feaster/Chokshi separation; local ASR was navigation-only, material passages and handoffs were manually checked, and no verbatim quote was promoted. (reliability: A for direct audio, B for navigation text; access: full audio / partial text; accessed: 2026-08-20; locator: exact official-audio timestamps)
• ★★ | a16z — The World Through an API | https://a16z.com/the-world-through-an-api/ | 2018-03-09 | 18-minute talk/video | API-economy thesis close to Kong/Mashape-era investments; requires deal-attribution separation. (reliability: A for direct voice; access: full video; accessed: 2026-08-20; locator: video timestamps)
• ★★★ | Software Engineering Daily — Software Chasms with Martin Casado | https://softwareengineeringdaily.com/2019/01/28/software-chasms-with-martin-casado/ | 2019-01-28 | long-form podcast with publisher transcript | Detailed technical-to-market transition and infrastructure history; strong early spoken route. (reliability: A for direct voice; access: full transcript/audio; accessed: 2026-08-20; locator: transcript/timestamps)
• ★★ | a16z Podcast — Gross Margins: Early to Late, What They Do (and Don't) Tell You | https://a16z.com/podcast/a16z-podcast-gross-margins-early-to-late-what-they-do-and-dont-tell-you/ | 2020-06-26 | firm podcast | Margin-screen evidence to compare with 2025 writing; complete official audio was speaker-checked with secondary text used only for navigation. (reliability: A for direct voice; access: partial text/full audio; accessed: 2026-08-20; locator: audio timestamps; show notes)
• ★★ | a16z Podcast — On Fear and Leadership: Product to Sales, CTOs to CEOs | https://a16z.com/podcast/a16z-podcast-on-fear-and-leadership-product-to-sales-ctos-ceos/ | 2020-12-04 | firm podcast | Founder-role and leadership-transition evidence relevant to Nicira's CEO sequence and board practice. (reliability: A for direct voice; access: partial text/full audio; accessed: 2026-08-20; locator: audio timestamps; show notes)
• ★★ | Software Engineering Daily — Cloud Cost with Martin Casado | https://softwareengineeringdaily.com/2021/06/25/cloud-cost-with-martin-casado/ | 2021-06-25 | podcast with publisher transcript route | Direct elaboration of the cloud-cost thesis; pair with the coauthored essay and later outcomes. (reliability: A for direct voice; access: full transcript/audio; accessed: 2026-08-20; locator: transcript/timestamps)
• ★★ | a16z Podcast — Cloud Wars, Company Wars, and Innovating Through Change | https://a16z.com/podcast/cloud-wars-company-wars-and-innovating-through-change/ | 2022-09-07 | multi-speaker firm podcast | Competitive transition and operator history; separate Casado's turns from co-speakers. (reliability: A for direct voice; access: partial text/full audio; accessed: 2026-08-20; locator: audio timestamps; show notes)
• ★★ | a16z — The Economic Case for Generative AI | https://a16z.com/the-economic-case-for-generative-ai/ | 2023-09-25 | presentation with transcript | Publisher transcript spans 00:00–14:39; 13:35 is the final chapter start, not a verified duration. Solo restatement of the 2023-08-03 Casado/Wang essay, so not independent thesis corroboration. (reliability: A for direct voice; access: full video/transcript; accessed: 2026-08-20; locator: publisher transcript timestamps)
• ★★★ | a16z — Why America Must Lead in AI Investment | https://a16z.com/why-america-must-lead-in-ai-investment/ | 2024-03-05 | policy interview/podcast with transcript | Direct policy and investment argument with full transcript; advocacy interest and co-speaker attribution remain explicit. (reliability: A for direct voice; access: full transcript/audio; accessed: 2026-08-20; locator: timestamped transcript)
• ★★ | a16z Podcast — The Future of Software Development: Vibe Coding, Prompt Engineering, AI Assistants | https://a16z.com/podcast/the-future-of-software-development-vibe-coding-prompt-engineering-ai-assistants/ | 2025-07-21 | multi-speaker firm podcast | Full 44:23 official audio and official YouTube; no publisher transcript, with third-party ASR navigation-only and all speaker turns preserved. (reliability: A for direct voice, B for navigation text; access: full audio/video / partial text; accessed: 2026-08-20; locator: official video/audio timestamps)
• ★★ | a16z Podcast — The State of American AI Policy: From Pause AI to Build | https://a16z.com/podcast/the-state-of-american-ai-policy-from-pause-ai-to-build/ | 2025-08-15 | multi-speaker firm podcast | Full 42:01 official audio and official YouTube; no publisher transcript, with third-party ASR navigation-only and independent policy sources used for comparison. (reliability: A for direct voice, B for navigation text; access: full audio/video / partial text; accessed: 2026-08-20; locator: official video/audio timestamps)
• ★★★ | a16z Podcast — Jack Altman and Martin Casado on the Future of VC | https://a16z.com/podcast/jack-altman-martin-casado-on-the-future-of-vc/ | 2025-09-03 original; 2025-09-04 a16z republication | Uncapped interview and a16z feed-drop edit | Original Uncapped GUID df6fdb44-887f-11f0-94e9-b3e2a6697529 is 52:10; a16z Simplecast edit ac6fc175-2685-49b6-b64b-1d66e87b4ae4 is 53:28. Complete official audio plus a full third-party timestamped navigation transcript; publisher identifies Jack Altman as host. (reliability: A for direct voice, B for navigation text; access: full audio / partial publisher text; accessed: 2026-08-20; locator: a16z edit timestamps)
• ★★ | a16z Podcast — Monopolies vs. Oligopolies in AI | https://a16z.com/podcast/monopolies-vs-oligopolies-in-ai/ | 2026-03-20 | wrapper around 2025 interview | One underlying 20VC interview: original 2025-07-28 (1:10:37), a16z feed drop 2025-08-28 (1:17:25), web wrapper 2026-03-20. Count once; no publisher transcript. (reliability: A for official audio/identity, B for navigation ASR; access: full original and a16z-distributed audio / partial text; accessed: 2026-08-20; locator: a16z feed-drop audio timestamps)

Supporting recording, transcript, and counterevidence routes used in T100.3

These are version, navigation, or independent-comparison routes for the 31 priority families above. They do not add to the priority denominator.

• Official Simplecast players for the 2016 a16z episodes: Infrastructure, Selling to Developers, Pricing, and Microservices. Exact publisher audio clocks used where creator transcripts were absent. (reliability: A for direct audio; access: full; accessed: 2026-08-20; locator: elapsed player timestamps)
• Official Simplecast players for Creating a Category, Gross Margins, and Fear and Leadership, plus the Creating a Category API/player record. Complete publisher audio; no creator transcript for these versions. (reliability: A for direct audio and publisher metadata; access: full audio / partial text; accessed: 2026-08-20; locator: elapsed player timestamps and API metadata)
• Official a16z/AI player for AI Is Upending SaaS Pricing. Complete 42:52 audio with partial public text. (reliability: A for direct audio; access: full audio / partial text; accessed: 2026-08-20; locator: episode player timestamps)
• Official a16z player records for Future of Software Development, State of American AI Policy, and Monopolies vs. Oligopolies. Exact IDs, durations, and a16z feed versions. (reliability: A for publisher metadata/audio; access: full audio; accessed: 2026-08-20; locator: player metadata and elapsed time)
• Current Colossus routes for Digital Infrastructure and Uncharted AI. Current publisher audio is full while page transcript text is partial. (reliability: A for direct audio; access: full audio / partial text; accessed: 2026-08-20; locator: chapters and enclosure timestamps)
• Software Engineering Daily publisher transcripts: SED756 — Software Chasms and SED1287 — Cloud Cost. (reliability: A; access: full; accessed: 2026-08-20; locator: PDF pages and printed timestamps)
• Official a16z YouTube recordings: Trends/Future Infrastructure, Software Secures the World, Going to Market, Navigating Uncertainty, World Through an API, Economic Case presentation, Future Software, and State of American AI Policy. (reliability: A for direct voice; access: full video/captions where noted; accessed: 2026-08-20; locator: video timestamps)
• Official 20VC episode pages for the 2022 venture-model interview and the 2025 market-structure interview. Full official RSS audio and publisher framing; no publisher transcript. (reliability: A for direct audio, B for show-note framing; access: full audio / partial text; accessed: 2026-08-20; locator: official audio clocks)
• Original Uncapped #23 routes: Spotify for Creators episode, official RSS, and Apple episode. These identify the 2025-09-03 original, GUID df6fdb44-887f-11f0-94e9-b3e2a6697529, and 52:10 runtime; the separate a16z edit is 53:28. (reliability: A for publisher feed/audio, B for directory metadata; access: full audio / partial text; accessed: 2026-08-20; locator: RSS item and player clocks)
• Original publisher routes behind later replays: Cognitive Revolution, Generalist, and Latent Space. Used to bind the original interviews, speakers, and clocks rather than treat later a16z/Turpentine versions as new talks. (reliability: A for original direct records, B where transcript is machine-generated; access: full media, full text for Cognitive Revolution/Latent and chapter map for Generalist; accessed: 2026-08-20; locator: original timestamps/chapters)
• Full publisher transcript for Is Non-Consensus Investing Overrated?. (reliability: A; access: full transcript/audio; accessed: 2026-08-20; locator: transcript timecodes)
• Secondary navigation text: Digital Infrastructure archival excerpt, Uncharted AI, Value Accrual, AI Pricing, Future Software, State Policy, Future VC, Monopolies, and AI Boom replay transcript. Used only for navigation/wording checks against exact media, not as independent evidence. (reliability: B/C; access: full or partial as disclosed in notes; accessed: 2026-08-20; locator: third-party timestamps)
• Firm disclosure route: a16z disclosures. Identifies interested-investor context; it does not enumerate all company-specific positions. (reliability: A for firm disclosure; access: full; accessed: 2026-08-20; locator: disclosure page)
• AI coding counterevidence: METR randomized experienced-developer study (2025-07-10). Relevant to mature-repository productivity, not every coding task. (reliability: A for research report; access: full; accessed: 2026-08-20; locator: results/method sections)
• Cloud-market counterevidence: FTC cloud-computing RFI synthesis and NBER cloud concentration paper. Used for switching, concentration, and platform-power comparisons rather than Casado attribution. (reliability: A for government/working-paper records; access: full; accessed: 2026-08-20; locator: findings/abstract and body)
• Open-model and AI-policy controls: NTIA open-model-weights report, California SB 1047 decision, and 2025 U.S. AI Action Plan. Primary policy records used to test episode characterizations. (reliability: A; access: full; accessed: 2026-08-20; locator: executive summary/decision text and named policy pillars)

Vehicle, deal-attribution, exit, and failure records

• ★★★ | Andreessen Horowitz — portfolio directory and disclosure | https://a16z.com/portfolio/ | live; accessed 2026-08-20 | official firm database | Firm candidate universe and warning that the list is incomplete and “exited” does not imply profit; no Casado hit-rate inference. (reliability: A; access: full; accessed: 2026-08-20; locator: disclosure below portfolio table)
• ★★★ | Andreessen Horowitz — canonical investment list | https://a16z.com/investment-list/ | live; accessed 2026-08-20 | dynamic firm database | Denominator/control route; dynamic output is not a stable complete person-linked export and firm disclosure says some positions are omitted. (reliability: A; access: partial/dynamic; accessed: 2026-08-20; locator: dynamic table and disclosure)
• ★★ | Forbes — Martin Casado profile | https://www.forbes.com/profile/martin-casado/ | updated 2026-05-27 | reported investor profile | Additional investment/recognition leads; neither complete denominator nor audited economics. (reliability: B; access: full; accessed: 2026-08-20; locator: “From the Editor” and investment-summary paragraphs)
• ★★★ | SEC — Andreessen Horowitz Fund VII Form D | https://www.sec.gov/Archives/edgar/data/1829358/000182935820000001/xslFormDX08/primary_doc.xml | 2020-12-02 | SEC filing | Vehicle identity/related-person record; no Casado allocation, carry, or performance. (reliability: A; access: full historically/automated 403 on 2026-08-20; accessed: 2026-08-20; locator: Items 2, 3, 13–15)
• ★★★ | Andreessen Horowitz — New Funds, New Era | https://a16z.com/new-funds-new-era/ | 2024-04-16 | firm fundraising announcement | $1.25B Infrastructure pool inside $7.2B; size is not deployment or performance. (reliability: A for firm claim; access: full; accessed: 2026-08-20; locator: allocation paragraph and disclosure block)
• ★★★ | Andreessen Horowitz — Why Did We Raise $15B? | https://a16z.com/why-did-we-raise-15b/ | 2026-01-09 | firm fundraising announcement | $1.7B Infrastructure allocation; no decision-right, carry, reserve, or return disclosure. (reliability: A for firm claim; access: full; accessed: 2026-08-20; locator: allocation list)

Official Casado-linked deal-announcement corpus — 43 records

Every row below is a primary record of what its named author(s) publicly wrote, not proof of sole decision rights, audited company metrics, ownership, or return.

• ★★★ | Martin Casado — SigOpt | https://a16z.com/announcement/sigopt/ | 2016-08-24 | investor-authored announcement | Optimization thesis and explicit deepening of a pre-existing seed position. (reliability: A; access: full; accessed: 2026-08-20; locator: header/byline/date, opening and closing investment paragraphs)
• ★★★ | Martin Casado — RapidAPI | https://a16z.com/announcement/rapidapi/ | 2016-11-21 | investor-authored announcement | API-market thesis and investment record. (reliability: A; access: full; accessed: 2026-08-20; locator: header/byline/date, thesis body, closing)
• ★★★ | Martin Casado — Mashape/Kong | https://a16z.com/announcement/mashape/ | 2017-03-23 | investor-authored announcement | Binds Mashape to Kong and prevents false double counting. (reliability: A; access: full; accessed: 2026-08-20; locator: identity/use and closing paragraphs)
• ★★★ | Martin Casado — ActionIQ | https://a16z.com/announcement/actioniq/ | 2017-10-23 | investor-authored announcement | Customer-data thesis, prior relationship, and disclosed enterprise diligence. (reliability: A; access: full; accessed: 2026-08-20; locator: relationship/diligence and closing paragraphs)
• ★★★ | Martin Casado — Astranis | https://a16z.com/announcement/astranis/ | 2018-03-01 | investor-authored announcement | MicroGEO thesis, outside-domain learning, investment and board. (reliability: A; access: full; accessed: 2026-08-20; locator: architecture and closing board paragraphs)
• ★★★ | Martin Casado — Imply | https://a16z.com/announcement/imply/ | 2018-03-13 | investor-authored announcement | Druid/analytics thesis and a16z-led Series A. (reliability: A; access: full; accessed: 2026-08-20; locator: opening Druid context and closing A paragraph)
• ★★★ | Martin Casado — Yubico | https://a16z.com/announcement/yubico/ | 2018-05-31 | investor-authored announcement | Root-of-trust thesis, product/reference diligence, capital-efficiency claim and board. (reliability: A; access: full; accessed: 2026-08-20; locator: opening, diligence and closing board paragraphs)
• ★★★ | Martin Casado — Eclypsium | https://a16z.com/announcement/eclypsium/ | 2018-12-04 | investor-authored announcement | Firmware-security thesis and seed lead. (reliability: A; access: full; accessed: 2026-08-20; locator: security thesis and closing seed paragraph)
• ★★★ | Martin Casado — Fivetran | https://a16z.com/announcement/fivetran/ | 2019-09-24 | investor-authored announcement | Automated-pipeline thesis, customer encounters, B lead and board. (reliability: A; access: full; accessed: 2026-08-20; locator: thesis/references and closing)
• ★★★ | Martin Casado — Preset | https://a16z.com/announcement/preset/ | 2019-10-10 | investor-authored announcement | Superset commercialization thesis. (reliability: A; access: full; accessed: 2026-08-20; locator: Superset/company distinction and closing)
• ★★★ | Martin Casado — Instabase | https://a16z.com/announcement/instabase/ | 2019-10-21 | investor-authored announcement | Prior-investment chronology, diligence, and workflow-platform thesis. (reliability: A; access: full; accessed: 2026-08-20; locator: “first invested,” diligence, and closing)
• ★★★ | Martin Casado and Ben Horowitz — Anyscale | https://a16z.com/announcement/why-were-investing-in-anyscale/ | 2019-12-17 | coauthored announcement | Ray/distributed-computing thesis, Series A lead and collective board statement. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, thesis and closing)
• ★★★ | Martin Casado — Fishtown Analytics/dbt | https://a16z.com/announcement/investing-in-fishtown-analytics-the-makers-of-dbt/ | 2020-04-22 | investor-authored announcement | Company-name continuity and analytics-engineering thesis. (reliability: A; access: full; accessed: 2026-08-20; locator: identity/thesis and closing)
• ★★★ | Martin Casado — Tecton | https://a16z.com/announcement/investing-in-tecton/ | 2020-04-28 | investor-authored announcement | Michelangelo/feature-store thesis and seed-to-A chronology. (reliability: A; access: full; accessed: 2026-08-20; locator: feature-store body and closing seed/A)
• ★★★ | Martin Casado — Rasa | https://a16z.com/announcement/investing-in-rasa/ | 2020-06-23 | investor-authored announcement | Open-source conversational-AI thesis, B lead and historical board. (reliability: A; access: full; accessed: 2026-08-20; locator: product/community and closing)
• ★★★ | Martin Casado — Material Security | https://a16z.com/announcement/investing-in-material-security/ | 2020-06-30 | investor-authored announcement | Email-security thesis and partnership. (reliability: A; access: full; accessed: 2026-08-20; locator: threat/product body and closing)
• ★★★ | Martin Casado — Isovalent | https://a16z.com/announcement/investing-in-isovalent/ | 2020-11-10 | investor-authored announcement | eBPF/Cilium thesis and founder-network evidence. (reliability: A; access: full; accessed: 2026-08-20; locator: adoption/team and closing)
• ★★★ | Martin Casado — Orbit | https://a16z.com/announcement/investing-in-orbit/ | 2020-12-02 | investor-authored announcement | Developer-community system-of-record thesis. (reliability: A; access: full; accessed: 2026-08-20; locator: community body and closing)
• ★★★ | Martin Casado — Parsec | https://a16z.com/announcement/investing-in-parsec/ | 2020-12-15 | investor-authored announcement | 2017 demo, community/enterprise evidence, B lead and board. (reliability: A; access: full; accessed: 2026-08-20; locator: demo/customer and closing)
• ★★★ | Martin Casado — Tackle | https://a16z.com/announcement/investing-in-tackle/ | 2021-03-09 | investor-authored announcement | Portfolio-board sourcing, references, B lead and board. (reliability: A; access: full; accessed: 2026-08-20; locator: discovery/traction and closing)
• ★★★ | Martin Casado — Tabular | https://a16z.com/announcement/investing-in-tabular/ | 2021-08-31 | investor-authored announcement | Iceberg/open-table thesis and Series A lead. (reliability: A; access: full; accessed: 2026-08-20; locator: team/thesis and closing)
• ★★★ | Martin Casado — EngFlow | https://a16z.com/announcement/investing-in-engflow/ | 2021-10-12 | investor-authored announcement | Build-system thesis, customer evidence and seed lead. (reliability: A; access: full; accessed: 2026-08-20; locator: customer and closing)
• ★★★ | Martin Casado — Truffle Security | https://a16z.com/announcement/investing-in-truffle-security/ | 2021-12-08 | investor-authored announcement | Secret-scanning thesis and A lead. (reliability: A; access: full; accessed: 2026-08-20; locator: risk thesis and closing)
• ★★★ | Martin Casado — Ambient.ai | https://a16z.com/announcement/investing-in-ambient-ai/ | 2022-01-19 | investor-authored announcement | Computer-vision physical-security thesis. (reliability: A; access: full; accessed: 2026-08-20; locator: problem/product and closing)
• ★★★ | Martin Casado — Metronome | https://a16z.com/announcement/investing-in-metronome/ | 2022-02-01 | investor-authored announcement | Usage-billing thesis, portfolio references, A lead and board. (reliability: A; access: full; accessed: 2026-08-20; locator: pricing/references and closing)
• ★★★ | Martin Casado — Convex | https://a16z.com/announcement/investing-in-convex/ | 2022-04-27 | investor-authored announcement | Backend thesis, A lead and board. (reliability: A; access: full; accessed: 2026-08-20; locator: product body and closing)
• ★★★ | Martin Casado — Fly.io | https://a16z.com/announcement/investing-in-fly-io/ | 2022-07-28 | investor-authored announcement | Edge/developer-cloud thesis and B lead. (reliability: A; access: full; accessed: 2026-08-20; locator: architecture and closing)
• ★★★ | Martin Casado and Jennifer Li — MotherDuck | https://a16z.com/announcement/investing-in-motherduck/ | 2022-11-15 | coauthored announcement | DuckDB/hybrid analytics thesis and A lead. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, thesis and closing)
• ★★★ | Martin Casado and Matt Bornstein — Coactive | https://a16z.com/announcement/investing-in-coactive/ | 2023-03-21 | coauthored announcement | Visual-data thesis and seed+A chronology. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, product/founder and closing)
• ★★★ | Martin Casado and Matt Bornstein — Distributional | https://a16z.com/announcement/investing-in-distributional/ | 2023-12-14 | coauthored announcement | AI-testing thesis, 25+ interviews and SigOpt relationship. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, diligence/relationship and closing)
• ★★★ | Martin Casado — Tigris | https://a16z.com/announcement/investing-in-tigris/ | 2024-02-20 | investor-authored announcement | Distributed object-storage thesis. (reliability: A; access: full; accessed: 2026-08-20; locator: storage body and closing)
• ★★★ | Martin Casado and Jennifer Li — Ideogram | https://a16z.com/announcement/investing-in-ideogram/ | 2024-02-28 | coauthored announcement | Generative-media thesis and seed+A chronology. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, model/product and closing)
• ★★★ | Matt Bornstein, Marco Mascorro, Rajko Radovanovic, and Martin Casado — Cursor | https://a16z.com/announcement/investing-in-cursor/ | 2024-08-22 | four-author announcement | Workflow/model-choice thesis and Series A; preserves team credit. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, usage/model/founder body and closing)
• ★★★ | Martin Casado and Sarah Wang — World Labs | https://a16z.com/announcement/investing-in-world-labs/ | 2024-09-13 | coauthored announcement | Spatial-intelligence thesis and partnership. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, world-model body and closing)
• ★★★ | Martin Casado — Braintrust | https://a16z.com/announcement/investing-in-braintrust/ | 2024-10-08 | investor-authored announcement | AI-evaluation thesis, product use and Series A lead. (reliability: A; access: full; accessed: 2026-08-20; locator: eval body and lead/user sentence)
• ★★★ | Yoko Li, Jennifer Li, and Martin Casado — Resend | https://a16z.com/announcement/investing-in-resend/ | 2024-12-04 | three-author announcement | Developer-email thesis, A lead and collective board statement. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, product body and closing)
• ★★★ | Justine Moore, Marco Mascorro, and Martin Casado — WaveForms AI | https://a16z.com/announcement/investing-in-waveforms-ai/ | 2024-12-09 | three-author announcement | Audio-model thesis and investment. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, model/team and closing)
• ★★★ | Yoko Li, Martin Casado, and Jennifer Li — Phota Labs | https://a16z.com/announcement/investing-in-phota-labs/ | 2025-09-16 | three-author announcement | Identity-preserving photo thesis and seed lead. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, identity/context and closing)
• ★★★ | Yoko Li, Guido Appenzeller, and Martin Casado — Relace | https://a16z.com/announcement/investing-in-relace/ | 2025-10-08 | three-author announcement | Coding-agent thesis, $23M A lead and collective board statement. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, stack body and closing)
• ★★★ | Justine Moore, Marco Mascorro, and Martin Casado — Shizuku AI | https://a16z.com/announcement/investing-in-shizuku-ai/ | 2026-02-09 | three-author announcement | AI-companion thesis and seed lead. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, thesis and closing)
• ★★★ | Yoko Li, Guido Appenzeller, and Martin Casado — QuiverAI | https://a16z.com/announcement/investing-in-quiverai/ | 2026-02-25 | three-author announcement | SVG-model thesis and seed lead. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, SVG body and closing)
• ★★★ | Marco Mascorro and Martin Casado — Deeptune | https://a16z.com/announcement/investing-in-deeptune/ | 2026-03-19 | coauthored announcement | RL-environment thesis and Series A lead. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, thesis and closing)
• ★★★ | Martin Casado, Yoko Li, and Guido Appenzeller — Runta | https://a16z.com/announcement/investing-in-runta/ | 2026-07-16 | three-author announcement | Agent-compute thesis, prior founder relationship and seed lead. (reliability: A; access: full; accessed: 2026-08-20; locator: exact byline, relationship and closing)

Previously mapped outcome and adverse controls

• ★★★ | Tackle CEO — company update | https://tackle.io/blog/tackle-company-update/ | 2023-06-21 | first-person company statement | Records 75 separations and admits 2021–22 overbuilding; adverse operating evidence, not proof of investor loss. (reliability: A for company statement; access: full; accessed: 2026-08-20; locator: “Today is a tough day,” “overbuilt,” runway, and model paragraphs)
• ★★★ | Unity Software Form 10-K — Parsec acquisition | https://www.sec.gov/Archives/edgar/data/1810806/000181080622000004/unity-20211231.htm | 2022-02-23 | SEC-filed acquisition note | $332.729M company cash consideration; no a16z cost, ownership, proceeds, or return. (reliability: A; access: full; accessed: 2026-08-20; locator: Note 6, Parsec table)
• ★★★ | Fivetran and dbt Labs — completed all-stock merger | https://www.fivetran.com/press/fivetran-dbt-labs-complete-merger-to-create-the-data-infrastructure-for-trusted-ai-agents | 2026-06-01 | issuer transaction announcement | Completed all-stock combination; exchange ratio, value, liquidity, and gain/loss undisclosed. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening and transaction-description paragraphs)

T100.4 vehicle, outcome, adverse, and base-rate controls

• ★★★ | a16z — Fund V | https://a16z.com/fund-v/ | 2016-06-10 | firm fundraising announcement | $1.5B multi-stage platform context; no performance or Casado-specific allocation. (reliability: A; access: full; accessed: 2026-08-20; locator: announcement and stage-allocation paragraphs)
• ★★★ | a16z — Fund VII and Growth II | https://a16z.com/fund-vii-and-growth-fund-ii/ | 2020-11-20 | firm fundraising announcement | $1.3B/$3.2B rounded amounts and separate-team description. (reliability: A; access: full; accessed: 2026-08-20; locator: opening allocation and team paragraphs)
• ★★★ | SEC — Fund VII Form D alternate XML | https://www.sec.gov/Archives/edgar/data/1829358/000182935820000001/xslFormDX01/primary_doc.xml | 2020-11-20 | regulatory filing | Exact $1.25B, 154 investors, first-sale date and managing chain; automated recheck may return 403. (reliability: A; access: full historically/automated 403 on 2026-08-20; accessed: 2026-08-20; locator: Items 3, 5, 12–14)
• ★★★ | SEC — Fund IX AI Infrastructure Form D | https://www.sec.gov/Archives/edgar/data/2011182/000110465924045700/xslFormDX08/primary_doc.xml | 2024-04-16 | regulatory filing | Vehicle identity, GP chain, offering and management-fee entitlement; not performance. (reliability: A; access: full; accessed: 2026-08-20; locator: Items 1, 3, 13, 15)
• ★★★ | SEC — Fund X AI Infrastructure Form D | https://www.sec.gov/Archives/edgar/data/2085104/000110465926002072/xslFormDX01/primary_doc.xml | 2026-01-09 | regulatory filing | Successor vehicle identity, GP chain, offering and fee entitlement. (reliability: A; access: full; accessed: 2026-08-20; locator: Items 1, 3, 13, 15)
• ★★★ | SEC/IAPD — A16Z Capital Management Form ADV | https://reports.adviserinfo.sec.gov/reports/ADV/160489/PDF/160489.pdf | amended 2026-03-30 | regulatory PDF | Ownership/control schedules, RAUM, Fund IX/X GAV/owners/valuation fields and multiplexer double-counting warning. (reliability: A; access: full; accessed: 2026-08-20; locator: Item 5.F; Schedules A/B; Schedule D §7.B.(1) fund IDs; Miscellaneous)
• ★★★ | Cursor — Series C | https://cursor.com/blog/series-c | 2025-06-06 | issuer financing post | $900M/$9.9B and issuer metrics; financing mark not proceeds. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening financing and ARR/customer paragraphs)
• ★★★ | Cursor — Series D | https://cursor.com/blog/series-d | 2025-11-13 | issuer financing post | $2.3B/$29.3B and issuer annualized revenue; unaudited. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening financing/revenue paragraphs)
• ★★★ | SpaceX — Cursor merger 8-K | https://www.sec.gov/Archives/edgar/data/1181412/000162828026043411/spaceexplorationtechnologi.htm | 2026-06-18 | SEC filing | All-stock mechanics and $60B implied Cursor equity value; not holder proceeds. (reliability: A; access: full; accessed: 2026-08-20; locator: Item 1.01, “Merger Consideration”)
• ★★★ | Cursor — Joining SpaceX | https://cursor.com/blog/joining-spacex | 2026-08-14 | issuer transaction post | Confirms completion; no holder allocation/return. (reliability: A; access: full; accessed: 2026-08-20; locator: opening two paragraphs)
• ★★★ | Yubico — ACQ Bure transaction archive | https://investors.yubico.com/en/investors/listing-nasdaq-first-north-growth-market/ | 2023/live | issuer transaction archive | Merger/listing chronology and documents; no security-level a16z realization. (reliability: A; access: full; accessed: 2026-08-20; locator: opening chronology and linked documents)
• ★★★ | Yubico/ACQ Bure — transaction announcement | https://investors.yubico.com/en/acq-bure-ab-is-merging-with-yubico-ab-merged-companies-intend-to-go-public-on-swedens-nasdaq-first-north-growth-market/ | 2023-04-19 | issuer transaction release | Approximate $800M company transaction value and structure; expected 6.1% combined-company a16z ownership and conditional 180-day lockup; not actual closing shares, sales, proceeds, distributions, or return. (reliability: A; access: full; accessed: 2026-08-20; locator: transaction overview and consideration lines; “Ownership structure” table; Andreessen Horowitz 180-day lockup paragraph)
• ★★★ | Yubico — first trading day | https://investors.yubico.com/en/first-day-of-trading-in-yubicos-shares-on-nasdaq-first-north-growth-market-following-completion-of-merger-with-acq-bure/ | 2023-09-20 | issuer release | Confirms merger completion/listing; listing is not realization. (reliability: A; access: full; accessed: 2026-08-20; locator: opening completion/trading paragraphs)
• ★★★ | Parsec — Series B | https://parsec.app/blog/about-our-series-b-and-parsecs-year-of-transformation | 2020-12-15 | issuer financing post | $25M round, syndicate and board relationship. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening financing/board paragraphs)
• ★★★ | Tackle — Series B | https://tackle.io/blog/tackle-series-b-funding/ | 2021-03-09 | issuer financing post | $35M B and issuer operating claims. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening and ARR/customer/volume paragraphs)
• ★★★ | Tackle — Series C | https://tackle.io/company-news/tackle-series-c-funding/ | 2021-12 | issuer financing post | Preemptive round and scaling intent; not outcome evidence. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: financing/scaling paragraphs)
• ★★★ | AppDirect — agreement to acquire Tackle | https://www.appdirect.com/about/press/releases/appdirect-and-tackle-io-to-unite-to-extend-leadership-in-b2b-subscription-commerce-with-native-hyperscaler-marketplace-integration | 2025-12-01 | acquirer release | Confirms transaction and omits terms. (reliability: A; access: full; accessed: 2026-08-20; locator: opening transaction paragraphs)
• ★★★ | Cisco — Isovalent acquisition completed | https://investor.cisco.com/news/news-details/2024/Cisco-Completes-Acquisition-of-Isovalent-to-Define-the-Future-of-Multicloud-Networking-and-Security/default.aspx | 2024-04-12 | acquirer release | Completion/rationale; consideration undisclosed. (reliability: A; access: full; accessed: 2026-08-20; locator: opening paragraphs)
• ★★★ | Postman — Orbit acquisition | https://blog.postman.com/announcing-postman-has-acquired-orbit/ | 2024-04-11 | acquirer post | Acquisition and 90-day product sunset; sunset not proof of loss. (reliability: A; access: full; accessed: 2026-08-20; locator: acquisition and final sunset paragraphs)
• ★★★ | Convex — Series A | https://news.convex.dev/series-a/ | 2022-04 | issuer financing post | Distinguishes Casado's prior angel/adviser role from firm Series A and board. (reliability: A for issuer statement; access: full; accessed: 2026-08-20; locator: opening financing/relationship paragraphs)
• ★★ | Forbes — Astranis decision-date profile | https://www.forbes.com/sites/alexknapp/2018/03/01/andreessen-horowitz-makes-its-first-space-investment-with-satellite-startup-astranis/ | 2018-03-01 | contemporaneous reported profile | Financing, board, architecture and prototype control. (reliability: B; access: full; accessed: 2026-08-20; locator: opening financing/board and architecture/prototype passages)
• ★★★ | Astranis — Arcturus anomaly | https://www.astranis.com/blog/arcturus-satellite-anomaly-alaska-service-update | 2023-08-10 | issuer technical update | Mission-impacting anomaly, reproduced supplier failure and fix. (reliability: A for issuer account; access: full; accessed: 2026-08-20; locator: opening anomaly and reproduced-failure paragraphs)
• ★★★ | Astranis — Series E/debt | https://www.astranis.com/blog/series-e-raise-advanced-high-orbit-spacecraft | 2026-05-06 | issuer financing post | Equity/debt and operating claims; not value/return. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening capital and five-satellite/backlog paragraphs)
• ★★★ | Tecton — Series C | https://www.globenewswire.com/news-release/2022/7/12/2478423/0/en/Tecton-Reports-Record-Demand-for-Its-Machine-Learning-Feature-Platform-as-It-Raises-100-Million-in-Funding-Led-by-Kleiner-Perkins-With-Participation-from-Strategic-Investors-Databr.html | 2022-07-12 | issuer-distributed release | $100M round/syndicate; issuer metrics unaudited. (reliability: A/B; access: full; accessed: 2026-08-20; locator: “Tecton Raises $100 Million” section)
• ★★★ | Databricks — Tecton joining | https://www.databricks.com/blog/tecton-joining-databricks-power-real-time-data-personalized-ai-agents | 2025-08-22 | acquirer post | Strategic transaction; terms absent. (reliability: A; access: full; accessed: 2026-08-20; locator: summary/opening and “Tecton + Databricks”)
• ★★★ | Uniphore — ActionIQ acquisition overview | https://www.uniphore.com/actioniq/ | transaction announced 2024-12; live page accessed 2026-08-20 | primary acquirer page | Confirms ActionIQ combination/product integration; transaction terms and investor proceeds undisclosed. (reliability: A for acquirer claim; access: full; accessed: 2026-08-20; locator: acquisition/integration overview; mutable page)
• ★★★ | NVIDIA — DeepMap acquired-status update | https://blogs.nvidia.com/blog/isaac-amr-platform/ | 2022-01-04 | primary acquirer post | Describes NVIDIA's recent acquisition of DeepMap and its integration into NVIDIA DRIVE; terms and holder proceeds absent. (reliability: A; access: full; accessed: 2026-08-20; locator: “DeepMap Delivers Mapping Advances,” first paragraph)
• ★★★ | Stripe — Metronome acquisition completed | https://stripe.com/gb/newsroom/news/stripe-completes-metronome-acquisition | 2026-01-14 | primary acquirer release | Confirms completion and product rationale; financial terms and shareholder allocations undisclosed. (reliability: A; access: full; accessed: 2026-08-20; locator: opening completion and integration paragraphs)
• ★★★ | Intel — SigOpt acquisition | https://download.intel.com/newsroom/archive/2025/en-us-2020-10-29-intel-to-acquire-sigopt-to-scale-ai-productivity-and-performance.pdf | 2020-10-29 | primary acquirer release PDF | Confirms acquisition announcement and rationale; terms and investor proceeds absent. (reliability: A; access: full; accessed: 2026-08-20; locator: p. 1, heading and opening transaction paragraphs)
• ★★★ | Intel — SigOpt acquired-status update | https://community.intel.com/t5/Blogs/Tech-Innovation/Artificial-Intelligence-AI/Kisaco-Research-designates-SigOpt-as-a-Leader-in-AI-Software/post/1335688 | 2021-05-14 | official Intel employee-domain post | States that Intel acquired SigOpt in October 2020; terms and investor proceeds remain absent. (reliability: A/B; access: full; accessed: 2026-08-20; locator: opening acquired-status paragraph)
• ★★★ | Databricks — agreement to acquire Tabular | https://www.prnewswire.com/news-releases/databricks-agrees-to-acquire-tabular-the-company-founded-by-the-original-creators-of-apache-iceberg-302163561.html | 2024-06-04 | issuer-distributed transaction release | Confirms announced agreement and strategic rationale; completion, consideration, and holder proceeds are not established by this release. (reliability: A/B; access: full; accessed: 2026-08-20; locator: headline and opening transaction paragraphs)
• ★★★ | Nokia — Rapid technology-assets and R&D-unit acquisition | https://www.nokia.com/newsroom/nokia-acquires-rapid-technology-and-rd-unit-to-strengthen-development-of-network-api-solutions-and-ecosystem/ | 2024-11-13 | primary acquirer release | Confirms technology-asset and R&D-unit acquisition, not clearly a whole-company acquisition; terms and holder proceeds absent. (reliability: A; access: full; accessed: 2026-08-20; locator: opening and acquired-assets/R&D-unit paragraph)
• ★★ | Tackle — company status | https://tackle.io/company/ | live; accessed 2026-08-20 | mutable issuer page | Identifies Tackle as acquired by AppDirect; no transaction price or holder economics. (reliability: A/B; access: full; accessed: 2026-08-20; locator: company-status/acquisition text)
• ★★★ | Preset — Series C | https://preset.io/blog/preset-series-c/ | 2026-03-09 | issuer financing post | $7.27M C; amount alone cannot prove down round. (reliability: A; access: full; accessed: 2026-08-20; locator: opening financing paragraphs)
• ★★ | TechCrunch — RapidAPI workforce reduction | https://techcrunch.com/2023/04/25/rapidapi-valued-at-1-billion-last-year-cuts-staff-by-50/ | 2023-04-25 | adverse-event report | Company-confirmed ~50% cut/CEO change; no shutdown or return evidence. (reliability: B; access: full; accessed: 2026-08-20; locator: opening/company-confirmation paragraphs)
• ★ | Fly.io founder comments | https://news.ycombinator.com/item?id=41441218 | 2024-09-04 | direct participant on third-party forum | Workforce adjustment preceding Casado's disclosed board role. (reliability: C; access: full; accessed: 2026-08-20; locator: founder comments)
• ★★★ | Fly.io — Series D/board | https://fly.io/news/fly-io-launches-computers-for-agents/ | 2026-07 | issuer post | $25M D, current claims and Casado board timing. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: financing/board and metrics paragraphs)
• ★★ | TechCrunch — WaveForms acquisition report | https://techcrunch.com/2025/08/08/meta-acquires-ai-audio-startup-waveforms/ | 2025-08-08 | transaction report | Acquisition/undisclosed terms; primary confirmation missing. (reliability: B; access: full; accessed: 2026-08-20; locator: opening/terms paragraphs)
• ★★ | Fortune — Deeptune acquisition report | https://fortune.com/2026/07/09/ai-unicorn-mercor-acquires-deeptune-brendan-foody-investor-a16z-openai-anthropic/ | 2026-07-09 | transaction report | Acquisition/undisclosed terms; primary confirmation missing. (reliability: B; access: full/partial publisher; accessed: 2026-08-20; locator: opening/terms paragraphs)
• ★★★ | Resend — Sunsetting new.email | https://resend.com/blog/sunsetting-new-email/ | 2026-07-08 | issuer product postmortem | Separate-brand/stack costs and integration into parent; product lesson only. (reliability: A; access: full; accessed: 2026-08-20; locator: “Why” and separate-stack paragraphs)
• ★★★ | World Labs — $1B new funding | https://www.worldlabs.ai/blog/funding-2026 | 2026-02-18 | issuer financing announcement | Confirms $1B in new funding but does not disclose a valuation or identify a16z in the named investor list. (reliability: A for issuer claim; access: full; accessed: 2026-08-20; locator: opening funding paragraph)
• ★★★ | NVCA/PitchBook — Q4 2021 Venture Monitor | https://nvca.org/wp-content/uploads/2022/01/Q4_2021_PitchBook_NVCA_Venture_Monitor.pdf | 2022-01 | industry report PDF | 2021 financing/exit regime. (reliability: A/B; access: full; accessed: 2026-08-20; locator: Executive Summary $329.6B/17,054/$774.1B)
• ★★★ | NVCA — Q1 2022 ecosystem update | https://nvca.org/press_releases/u-s-venture-ecosystem-normalizes-after-record-year-in-2021/ | 2022 | industry release | Exit reset at $33.6B after three >$192B quarters. (reliability: A/B; access: full; accessed: 2026-08-20; locator: exit-market paragraph)
• ★★★ | NVCA/PitchBook — Q4 2025 Venture Monitor | https://nvca.org/wp-content/uploads/2026/01/q4-2025-pitchbook-nvca-venture-monitor.pdf | 2026-01 | industry report PDF | Fundraising, close-ratio and LP-liquidity context. (reliability: A/B; access: full; accessed: 2026-08-20; locator: Executive Summary/fundraising/liquidity)
• ★★★ | NVCA/PitchBook — Q1 2026 Venture Monitor | https://nvca.org/wp-content/uploads/2026/04/Q1-2026-PitchBook-NVCA-Venture-Monitor.pdf | 2026-04 | industry report PDF | AI deal-value/count concentration. (reliability: A/B; access: full; accessed: 2026-08-20; locator: Executive Summary AI concentration)

Completed-peer comparison controls

• ★★ | David H. Hsu and Martin Kenney — Organizing Venture Capital: The Rise and Demise of American Research & Development Corporation, 1946–1973 | https://kenney.faculty.ucdavis.edu/wp-content/uploads/sites/332/2018/03/Organizing-Venture-Capital_-The-Rise-and-Demise-of-American-Research-and-Development-Corporation.pdf | 2005 | peer-reviewed historical study | Completed-peer evidence for ARD sourcing, instrument mix, portfolio concentration, and the DEC/non-DEC return boundary; Casado comparison only. (reliability: B; access: full; accessed: 2026-08-20; locator: p. 594 and n. 11 for sourcing; Table 3 for instruments/portfolio construction; p. 599 for DEC and non-DEC return estimates)
• ★★ | Harvard Business School Baker Library — Financing New Ideas | https://www.library.hbs.edu/hc/doriot/innovation-vc/financing-new-ideas/ | 2014 exhibit | institutional archive exhibit with primary-source excerpts | Reproduces Doriot's selection and company-building language; completed-peer comparison only. (reliability: B; access: full; accessed: 2026-08-20; locator: “Financing New Ideas” exhibit panels and reproduced Doriot passages on patents/know-how, technical defensibility, construction, and speculation)
• ★★ | Tom Nicholas — The Origins of High-Tech Venture Investing in America | https://www.hbs.edu/ris/Publication%20Files/VC_Tech_4b7a1b2a-ef2a-4867-a41e-2ba84cb120f2.pdf | 2016 | scholarly chapter | Completed-peer evidence for ARD selection criteria, operating intervention, staged DEC financing, and Island Packers' adverse economics. (reliability: B; access: full; accessed: 2026-08-20; locator: printed pp. 230–231 for selection criteria; pp. 231–232 for intervention and Island Packers; p. 234 for staged DEC financing)
• ★★ | Aulikki Olsen — Aulikki Olsen Oral History | https://bitsavers.org/pdf/dec/history/Oral_Histories/Aulikki_Olsen_910305_910618.pdf | interviews 1991-03-05 and 1991-06-18 | approved oral-history transcript mirror | First-person completed-peer evidence on Doriot's demanding diligence, mentorship, family interviews, and founder-control tension. (reliability: B for archive mirror of retrospective testimony; access: full; accessed: 2026-08-20; locator: transcript p. 9 and surrounding diligence/mentorship discussion)
• ★★ | U.S. House of Representatives — Small Business Investment Incentive Act Hearings, Part 2 | https://demo.sechistorical.org/collection/papers/1970/1979_1107_HouseIncentive_02.pdf | 1979-11-07 | government-hearing record mirror | Completed-peer evidence on ARD's public-company constraints and affiliate/employee incentive problems. (reliability: B for archive mirror of a government record; access: full; accessed: 2026-08-20; locator: printed pp. 339–340, including p. 339 n. 12)
• ★★ | Reid W. Dennis — Reid Dennis: Pioneer Venture Capitalist | https://digicoll.lib.berkeley.edu/record/218448/files/dennis_reid.pdf#page=19 · https://digicoll.lib.berkeley.edu/record/218448/files/dennis_reid.pdf#page=33 · https://digicoll.lib.berkeley.edu/record/218448/files/dennis_reid.pdf#page=41 | interviews 2009-01-13 and 2009-02-10; published 2009 | interviewee-reviewed Bancroft Library oral-history transcript | Completed-peer evidence for inherited/personal access, referrals, syndication, portfolio philosophy, vehicle evolution, authority, and succession. (reliability: A for authenticated Dennis testimony and interview metadata; B for retrospective chronology, causality, and unaudited economics; access: full; accessed: 2026-08-20; locator: printed pp. 13–28/PDF pp. 19–34 for access, referrals, and The Group; printed pp. 27–28/PDF pp. 33–34 for portfolio philosophy; printed pp. 35–48/PDF pp. 41–54 for IVA/IVP vehicle evolution and succession)
• ★★ | Reid W. Dennis — Something Ventured Interview, Transcript 2 | https://stacks.stanford.edu/file/mj877gg1809/Reid_Dennis_Transcript_2.pdf#page=17 · https://stacks.stanford.edu/file/mj877gg1809/Reid_Dennis_Transcript_2.pdf#page=37 | recorded 2008-09-25; transcript file dated 2009-08-27 | raw Stanford/Smithsonian filmed-interview transcript | Completed-peer evidence for technical-to-customer-to-security underwriting, Reference Technology, held-to-zero positions, and the Apple price pass. (reliability: A for directly recorded testimony and repository metadata; B for retrospective claims; C for exact spellings/numbers in the unpolished transcript; access: full; accessed: 2026-08-20; locator: transcript 2 PDF pp. 17–18 for Reference Technology and held-to-zero discussion; pp. 37–42 for technical/customer/security sequencing and the Apple-price account)
• ★★ | Burton J. McMurtry — Early Bay Area Venture Capitalists Oral History | https://digicoll.lib.berkeley.edu/record/218493/files/mcmurtry_burt.pdf#page=40 | interviews 2009-06-09 and 2009-06-30; published 2009 | participant Bancroft Library oral-history transcript | Completed-peer evidence separating sourcing, disclosure/recusal, collective approval, and board work in the ROLM case. (reliability: A for authenticated participant testimony; B for retrospective accuracy; access: full; accessed: 2026-08-20; locator: printed pp. 30–32/PDF pp. 40–42)

Coverage assessment

Corpus, categories, and search variants

• Mapped corpus: 220 starred/priority source rows plus 16 supporting annotated route groups yield 236 annotated rows and 274 unique external URLs. They span official identity/archives; biography and public service; all 11 direct-corpus research works; Nicira/VMware records; 16 direct writings; 31 spoken appearances; the 43-record official deal-announcement corpus; vehicles, boards, outcomes, failures, market context, skeptical controls, and eight completed-peer comparison works. The direct-corpus sweep separately remains 60 unique works/routes: two official archives, 16 writings, 11 research papers, and 31 spoken appearances. Supplementary deal, profile, version, attribution, counterevidence, and peer-comparison routes do not add to that direct-work denominator.
• Names and date variants searched: Martin Casado, Martín Casado, Nicira, Illuminics/Illuminics Systems, Quova, VMware NSX, OpenFlow, Open vSwitch, SANE, Ethane, NOX, Onix, MuralNet, Andreessen Horowitz/a16z, Infrastructure, board/director, investment/led/Series/exit/acquired/IPO, podcast/interview/transcript/video, lawsuit/complaint/SEC/enforcement/criticism/layoff/shutdown/write-off. Chronology searches tested 1976 birthplace, 1998–2003 versus 2000–2006 LLNL tenure, 2005 versus 2007 MS, 2007 Nicira formation, 2009 a16z encounter, 2012 acquisition measures, 2014 versus 2015 NEC award, and 2016 a16z start.
• Source classes searched: official/personal/firm archives; university and laboratory pages; peer-reviewed papers and award indexes; SEC EDGAR and adviser databases; issuer/acquirer and portfolio-company announcements; congressional testimony; nonprofit records; publisher transcripts/audio/video; high-quality press; legislative/regulator records; venture benchmarks; practitioner criticism; legal/enforcement keyword searches. No credible person-specific lawsuit or enforcement record surfaced; that is a no-result, not proof of absence.

Direct voice, archives, and access boundaries

• Exhausted or mapped direct routes: the live a16z author/team archive, a16z podcast and video pages, the Stanford personal archive and publication index, LLNL histories, House and Senate statements, MuralNet, Software Engineering Daily transcripts, Invest Like the Best transcripts, direct company announcements, academic proceedings, and SEC filings were searched. Available full pages, PDFs, transcripts, audio, and video were opened; quotation use must bind versions and exact timestamps.
• Access barriers: the dynamic a16z investment list did not yield a stable complete person-linked export; the current Form ADV PDF was retrieved and mapped, while one Fund VII SEC XML route returned an automated-client 403 on recheck after earlier retrieval; the Wall Street Journal, Economist, and The Information routes remain paywalled/partial; several podcast pages expose full audio but only show notes. No blocked body was reconstructed from search snippets.
• Archival gaps: no original Illuminics–Quova transaction announcement, exact Nicira incorporation/first Form D, complete Nicira cap table, archived VMware title chronology, private a16z investment-committee record, LP communication, or Casado-specific cash-flow record was located. Historical snapshots of the a16z profile, portfolio, and investment-list databases remain the best public route to roster drift and omitted positions.

Candidate deal denominator, attribution, and economics

• Final candidate reconciliation: as of 2026-08-20, 43 announcement claims plus 25 current/former board claims, less 22 overlapping companies, yield 46 included companies. The three board-only companies are DeepMap, Netlify, and Pindrop. Adding two excluded founder/operator candidates—Nicira and Illuminics—produces 70 candidate claim records across 48 candidate companies. The board roster is therefore neither a complete investment denominator nor a person-role ledger; the exact profile and investment-list URLs are mapped once in their canonical source rows above.
• Inclusion boundary: Illuminics and Nicira are founder/operator outcomes, not investments selected by Casado. Firm investment, Infrastructure-team participation, Casado authorship, board service, Casado-led sponsorship, sole approval, vehicle, check, ownership, reserves, exit influence, company consideration, investor proceeds, and LP distributions are separate fields. Every silent role remains unknown.
• Economic boundary: the public record does not support a personal hit rate, MOIC, IRR, DPI, or complete loss count. Acquisition values and private marks are company outcomes; “acquired,” “IPO,” or public listing does not prove profitability or realization. The highest-value missing records are fund/check/ownership/dilution/preference/distribution data and a complete searched set of passes, shutdowns, down rounds, recaps, write-offs, secondaries, and below-cost exits.

Historical context, peers, and highest-value retrieval leads

• Historical context: the dated technical papers and independent SDN history establish the pre-Nicira architecture and collaboration set; contemporaneous 2012 reporting and SEC filings freeze financing, product, competition, and transaction measures; PitchBook-NVCA vintages supply later venture-market conditions; FTC, legislative, cloud-cost, policy, and benchmark records provide external tests of AI and infrastructure claims. Later data must never be leaked into an earlier decision record without a retrospective label.
• Completed peers: Trae Stephens is the closest completed modern operator/founder-investor comparison for attribution, conflicts, private marks, and missing returns. Georges Doriot and Reid Dennis provide completed long-duration vehicle and governance contrasts but are not technology-era peers. Marc Andreessen, Ben Horowitz, and Peter Thiel remain incomplete in this repository and therefore are people/institution sources, not completed-peer comparators.
• Highest-value next retrievals: export and archive the a16z investment database; recover historical Form ADV/IAPD snapshots; reconcile every Casado announcement and board date to a company-primary source; obtain official transcripts/captions for audio-only priority talks; legitimately retrieve the two paywalled guest essays; find original Illuminics/Quova and Nicira incorporation/financing instruments; search acquirer filings for purchase-price notes; audit every candidate company for adverse outcomes; and map fund, decision rights, reserves, ownership, seller liquidity, and LP cash flows. These are named unresolved evidence requirements, not grounds to infer a clean record today.

Source notes