Tom Brown

Who he is

Affiliation (role check run 2026-10-03 — v1's "[TBD: exact current title]" resolved): Co-founder and Chief Compute Officer, Anthropic — leads the organization "responsible for securing, scaling, and effectively using its compute resources" (title on Anthropic's leadership page as fetched for the McCandlish file 2026-09-21; Crunchbase; introduced as such by Commerce Secretary Howard Lutnick on stage at the G20 Innovation Summit, 2026-09-02). Age 39; MIT (MEng, computer science / brain and cognitive sciences — secondary sources). Career: first employee at a language-learning startup, founding engineer at MoPub (sold to Twitter), co-founder and CTO of Grouper (YC W12, a group-dating app) — then six months of self-study ("I got a B- in linear algebra") and monthly emails to Greg Brockman offering to "mop floors," which got him into OpenAI as one of its first ~20 employees (Business Insider via Yahoo, 2025-08-20). Co-author of the original RLHF paper (2017), first author of "Adversarial Patch" at Google Brain (2017), lead engineer and first author of GPT-3 (2020). Co-founded Anthropic in 2021. Bloomberg's index lists him at roughly $8B net worth on his Anthropic stake after the $965B round (Bloomberg profile, paywalled — figure via Business Insider, 2026-07-01); reported to have pledged 80% of it (same source; unverified against a pledge registry).
One-line position: Scaling works and is predictable; the frontier is an engineering problem; compute — and now the power and permits behind it — is "very clearly the bottleneck for all of our progress"; safety-motivated people should be the ones doing the scaling. New in 2026: the engineer became the company's negotiator — Lutnick's export-control letters went to Brown, not to Dario Amodei.

Discipline & technical bet

Infrastructure engineer turned compute strategist; the only persona on the bench whose job is literally buying and running the frontier. The bet has three legs. (1) Pretraining scale still pays: he told the G20 audience the scaling curve now holds "over 15 orders of magnitude" and that a model as smart as last year's best is "roughly 20 times cheaper to access" each year (2026-09-02). (2) Multi-vendor silicon: Anthropic trains and serves Claude on AWS Trainium, Google TPUs and NVIDIA GPUs so it can "match workloads to the chips best suited for them" (Anthropic, 2026-04-06); his own 2024 post on going "deep into the Neuron stack to get as close as possible to 100% peak theoretical performance" is the engineering version of the same bet (X, 2024-11-22). (3) Power, not chips, is the next wall: "the biggest thing by far is … making it possible to build more data centers, more compute in your countries" — permitting, labor, and a "ratepayer pledge" to bring new generation behind the meter and give it back to the grid (G20, 2026-09-02). Revealed posture: in eleven months Anthropic signed for roughly 5 GW of Trainium, multiple GW of TPUs from 2027, up to 1 GW of NVIDIA via Azure, ~325,000 GPUs at SpaceX's Colossus, $50B of its own data centers, and supply deals with all three HBM memory makers — a portfolio built so that no single vendor can throttle it.

Key claims (Says)

• Scaling laws are predictive, not retrospective — loss improves as a smooth power law in compute, data, and parameters; you can budget capability in advance (co-author, "Scaling Laws for Neural Language Models," 2020-01). Held — Epoch AI measures frontier training compute growing 4–5× per year through 2024 (Epoch) and Brown's own 2026 restatement extends the curve to "15 orders of magnitude" (G20, 2026-09-02). Watch item: Epoch projects growth easing toward 3–4×/yr from 2026 as power and capital bind — the first real test of "predictive."
• In-context learning emerges from scale — GPT-3's few-shot ability appeared without task-specific training ("Language Models are Few-Shot Learners," 2020-05). Held — unchallenged in the window; the 2026 frontier (Fable 5 / Mythos 5, June 2026) extends the same mechanism rather than replacing it (CNBC, 2026-06-30).
• Frontier AI is a systems-engineering problem — distributed training, cluster reliability, and infrastructure determine who can operate at the frontier. Held, and hardened — his words: building compute is "very clearly the bottleneck for all of our progress" (G20, 2026-09-02); his title is now Chief Compute Officer; and the reported S-1 puts Anthropic's binding infrastructure commitments near $518B (Reuters via Yahoo, 2026-09-29 — not yet on EDGAR; see Sources).
• Compute is the strategic chokepoint; hence multi-cloud. Held — and the chokepoint moved one layer down (power, permits) and one layer out (export law). Deals since v1: Amazon $100B+ over ten years for up to 5 GW of Trainium (2026-04-20); Google/Broadcom multi-GW TPUs from 2027 (2026-04-06); Microsoft/NVIDIA $30B Azure, up to 1 GW (2025-11-18); SpaceX Colossus (2026-05-06); $50B Fluidstack build (2025-11-12); Micron, Samsung and SK Hynix as HBM suppliers and investors (2026-06-22). Then the June export order showed the chokepoint can be a government letter: Anthropic's two best models went offline worldwide for 18 days (CSIS).
• Safety-motivated scaling — if scale is the path to powerful AI, safety-focused labs must hold the frontier rather than cede it (normative; Anthropic founding premise). Unfalsifiable as stated — rewritten: observable = does Anthropic withhold or gate a model at visible commercial cost? One data point in the window: Mythos was announced in April 2026 with restricted access on cyber/bio grounds, and Fable 5 shipped as the guardrailed version (Wikipedia timeline, secondary). Reads Held on that one observable; still indistinguishable from brand positioning from outside.
• Small elite teams suffice — frontier models can be trained by dozens of strong engineers, not thousands (v1 flagged "weak public articulation"). Wrong — by his own count: "Anthropic, we have like 4,000 people or something like that" (G20, 2026-09-02); Revelio Labs has 3,950 in March 2026, up from 546 in 2023 (Revelio); Tracxn 5,994 by August. A sevenfold headcount in three years is not the small-team model, and the compute org itself now spans silicon, power, and trade negotiation. No quote from him ever asserted the claim — see Delta log.
• Capability gains are not slowing on trend (v1: "[TBD: any direct public statement]"). Held — now stated in his own voice: models are "a once-in-a-generation genius on math … a top 100 current mathematician in many of the subfields"; same-intelligence cost falls ~20× a year (G20, 2026-09-02). Revealed: run-rate revenue ~$9–10B at end-2025 → $30B (April) → $47B+ (May) (Anthropic; TechCrunch, 2026-05-28). Rule-25 discount applies: he is describing his own product.
• Working on the problem directly beats earning-to-give — EA-style career reasoning (v1: "[TBD: primary source]"). Held — sourced: he describes switching "from engineering to try to help out learning about AI" (G20, 2026-09-02); the six-month self-study and the Brockman emails are on record (BI, 2025-08-20); the 80% wealth pledge is reported (BI, 2026-07-01) but the Giving What We Can listing cannot be confirmed as the same Tom Brown (longtermwiki caveat). Pledge registry is still the missing primary.
• NEW — Power and permitting are the policy ask: data centers "produce jobs," "can be taxed," and Anthropic's design "bring[s] on more power to the grid" via the ratepayer pledge (G20, 2026-09-02). Empirical on the grid claim; the observable is whether Anthropic-contracted sites add net generation or draw from existing supply — not yet checkable.
• NEW — Jobs: the Industrial Revolution analogy — "the Industrial Revolution didn't put everybody out of a job, and the invention of electricity didn't put everyone out of a job" (G20, 2026-09-02). Normative-by-analogy, no data; the same aggregate-vs-margin problem that graded Ng Wrong at the hiring margin applies, and Brown has offered less than Ng did.

Notable predictions — with falsifiable checks

• (2020-01) Scaling laws: loss follows a power law over many orders of magnitude. Held — vindicated for pretraining through the GPT-4 class (v1) and restated by Brown at 15 orders of magnitude (2026-09-02); inference-time scaling added a second axis the paper did not cover. Check: Epoch's frontier-compute growth estimate for 2027 — ≥3×/yr with capability-per-dollar still improving ~10× or better → Held; growth under 2×/yr or a flat frontier for 18 months → Drifted.
• (2020-05) Further scale yields broader few-shot competence. Held — Fable 5 / Mythos 5 (June 2026) and his own math-capability claim. Check: a top-3 lab publicly abandoning a larger pretraining run in favor of post-training only, before end-2027 → Drifted.
• (v1) "Few personally attributed dated public predictions exist." Drifted — true until 2025; false since. One transcript (G20, 2026-09-02) carries four dated calls, logged below. The bench's "thin record" was partly a search problem: his X handle is @nottombrown, and two podcasts (Lightcone 2025-08-21, Salesforce Ventures 2023-12-19) existed before the June 2026 review.
• NEW (2026-09-02) "Within the next 12 months … they could be a once-in-a-generation scientist for key scientific fields." Check by 2027-09-02: a peer-reviewed result in a named field where the model's contribution is credited as the discovery, not the assistant → Held; benchmark-only evidence → Drifted; nothing → Wrong.
• NEW (2026-09-02) Cancer: "I would be astonished if we haven't made material progress on that in five years." Check by 2031-09: an approved therapy or a Phase III result with a frontier model credited in discovery → Held. Soft wording ("material progress") — grade strictly or retire.
• NEW (2026-09-02) Human-level intelligence "not in 50 years, but on the order of under 10 years" — his stated shift from a within-our-lifetimes prior. Check: reuse Ng's Turing-AGI Test (multi-day work "as well as a skilled human," 2026-01-02); a widely accepted pass before 2036-09 → Held; none by then → Wrong. Converges with Kaplan's timeline and splits from Ng's "many decades."
• NEW (2026-09-02) "The doubling time of the overall economy … is like 4x per year. So it doubles every six months." Ambiguous as transcribed — the AI economy, not GDP. Observable: Anthropic run-rate 4× year over year ($47B May 2026 → ~$190B by May 2027) → Held for his firm; ≤2× → Wrong. Flag: the transcript is a third-party transcription; confirm wording against video before grading.
• (2023-12-19) "If you need to get 99.9%" reliability "it might take a year or something like that" (Salesforce Ventures). Unfalsifiable without a task named — retired.

Revealed behavior (Does)

• Bought the frontier five ways: Trainium (up to 5 GW, over one million Trainium2 chips in use, ~1 GW of Trainium2/3 by end-2026), TPUs (multiple GW from 2027), NVIDIA via Azure (up to 1 GW), NVIDIA via SpaceX, and his own sites via Fluidstack ($50B, Texas and New York). Secondary reports add TeraWulf, AMD (2 GW MI450) and Nscale in Jul–Aug 2026 (Wikipedia timeline, unverified here).
• The SpaceX contract, read from the filing (Aschenbrenner rule — primary beats press): SpaceX's S-1/A says the May 2026 Cloud Services Agreements cover "approximately 325,000 NVIDIA GPUs" across Colossus I and II, at $1.25B per month through May 2029, ramping at a reduced fee in May–June 2026 — and that "the agreement may be terminated by either party upon 90 days' notice" (SpaceX S-1/A, 2026-06-03, p. F-63). Press reported 220,000 GPUs and a $40B+ total; the filing shows a bigger pool and a 90-day exit. That exit clause is the multi-vendor thesis written into a contract.
• Became the negotiator: arrived at Commerce on 2026-06-15 with policy head Sarah Heck to open emergency talks on the 06-12 export directive; engaged Lutnick and National Cyber Director Sean Cairncross directly; Lutnick's letter lifting the controls was addressed to Brown, not Amodei (Wired/BI via AI Weekly; BI, 2026-07-01). Controls withdrawn 06-30, access restored 07-01 (CNBC). Ten weeks later Lutnick interviewed him on stage at the G20.
• Went on record for the first time at length: Lightcone podcast (YC, 2025-08-21 — "humanity's largest infrastructure buildout ever"), a Business Insider career piece (2025-08-20), the G20 stage (2026-09-02). Still no written essay, no letter, no blog in the window.
• Signed the ratepayer pledge — Anthropic's data-center design brings new generation behind the meter (G20, 2026-09-02). Text of the pledge not located.
• Kept the Trainium bet after it got hard: his 2024 post committed the team to the Neuron software stack; by April 2026 Anthropic was running over a million Trainium2 chips and signed for Trainium3 and 4 (Anthropic, 2026-04-20).
• Did not weigh in on the labor data, the AGI-timeline debate in writing, or Anthropic's Responsible Scaling Policy — compute and policy access are the whole public footprint.

Feels

Wants to be the person who makes the curve keep going — more chips, more power, closer to peak FLOPs — and wants it to be true that the people doing it care about safety. Fears the wall (power, permits, a government letter) more than the model; the June shutdown is the first time his bottleneck was a regulator. Underneath: the self-taught engineer who still frames his career as "trying to help out."

Hears

Dario and Daniela Amodei, Kaplan and McCandlish on what the compute should buy; AWS Annapurna, Google TPU and NVIDIA teams on what it costs; Lutnick, Cairncross and Commerce since June. Downstream of the Christiano/Amodei scaling-and-safety lineage he joined at OpenAI; upstream of nobody in public — he quotes figures, not thinkers.

Sees

The real bill of materials of the frontier: chip allocations, power contracts, utilization, and the terms of ~$500B of commitments nobody outside sees whole. Also the inside of a national-security negotiation over a model. Blind spots: anything outside the lab — labor markets, distribution, the open-weight ecosystem.

Incentive map

One vehicle, one claim set, maximal alignment. Anthropic equity (~$8B reported) rises with "capability is not slowing" and "compute is the bottleneck" — both are the pitch to the capital markets ahead of an IPO targeting ~$2T. The compute job pays him to say "build more data centers"; a Commerce Secretary interviewing him on stage pays Commerce to hear it. The pledge (80%, reported) changes where the money goes, not what he is paid to believe. Can't say: "we overbought," "a vendor has us," "the curve is bending," "the June order had a point." Rule-25 discount: high on capability-trend and compute-demand claims (he is the buyer and the beneficiary); moderate on power/permitting (industry-wide ask); lowest on the career-path material and the scaling-law history, which cost him nothing. Compared with Huang — same demand story, opposite side of the invoice.

Theories aligned with

• AI alignment / x-risk — moderate, infrastructure-operator variant; risk real enough to organize a career around, not enough to stop scaling (carried)
• Techno-optimism — conditional, safety-gated kind via Anthropic's institutional position (carried); the G20 cancer and "once-in-a-generation scientist" calls move him toward the unconditional end
• Jevons' paradox — adjacent in v1, now his own numbers: same-intelligence cost falls ~20×/yr while his commitments grow to ~5 GW and ~$500B. Cheaper compute, more compute — the paradox in one procurement ledger
• Scaling-as-engineering camp — with Kaplan, McCandlish, Sutskever (pre-SSI)

What he's reacting against

• The view that AGI progress requires fundamental scientific breakthroughs rather than disciplined engineering at scale — carried; restated as "the stupid thing that works" (secondary attribution)
• Architecture-first research culture — GPT-3 showed capability gains from scale with essentially no architectural novelty — carried
• Scaling skepticism (LeCun-style "LLMs are an off-ramp") — carried; his career is the empirical counter-bet
• The idea that safety-concerned researchers should stay away from capabilities work — carried; co-founded Anthropic on the opposite premise
• Credentialism in ML — carried, now sourced: "It seemed at the time that you needed to be top superstar to try to help" (BI, 2025-08-20); v1's "[TBD: verify narrative details]" resolved
• NEW: permitting and grid friction — "making it possible to build more data centers" is his first public policy position (G20, 2026-09-02)
• NEW: doomer framing of jobs — by historical analogy only (G20, 2026-09-02)
• NEW, by action not words: single-vendor dependence — five compute suppliers, three memory suppliers, a 90-day exit clause

Where he overlaps / splits (with Rich)

• Overlaps with Dario Amodei on the full Anthropic stack — carried; new: Brown now also fronts the government relationship when the dispute is technical, Amodei when it is political (G7 2026-06-17)
• Overlaps with Kaplan and McCandlish — the scaling trio; Brown's "under 10 years" (2026-09-02) now sits beside Kaplan's dated timeline — both on the bench, both quarterly
• Overlaps with Sutskever on scaling conviction; splits sharper than v1: Sutskever says scaling is over because generalization is missing; Brown says 15 orders of magnitude and counting
• Overlaps with Olah institutionally — complementary bets, carried
• Splits with LeCun — carried, the long and the short of autoregressive scaling
• Distinct from Noam Brown (no relation) — carried; searches on "Tom Brown" still contaminate both ways
• Splits with Toner on lab self-regulation — carried, sharpened: June 2026 is a case where the lab negotiated its own terms with the regulator and won in 18 days
• Splits with Ng — new: Ng calls training "the riskiest" layer and AGI "many decades away"; Brown is the training layer, says under ten years, and was the Anthropic executive in the room during the export dispute Ng wrote three letters against
• Overlaps with Huang on compute demand; splits on vendor — Brown is the hyperscalers' and NVIDIA's single largest reason to compete with each other
• Overlaps with Musk as counterparty, not ally — Anthropic is SpaceX's anchor compute customer (S-1/A, 2026-06-03) while Grok-5 trains next door at Colossus II
• Splits with Clark on role, not substance — Clark writes the policy case; Brown closed the deal
• For Rich's bench: use Brown as the demand-side index for the compute layer (Jensen's mirror) — every GW figure he signs is a revealed forecast; grade his capability claims against third-party evals and his jobs remarks against Brynjolfsson's payroll data, never against his analogies.

Track record

• Lead author, GPT-3 ("Language Models are Few-Shot Learners," 2020-05) — the result that made scaling the field's center of gravity (carried)
• Co-author, scaling-laws paper (2020-01) — framework vindicated for pretraining (carried)
• NEW: co-author, "Deep Reinforcement Learning from Human Preferences" (Christiano et al., 2017-06) — the original RLHF paper; v1 missed it; it is his strongest safety-lineage credential
• "Adversarial Patch" (Google Brain, 2017-12) — early safety-adjacent credibility (carried)
• Built Anthropic's compute capability from founding (2021) — carried; by 2026 the evidence is ~5 GW contracted, a $965B valuation, and run-rate revenue from ~$10B to $47B+ in five months
• Project Rainier: ~500,000 Trainium2 chips online in under a year (announced re:Invent 2024, operational Oct 2025, secondary); Anthropic now reports over one million Trainium2 in use (2026-04-20) — a bet on unproven silicon that paid
• NEW: the June 2026 export negotiation — two weeks, full restoration; the first time a frontier model was pulled by export order and the first time it came back
• Pattern: right on scaling for six years; every TBD in v1 resolved in his favor except the team-size inference, which was the bench's, not his

Empirical vs normative

• Empirical: scaling-law claims, in-context learning, engineering-bottleneck thesis, compute-and-power bottleneck, cost-per-intelligence falling ~20×/yr — well-evidenced, his strongest ground (carried, extended)
• Normative: safety-motivated scaling, EA-style career logic, "build more data centers," jobs-by-analogy — asserted via institutional choices and one stage interview more than argument (carried)
• Anthropic co-founder with ~$8B at stake ahead of an IPO — capability-trajectory and compute-demand claims need CONVENTIONS rule 25 discounting (carried, now quantified)

Weak spots / open questions

• Thin public record — carried, narrowing: two podcasts and one transcript; still no writing. Risk of projecting Anthropic's institutional voice onto the individual remains, and v1 did exactly that on team size
• Scaling laws and the 2024–2025 bend: now answered in his favor by assertion ("15 orders of magnitude") but not by data he has shown; Epoch's projected 3–4×/yr deceleration from 2026 is the open empirical question
• "Safety-motivated scaling" — still indistinguishable from competitive scaling from outside; Mythos gating is one data point, the June negotiation (lab argues for fewer restrictions) cuts the other way
• Silent on economics and distribution — carried; the Industrial Revolution analogy is his only labor statement
• Engineering-bottleneck thesis vs algorithmic efficiency — carried; his own "20× cheaper per year" figure is mostly algorithmic and he does not separate the two
• Does the small-team model hold at gigawatt scale? Answered: no — 4,000 people. New question: does a 5-GW, $500B-committed lab still have the option to slow down, or has the compute bill made "safety pause" a contractual impossibility? The 90-day SpaceX clause says some of it can be shed; the ~80% non-cancelable share reported in the S-1 says most of it cannot
• NEW: the G20 transcript is a third-party transcription with at least one garbled passage ("doubling time … 4x per year"); no video reviewed. Confirm before grading the economy call
• NEW: the 80% pledge and the ~$8B figure are both reported, neither primary

Rich's take

• (your synthesis here)

Delta log

2026-10-03 — v1→v2 migration + validation

• Grades: 8 Held / 1 Drifted / 1 Wrong (plus 1 normative claim rewritten with an observable, 4 new live predictions with dated checks, 1 vague prediction retired). The June 2026 v1 pass is treated as never-reviewed per the batch rule; it carried four "[TBD]" placeholders, all now resolved.
• Wrong: "Small elite teams suffice" — Brown: "we have like 4,000 people" (G20, 2026-09-02); 546 → 3,950 employees 2023–Mar 2026 (Revelio Labs). The claim was never his — v1 inferred it from the 2021 founding team and flagged "weak public articulation." Logged as a bench miss: an inferred belief graded as if stated.
• Drifted: "Few personally attributed dated predictions exist" — one transcript now carries four (G20, 2026-09-02); two podcasts predate the v1 review (Lightcone, 2025-08-21; Salesforce Ventures, 2023-12-19).
• Held: scaling laws predictive (Epoch; G20); in-context learning (CNBC, 2026-06-30); systems-engineering bottleneck (G20; Reuters via Yahoo, 2026-09-29); compute chokepoint / multi-cloud (Anthropic, 2026-04-20; 2026-04-06; 2025-11-18; SpaceX S-1/A, 2026-06-03); capability not slowing (G20; TechCrunch, 2026-05-28); direct-work career logic (BI via Yahoo, 2025-08-20; BI, 2026-07-01); 2020-01 power-law prediction; 2020-05 few-shot prediction.
• Most surprising delta: the compute engineer became Anthropic's lead negotiator with the U.S. government. Lutnick's export letters were addressed to Brown, not Amodei; he opened talks at Commerce on 2026-06-15 and the controls were withdrawn 06-30 (Wired/BI via AI Weekly; CSIS). Ten weeks later the same Secretary interviewed him on the G20 stage. v1 had him as an infrastructure operator with "unusually thin" public footprint.
• Primary-source check (Aschenbrenner rule): the SpaceX S-1/A on EDGAR shows ~325,000 GPUs across Colossus I and II, $1.25B/month through May 2029, terminable on 90 days' notice — press had 220,000 GPUs and no exit clause. EDGAR full-text search run 2026-10-03 for "Anthropic, PBC" in S-1/DRS forms returns only SpaceX and a SPAC: no Anthropic S-1 is public; the $518B commitment figure is Reuters-reported and ungraded until it is.
• Role check: title confirmed Chief Compute Officer (leadership page via McCandlish file 2026-09-21; Crunchbase; G20 introduction). Added: age 39, MIT, MoPub, Grouper (YC W12), OpenAI first ~20, RLHF co-author, ~$8B reported stake, 80% pledge reported. Google Brain dates conflict across secondary sources (2017–18 sabbatical vs 2016–19) — left approximate.
• Cross-grades: Ng–Brown split added (Ng's three anti-Anthropic letters vs Brown in the room); Sutskever split sharpened; Musk logged as counterparty; Kaplan timeline convergence noted; Jevons moved from "adjacent" to "his own numbers."
• Tier: quarterly — an IPO that would publish his holdings and the compute ledger is targeted for November 2026; compute deals landed roughly monthly in the window; Epoch's deceleration check falls inside a year. Confidence stays medium: the record is on the record now, but it is two transcripts and no writing, and the one Wrong grade was the bench projecting onto him.

Sources

• Primary — his words: G20 Innovation Summit, interviewed by Howard Lutnick (2026-09-02; third-party transcript published 09-04) · Lightcone Podcast, YC (2025-08-21) and YC's post · Business Insider career piece via Yahoo (2025-08-20) · Salesforce Ventures conversation (2023-12-19) · X — Trainium/Neuron (2024-11-22) · @nottombrown
• Primary — filings and company announcements: SpaceX S-1/A (EDGAR, 2026-06-03) — Anthropic Cloud Services Agreements · Anthropic–Amazon (2026-04-20) · Anthropic–Google/Broadcom (2026-04-06) · Anthropic–SpaceX (2026-05-06) · Microsoft/NVIDIA (2025-11-18) · $50B infrastructure (2025-11-12)
• Papers: Language Models are Few-Shot Learners (2020-05, first author) · Scaling Laws for Neural Language Models (2020-01, co-author) · Deep RL from Human Preferences (2017-06, co-author) · Adversarial Patch (2017-12, first author)
• Export-control episode: Forbes (2026-06-16) · CSIS · CNBC (2026-06-30) · AI Weekly on Wired/BI reporting · Business Insider, "The engineer who became Anthropic's political secret weapon" (2026-07-01, syndicated copy)
• Market-level data: Epoch — frontier compute 4–5×/yr · Epoch — frontier labs' compute share (2026-05-20): Anthropic ~1M+ H100e, ~1.4 GW, 5–6 GW forecast end-2026 · Revelio Labs headcount · TechCrunch — $65B at $965B (2026-05-28) · CNBC · Blocks & Files — Micron (2026-06-23)
• IPO reporting (unverified against EDGAR as of 2026-10-03): Reuters via Yahoo — $518B commitments, Class F founder shares (2026-09-29) · "S-1 expected after Labor Day still hasn't arrived" · EDGAR full-text search, forms S-1/DRS, "Anthropic, PBC", 2026 — no Anthropic filer
• Secondary bios (used for dates only, flagged where they conflict): Crunchbase · Bloomberg Billionaires (paywalled) · frontierminds.ai · FounderFiles · longtermwiki (GWWC caveat) · Wikipedia — Anthropic timeline · compute-deal timeline compilation
• Carried from v1: papers above; Anthropic compute posts; the caveat that the persona leaned on papers and institutional behavior more than statements — now less true than it was

migrated v1→v2 2026-10-03 · backup: _archives/tom-brown.html.bak-20261003 · converts-from: people/tom-brown.md · Singularity 🧠 / People