Eric Schmidt

Who he is

Affiliation (hat count run 2026-09-21 — eleven hats, up from five in v1): Chair & CEO, Relativity Space (since 2025-03-10; took a controlling stake in the 3D-printed-rocket company, replacing founder Tim Ellis — his first CEO job since Google; reported ~$800M of his own money, ~$3B of Terran R launch contracts on the books, maiden flight "late 2026" with industry expecting 2027; NASA picked Relativity on 2026-06-18 to build and fly the Aeolus Mars orbiter for ~2028). CEO, Swift Beat LLC (the drone company v1 knew as White Stork, 2023–; renamed via "Project Eagle"; signed a co-production memorandum with Ukraine's Ministry of Defense on 2025-07-03 in Denmark with Zelensky present — priority supply "at cost price," "hundreds of thousands" of drones planned). Chair, Special Competitive Studies Project (SCSP, 2021–; keynoted its AI+ Expo 2026-05-06/09, ~20,000 attendees). Co-founder, Schmidt Sciences (with Wendy Schmidt; 2026 AI programs below). Co-founder, Hologen (medical AI, "large medicine models"; raising $150M Series A at ~$850M, STAT 2026-01-28). Board chair, SandboxAQ (AI/quantum; $450K of federal lobbying since 2023 per Revolving Door Project). Hillspire family office (22 private AI companies since 2019, $5B+ of rounds — including Anthropic, Holistic AI and SandboxAQ; 2026 adds General Intuition's $320M Series A; Gradium's seed was 2025-12, extended 2026-07). Innovation Endeavors (VC; Rebellion Defense). Political donor: $3M+ to the committee fighting California's Prop 40 billionaire tax (Fortune, 2026-08-21). Historical: CEO Google 2001–2011, Executive Chairman Alphabet 2011–2017; Chair NSCAI 2018–2021; founder Schmidt Futures. Author: The Age of AI (2021), Genesis (2024-11, with Kissinger — who died 2023-11-29 — and Craig Mundie), Superintelligence Strategy (2025-03, with Dan Hendrycks and Alexandr Wang).
One-line position (updated): AI is primarily a great-power-competition problem — but as of March 2025 he says the winning move is not a race to a superweapon: deter rivals from destabilizing AI projects (MAIM — mutual sabotage, the AI version of nuclear deterrence), keep weaponizable AI away from rogue actors, and out-deploy China in the real economy and on the battlefield. He now personally builds the rockets, drones and medical AI that thesis calls for. v1's line ("out-build and out-secure China through compute, energy, export controls and tighter government-industry coupling") is the 2021–24 version; the 2025–26 version keeps the compute-and-energy half and softens the race-and-controls half.

Discipline & technical bet

A non-technical thinker (in the sense the template means — he ran Google, he did not build the models) who has taken on three concrete, capital-heavy technical wagers since v1's last honest review. (1) Mass-produced autonomous drones with onboard vision — Swift Beat's Bumblebee interceptor (claims 70%+ autonomous hit rate), the Hornet fixed-wing striker (11-lb payload, 6,000+/month planned), $400–500 kamikaze quadcopters; all navigate by camera in GPS-jammed, comms-denied airspace, which is the whole bet: cheap mass beats exquisite few (Noema, 2026-06-16; DroneXL, 2026-09-11). (2) Reusable heavy-lift launch to put solar-powered data centers in orbit — the stated reason he bought Relativity (TechSpot / Slashdot, 2025-05), and the operational answer to his own claim that Earth's grid cannot carry AI. Terran R (33,500 kg to LEO) has to fly first; Relativity's only prior rocket never reached orbit. (3) "Large medicine models" at Hologen — AI that models between-patient variation to "de-risk billion-dollar clinical trials." On the underlying science he is still a straight scaling-law believer — "if you put more data and more electricity and more chips, you get this emergent behavior… it's happening very quickly" (Harvard Kennedy School, 2025-12-01) — which puts him on the opposite side of Sutskever's "age of scaling is over" turn. The pattern to watch: each of his newest theses (AI eats the grid; drones decide wars; the US must out-deploy China) now has a company attached.

Key claims (Says)

• Empirical: PRC is ~2 years behind US frontier capability as of ~2024–2025, closing rate sensitive to chip export controls (v1 carried this as [TBD: cite]; the primary is his own Jan-2024 "two or three years ahead" framing, per TechCrunch's recap). Wrong — retired by its own author. Market-level data: Chinese models have trailed the US frontier by 7 months on average since 2023, range 4–14 (Epoch AI); Stanford's 2026 AI Index puts the top-model gap at 2.7% as of March 2026 (per secondary summaries; primary not fetched); Chinese open-weight models sat ~4.4 months behind US closed frontier models in September 2026 (AndroidHeadlines, 2026-09). He got there first: DeepSeek is "a turning point" (WaPo op-ed, 2025-01-28); "we're no longer in the era when China is far behind us" (NYT with Selina Xu, 2025-05-05). Graded on the 2024 statement, per the signature-framing rule; the revision is logged, not blended.
• Empirical: AI compute demand will require nation-scale energy build-out — multi-GW campuses, new transmission, accelerated nuclear (SMRs + restarts). Held (mechanism). His House testimony (2025-04-09) put numbers on it: "an additional 29 gigawatts of power by 2027 and 67 more gigawatts by 2030," 10-GW campuses being planned, and the line that travelled — demand "will go from 3% to 99% of total generation." Big-four hyperscaler capex went $410B (2025) → ~$690–725B planned for 2026 (CNBC, 2026-02-06); ~10 GW of nuclear-for-data-center deals are signed; DOE lent Constellation $1B to restart Three Mile Island's Crane unit (Utility Dive, 2025-11). Valence shifted: by 2026-03-06 he was writing that tech should co-locate its own generation so households stop paying ($11.6B of approved rate increases, up to 25% in data-center regions; only 5 of 16 expected GW for 2026 under construction — Fortune). "Energy in all forms, quickly" became "energy, but not on the ratepayer's bill." Note the Relativity pairing: "Earth can't power AI" is also the sales pitch for orbital data centers.
• Empirical: agentic systems and recursive AI-improving-AI compress capability timelines materially within ~2–5 years of 2024. Drifted — the agent half held (below); the recursion half he reset himself: "Recursive self improvement is when it's learning on its own. This is not true today" — now "within 4 years" (Harvard Crimson, 2025-12-02). Outside evidence agrees: Princeton's shadow-evaluation of Claude Opus 4.8 on unpublished NeurIPS-2026 problems found agents "capable of all the engineering" and "unambiguously bad at carrying out the research itself" (MIT Technology Review, 2026-08-18).
• Empirical: open-weights frontier models (models whose trained parameters are published for anyone to download), once released, are immediately available to PRC, Russia and non-state actors — irreversible diffusion. Wrong on valence, Held on mechanism — graded separately per the Andreessen rule. v1 attributed this with no citation and cast it as his reason to oppose open weights. His dated record runs the other way: the DeepSeek op-ed's first recommendation was that America "develop more open source AI models" (WaPo, 2025-01-28); "I don't see the open-source leadership in America, and I see nothing but open-source leadership in China" (Harvard, 2025-12-01); and on Moonshots (Nov 2025) the fear is that "the vast majority of governments… will end up standardizing on Chinese models not because they're better, but because they're free" (TechTimes, 2025-11-12). The mechanism — free weights diffuse globally and irreversibly — is exactly what he now says, with Qwen downloads passing Llama as the datapoint; but he wants the US to be the one diffusing. The narrower Superintelligence Strategy nonproliferation pillar (secure weights of weaponizable-capability models) survives; the blanket anti-open-weights stance v1 gave him does not.
• Normative: US should expand chip export controls, restrict outbound AI investment, and tie cloud-compute access to know-your-customer regimes (KYC — verify who is renting the chips). Drifted. Chip nonproliferation stays in the March-2025 paper, but the May-2025 NYT essay argues the controls "only fuelled China's determination" to build its own supply chain, and no Schmidt statement on the August-2025 H20 export reversal was found in a dated search (searched 2026-09-21; absence noted, not asserted as position). Endorsement softened from "expand" to "keep, with doubts."
• Normative: government should fund and partially direct frontier compute and fundamental research (Schmidt Sciences as a private prototype of this thesis). Held (his side of it) — Schmidt Sciences in 2026 alone: $3M+ to 19 labor-economics studies of AI at work (2026-01), a $1–5M+ "Science of Trustworthy AI" RFP (deadline 2026-05-17), a multi-agent-safety program (deadline 2026-08-08); his House testimony asked for a Technology Competitiveness Council. The state's side is thin: federal nondefense AI R&D was $2.98B enacted for FY2024 and $3.32B requested for FY2025 (NITRD) against NSCAI's ~$8B/year ask; NAIRR is still formally a pilot with an operations-center solicitation that closed 2026-02-04.
• Normative: democracies need a coordinated AI bloc — coalition diplomacy as well as competition. Unfalsifiable — rewrite or retire. No 2025–26 Schmidt statement restating it was found; the closest in-window evidence is procedural (SCSP's 2026 Expo hosted "The China Challenge: The View from Taiwan"). Proposed rewrite: "by end-2027 an allied compute-sharing or model-access agreement exists that Schmidt or SCSP publicly advocated for." Until then, not graded.
• Mixed: existential / catastrophic AI risk is real but secondary to the geopolitical risk of authoritarian-led AI — the Kissinger framing of AI as a new "epistemology" of state power. Drifted — the hierarchy flattened. Superintelligence Strategy (arXiv 2503.05628, 2025-03-07) treats "capable AI hackers and virologists" in rogue hands and loss of control as first-order, not downstream; at Harvard he said "somebody's going to have to raise their hand and say, 'We just went too far'"; his foundation now funds alignment directly. Still not a doomer — but "secondary" no longer describes the ordering.
• Empirical / contested: AI-driven productivity gains will be large but distributed in ways that favor capital and high-skill knowledge work — labor displacement is real but not his primary lens. Drifted — the lens moved to the front, and he contradicted himself on the way. Riyadh, 2025-10-30: predictions of a billion job losses are "almost certainly false… for every job loss, there's probably more than one job created" (The National). Then 2026-07-13: he signed the 200-signatory "We Must Act Now" statement warning of "large-scale job displacement" and published an NYT essay ("Our Century's Biggest Challenge") the same week, citing Gallup's 80% who want regulation even if it slows progress (SB Independent, 2026-07-14). Market-level data that moved him is the same Canaries payroll evidence on this bench: early-career workers in AI-exposed jobs −13% relative since late 2022, developers aged 22–25 down ~20% from peak.
• Normative: a "decisive period" of 2–5 years (from ~2024) sets the long-run balance of power; choices made now are durable. Drifted — he keeps issuing short clocks (AGI 3–5 years; recursive self-improvement 4 years) but the March-2025 paper replaced "win the decisive window" with a deterrence equilibrium in which no state can safely sprint — a stable-balance frame, not a lock-in frame. Check: if by end-2029 the US–China frontier gap (Epoch measure) is within ±6 months of today's, the durable-lock-in claim is Wrong; if it has widened past 18 months in either direction, Held.

Notable predictions — with falsifiable checks

• (2024-08, Stanford ECON 295) Within ~1 year, AI agents will "do real work," not just chat. Held (mechanism; "real work" still partial) — Gartner's 2026 CIO survey: 17% of organizations have agents deployed, 80% of enterprise applications shipped in Q1-2026 embed one (up from 33% in 2024). Valence caveat: only ~23% of organizations are scaling an agentic system (McKinsey). v1's "[TBD: verify against benchmark]" is closed with the Gartner/McKinsey numbers.
• (2024-08) GPU/compute shortage will dominate the next several years; ~$300B+ in data-center capex. Held — and undershot by 2x: $410B in 2025, ~$725B planned for 2026 across Amazon/Google/Meta/Microsoft (CNBC, 2026-02-06). Check for 2027: does combined capex fall below 2026's level? If so the shortage claim ends.
• (2024) PRC ~2 years behind US frontier; gap narrows if chip controls fail. Wrong — graded under Key claims; the gap was 4–14 months on Epoch's series and he abandoned the figure in January 2025.
• (2021, NSCAI final report) US must invest ~$40B over 5 years in AI R&D; create a National AI Research Resource. Drifted (partially adopted) — NAIRR pilot launched 2024, 600+ research teams served, transition to a permanent operations center solicited (NSF 25-546, closed 2026-02-04) but still a pilot in September 2026; nondefense AI R&D ran ~$3.0–3.3B/yr against the ~$8B/yr implied ask. Check: NAIRR-OC award announced and funded by end-2026.
• (2021, The Age of AI) AI will reshape diplomacy, warfare, and epistemology within a decade. Warfare branch: Held — his own figure is that drone units caused 96% of Russian casualties in March 2026 (Noema, 2026-06-16), Russia produces 404+ Shaheds a day aiming for 1,000 (WaPo, 2026-09-09), and he is a supplier in that war. Diplomacy and epistemology branches: unfalsifiable as written — per-branch rewrite: "by 2031, a treaty or formal deterrence arrangement whose subject is AI capability exists between the US and China" (diplomacy); "by 2031, a national election's official result is contested primarily on grounds of AI-generated evidence" (epistemology). Not graded until then.
• (2024-08, Stanford) Within "a year or two" recursive self-improvement of foundation models becomes the central capability driver. Wrong (author-reset) — the window closed in August 2026; in December 2025 he said it "is not true today" and moved the date to four years; the Princeton/MIT-TR evidence (2026-08-18) says agents cannot yet do the research part. Successor check below.
• (2023–2024) Nuclear restarts and SMRs become a dominant new-power source for AI data centers. Drifted — deals yes, electrons no: 13 named nuclear-for-data-center deals totaling ~10 GW of commitments; Crane (Three Mile Island) restart now targeted for H2-2027 with an NRC licence decision expected May 2027 (WITF, 2026-07-28); no SMR has delivered a megawatt to a data center by September 2026, and he told Congress SMR approval takes "12 years." Check: first SMR-generated MW delivered to a data center by end-2030; "dominant" needs nuclear ≥25% of new data-center supply additions by 2032.
• NEW (2025-04-10, AI+Biotech Summit, DC) "In the next one year, the vast majority of programmers will be replaced by AI programmers." Wrong — the year ended 2026-04-10. Market-level data: BLS projects software-developer employment +15.8% for 2024–34 and "programmer" titles down only 6%; Indeed software-engineering postings sit ~49% below February 2020 but employment did not collapse; the real signal is at the entry margin (22–25-year-old developers −20% from peak, Stanford Canaries). Mechanism partly right at the junior end, headline claim wrong. Cross-grade: this is a win for Frey's displacement-at-the-margin thesis, not for "vast majority replaced."
• NEW (2025-04, "San Francisco consensus") AGI — a single model rivalling the best mathematician, physicist, writer or artist — in 3–5 years (2028–2030); superintelligence by ~2031, defined as "it can prove something that we know to be true, but we cannot understand the proof." Live. Check: by end-2028 one model is credited with an original result in each of two of those fields that domain experts rank in the top human tier; by end-2031 a proof accepted as true that no human team can follow. His own definition makes this one of the few gradable superintelligence claims on the bench.
• NEW (2025-12-01, Harvard) AI mathematicians "next year" (2026); recursive self-improvement within four years (by ~end-2029). Live. Check: by end-2026 an AI-originated theorem or conjecture published and accepted by the mathematical community without a human co-discoverer; by end-2029 a frontier lab attributes a capability jump primarily to AI-conducted research. Correction factor: his 2024 version of this clock ran at least 2x fast.
• NEW (2025-11, Moonshots) Most governments and countries will standardize on Chinese open-source models because they are free. Live. Leading indicator already in: Qwen ~385M Hugging Face downloads vs Llama ~346M (Bloomberg data via Digitimes, 2025-11-17). Check: by end-2027, a count of national-government or sovereign-AI deployments shows Chinese open weights in the majority; if a US open-weight family regains the download lead, mechanism holds but his "no US leadership" premise falls.
• NEW (2025-04-09, House testimony) Data centers require +29 GW by 2027 and +67 GW by 2030. Live. Check: LBNL/EIA data-center load at end-2027 vs a ~2024 baseline; his own March-2026 piece already conceded only 5 of 16 expected 2026 GW were under construction, so the risk is a demand forecast the supply cannot meet — grade demand and supply separately.
• NEW (2026-09-09, WaPo op-ed) Russia is erasing Ukraine's battlefield edge — jet-powered Geran-3s outrun interceptors, a Rassvet satellite network replaces Starlink, and mass wins over sophistication ("hundreds of thousands" of good-enough systems beat hundreds of exquisite ones). Live, and self-serving: he sells the interceptors the argument calls for, which the op-ed did not disclose (DroneXL, 2026-09-11). Check: Ukrainian interceptor kill rate against jet-powered Shaheds reported by mid-2027; Swift Beat's Hornet output vs the 6,000/month plan.
• NEW (2025-10-30, Riyadh) AI will create more jobs than it destroys. Live — and contradicted by his own July-2026 signature. Per the Musk rule, this pattern gets a forward check rather than a narrative: by end-2028, payroll employment in the top-quartile AI-exposed occupations (Canaries methodology) is above its 2025 level (Held) or below (Wrong); whichever way it lands, log which of his two 2025–26 statements he repeats next.

Revealed behavior (Does)

• Went back to operating, three times over, after 14 years of advising. Bought control of Relativity and took the CEO chair (2025-03-10), signed a wartime co-production deal for Swift Beat with a head of state in the room (2025-07-03), co-founded and is raising for Hologen (2026-01). Time and money both say "builder," whatever the op-eds say about strategy. The May-2026 v1 review recorded none of it.
• Put his own money where the Superintelligence Strategy pillars are. Competitiveness: rockets, drones, medical AI, 22 AI startups incl. Anthropic. Nonproliferation/safety: Schmidt Sciences' 2026 trustworthy-AI and multi-agent-safety RFPs. Deterrence: nothing personally investable, and correspondingly the pillar he talks about least since March 2025.
• Writes the policy column for the market he sells into. WaPo drone op-ed (2026-09-09) with no Swift Beat disclosure; NBC interview (2026-09-14) on the same theme; Fortune energy piece (2026-03-06) from the owner of a would-be orbital-data-center launcher. Not deception — but the byline and the business now overlap on every subject he writes about.
• Spent $3M+ against a 5% one-time billionaire wealth tax (Prop 40) and, per Fortune, is at odds with Sergey Brin over how the coalition's $100M+ is spent; an adviser said it is his "only current involvement in politics" (2026-08-21). A direct personal-balance-sheet position, logged as such.
• Kept the DC machine running: SCSP's AI+ Expo doubled to ~20,000 attendees; his 2026 keynote ("Delivering AI Advantage," with Thom Shanker) was staged with the Chairman of the Joint Chiefs. Content of the keynote not retrievable in text at review time — noted as a gap, not filled.
• Declined to confirm his own product's numbers — asked about the reported 90% Shahed-kill share, he "would rather not discuss specific wartime details" (hvylya, 2026-09-09). Treat all Swift Beat performance figures as unverified third-party claims.

Feels

The operator's itch, finally scratched: he spent 2011–2024 as the man who advised everyone and ran nothing, and 2025 reads like a decision never to be that again. The fear underneath is the one he keeps naming about America — the slow incumbent that loses to the hungrier newcomer, which is the story of Google and OpenAI, which is his story. He wants it to be true that the person who ran the largest information company can also run a war supply chain and a rocket program at 71.

Hears

Selina Xu (his China lens and NYT co-author), Dan Hendrycks and Alexandr Wang (the MAIM paper), SCSP staff and the Ylli Bajraktari NSCAI network, Ukrainian commanders and Rustem Umerov (defence minister at the July-2025 signing), Kissinger's ghost through Genesis, the Moonshots crowd (Diamandis/Blundin — where the Chinese-open-source claim was made), and Hillspire portfolio founders including Anthropic's. A diet heavy on national-security people and light on economists — which is why the July-2026 labor turn arrived through a signed letter rather than his own analysis.

Sees

Battlefield drone telemetry from a live war that no other bench member has — kill rates, Russian adaptation cycles, what a $400 quadcopter does to a Patriot's economics. Plus the cap tables: Anthropic from inside, SandboxAQ and Relativity from the chair, 22 AI startups' board decks, and a Washington policy network he built and funds. The blind spot is the same as the vantage: he sees the war from the supplier's seat and the AI economy from the investor's.

Incentive map

Re-derived this cycle: maximal, and structurally different from v1. v1 flagged "significant investments in defense-AI and frontier compute" as a CONVENTIONS rule-25 conflict. It is now an operating conflict on every axis he speaks about. Who pays him / what he sells: launch (Relativity — needs Terran R to fly, NASA and NSSL contracts, and ~$800M of his own capital returned), interceptor and strike drones (Swift Beat — Ukraine MoD, at-cost pricing that presumably needs volume or follow-on Western buyers to pay), medical AI (Hologen — a $150M round in progress), quantum/AI (SandboxAQ — federal lobbying, ongoing litigation alleging inflated investor figures per Revolving Door Project, 2026-08-17), and an Anthropic position that gains when closed US frontier labs win. What he can't say: "Ukraine can't be saved by drones," "Terran R will slip," "AI power demand is overstated" (kills the orbital-data-center thesis), "Chinese open models are fine to standardize on," "a wealth tax is affordable." Pairings to log (Ng rule — newest vehicle matches newest thesis): the 99%-electricity testimony (2025-04) and the Relativity purchase (2025-03) are one month apart; the "mass beats sophistication" drone doctrine (2026-09) is Swift Beat's product spec; "Silicon Valley should stop obsessing over superhuman AI and deploy in the real economy" (NYT, 2025-08) is the portfolio thesis of a man whose 2025–26 checks went to robotics-training data, voice AI and clinical-trial AI. Independence of sources: the Moonshots podcast is a bench-internal channel (Diamandis, Blundin) — his Chinese-open-source claim and their coverage of it are not independent confirmations. Discount: maximal on drones, energy and launch timelines; high on China-gap and open-source claims (they now serve a "fund US open models" agenda he invests around); moderate on the historical scientific-strategy record; low on the March-2025 MAIM argument, which cuts against his own defense-vendor interest and is therefore the most credible thing he has said in-window.

Theories aligned with

• AI alignment / x-risk — national-security variant; v1 placed him "closer to Aschenbrenner than to Yudkowsky or Bengio." Revised: since Superintelligence Strategy he is the most prominent public opponent of the Manhattan-Project frame Aschenbrenner popularized — "what begins as a push for a superweapon and global control risks prompting hostile countermeasures" (TechCrunch, 2025-03-05). Still a hawk, now a deterrence hawk rather than a race hawk.
• Techno-optimism — conditional and US-state-coupled, not market-libertarian; closer to a Kissingerian "responsible great-power technology" frame. Unchanged, with a 2026 labor asterisk.
• Adjacent to a not-yet-named "national-AI-project" position [[national-security-ai-theory]] also implied by Aschenbrenner — now the split inside that position (deterrence vs sprint) is the theory's live fault line, and Schmidt named it.
• NEW: "application-first China" — his Aug-2025 NYT argument (America builds the most powerful systems, China builds the more powerful economy) is Kai-Fu Lee's AI Superpowers thesis arriving in Washington seven years late.

What he's reacting against

• Treating AI as purely a commercial / consumer-tech story — argues the national-security stakes dominate
• Light-touch governance frames that assume US labs and PRC firms operate in symmetric regulatory worlds
• Open-weights releases of frontier models — v1 said he treats them as proliferation risk, not democratisation. Corrected this cycle: what he is reacting against in 2025–26 is the absence of US open-weight leadership and the prospect of the world defaulting to free Chinese models; the residual worry is weight security for weaponizable capability, not openness as such
• Slow / decoupled US industrial policy — power, chips, fabs, and talent moving too slowly relative to PRC capacity build (now with a ratepayer twist: don't put it on households' bills)
• Lab self-regulation as a sufficient safety regime — supports labs but argues only USG can underwrite security and infrastructure at scale
• NEW (2025-03): a US "Manhattan Project for AGI" — the very policy the congressional US-China commission and the Energy Secretary floated; he argues it invites sabotage and escalation
• NEW (2025-08): Silicon Valley's superintelligence obsession as a strategy — it distracts from deploying existing AI "from manufacturing and agriculture to robotics and drones," where China is winning
• NEW (2026-09): exquisite-but-few Western weapons procurement — Lockheed's 600 Patriots a year against 1,000 Shaheds a day is the arithmetic he keeps repeating

Where he overlaps / splits (with Rich)

• Overlaps with Aschenbrenner on US-vs-PRC framing, weight-security worries, and aggressive timelines; splits — v1 said on the form of state involvement (public-private vs classified Manhattan program). Now a head-on split: Schmidt's MAIM paper argues the Manhattan program is the destabilizing move; both cannot be right about what a sprint does to the other side
• Overlaps with Altman on industrial-policy framing (compute as strategic) and on US leadership — he cited Stargate approvingly in the DeepSeek op-ed; splits on tone and on how much existential-risk discourse should drive policy, and now on the superintelligence-first strategy itself (NYT, 2025-08)
• Overlaps with Amodei on democracies-must-lead — and is an Anthropic investor via Hillspire, which v1 did not record; splits on whether safety-focused labs alone can carry the security perimeter (Schmidt: no, USG must) and on open weights (Anthropic is closed; Schmidt wants US open models — his money and his op-eds point different ways here)
• Overlaps with Hassabis on science-for-good framing (Schmidt Sciences; Hologen); splits on geopolitics — Schmidt is more hawkish, and now a defense vendor
• Splits sharply with Andreessen on regulation — Schmidt endorses export controls, KYC for compute, and active industrial policy that Andreessen treats as capture; converges unexpectedly on wanting more US open-weight models, for opposite reasons (Andreessen: check on state power; Schmidt: instrument of state power)
• Splits with Mostaque, Buterin on open-source frontier — v1 said "Schmidt sees open weights as adversarial diffusion." Narrowed: he now wants US open weights as a diffusion tool; the remaining split is who should own the openness, not whether
• Splits with Acemoglu, Autor, Frey on frame — labor and institutional outcomes are downstream of who wins the capability race. Narrowed in July 2026: he signed the displacement statement and funds 19 labor-economics studies; the Canaries data that regraded Brynjolfsson and vindicated Frey is what he is now citing
• Splits with Zuboff, Crawford, Gebru, Sun on emphasis — power-and-harm critique is acknowledged but subordinated to the state-competition frame; unchanged, though the Prop 40 fight gives the critique a fresh exhibit
• Splits with Yudkowsky on pause/moratorium — building faster (with US leadership) is treated as the safer strategy; MAIM is his answer to "then how do you stop a race?" (mutual sabotage, not mutual restraint)
• NEW — overlaps with Karp, Luckey as the defense-AI vendor cohort; splits on scale philosophy — Schmidt's $400 drones vs Anduril's and Palantir's platform economics, and Schmidt argues publicly against the Manhattan-style state program Karp's integrators would run
• NEW — splits with Sutskever on scaling: Schmidt restated the scaling law as the mechanism of what is coming (Harvard, 2025-12) the month after Sutskever buried it
• NEW — overlaps with Kai-Fu Lee on China's application-first advantage; with Zuckerberg his "no US open-source leadership" is the outside verdict on Llama's stall
• NEW — overlaps with Naam on grid-as-binding-constraint; his +29/+67 GW sit well below Naam's ~100 GW band — same direction, different order of magnitude, worth reconciling on the Naam file

Track record

• Built Google from start-up CEO into a trillion-dollar firm (2001–2017) — operating record is strongest in the roster
• NSCAI delivered final report 2021-03 — most influential US AI-strategy document of its era; many recommendations adopted (CHIPS Act 2022, AI EO 2023, export controls 2022–2024); critics argue the framing accelerated arms-race dynamics — and in 2025 he co-authored the paper arguing against the race he helped start
• SCSP (2021–) — sustained policy throughput; widely read in DC, NATO, and allied capitals; AI+ Expo now ~20,000 attendees (2026)
• Books with Kissinger (The Age of AI 2021; Genesis 2024) — set the "AI as new epistemic regime" frame now common in elite discourse
• Capital-allocator track record (Schmidt Futures, Schmidt Sciences, Hillspire's 22 AI companies incl. Anthropic, defense-tech investments) is significant but harder to score — conflict-of-interest watch per CONVENTIONS rule 25
• Forecasting record on AI timelines pre-2022 was relatively cautious; updated faster than most peers once GPT-4-era capabilities emerged — and has since overshot: "vast majority of programmers replaced within a year" (2025-04) and "recursive self-improvement within a year or two" (2024-08) both expired unmet; he reset the second himself. Apply a ~2x slow-down factor to his unscored clocks
• Called the China gap early and correctly in the revision: "two or three years" (early 2024) → "turning point" (2025-01-28, the day after the DeepSeek market shock) → "no longer far behind" (2025-05). Wrong number, fast update — the opposite failure mode from most of the bench
• Relativity Space (2025-03–): unscored — a rocket company whose only prior vehicle never reached orbit, a maiden flight one to two quarters away, a NASA Mars contract already in hand. Terran R is the first Schmidt operating bet with a binary public outcome since Google
• Swift Beat (2023–): the reported 90% Shahed-kill share and 70% autonomous hit rate are supplier-side or unconfirmed; he declined to confirm them himself. Unscored pending an independent Ukrainian or allied assessment

Empirical vs normative

• Empirical: capability timelines (AGI 3–5 years, SI by ~2031, RSI by ~2029), energy/compute demand (+29 GW 2027, +67 GW 2030), the gap with PRC (now conceded to be months), agentic-AI emergence (held), free Chinese models winning global adoption (live), Russia out-scaling Ukraine (live, self-interested)
• Normative: US should lead by deterrence plus deployment rather than a sprint; democracies should coordinate (unfalsifiable as stated); the US should field open-weight models (reversal from v1's reading); government should partially direct frontier compute and research; tech should self-supply power; a one-time 5% wealth tax on billionaires should fail
• Conflict-of-interest note (CONVENTIONS rule 25): v1 said "consistent with thesis but also benefits directly from the policies he advocates — weight accordingly, not dismissively, but skeptically." Upgraded: he now operates the beneficiaries (Relativity, Swift Beat, Hologen, SandboxAQ chair) and writes undisclosed op-eds in their markets. The one in-window position that cuts against his interest — MAIM, arguing against a state AI sprint that would be a bonanza for defense vendors — is therefore the one to weight highest

Weak spots / open questions

• Heavy reliance on the China-as-peer-competitor frame — v1 said "if DeepSeek-class PRC progress turns out to be largely indigenous and chip-control-resistant, the export-control thesis weakens." It did, and he said so (NYT, 2025-05); what replaced it — fund US open models, out-deploy China — has not yet been tested against anything
• Personal investment exposure to defense-AI and compute infrastructure creates pervasive conflict-of-interest pressure across his policy positions — hardest to score independently of his portfolio. Now operating exposure, not just investment exposure (see Incentive map)
• 2024-08 Stanford talk leak (where he said AI start-ups should "steal" content and let lawyers clean up later, and blamed Google's AI lag on remote work — comments he partially walked back) suggests his unfiltered views may be more cynical / instrumentally aggressive than his published frame — the Prop 40 spending is the 2026 exhibit for the same gap between the public-interest byline and the private position
• Operational silence on distributional outcomes — v1: Schmidt rarely engages Acemoglu / Autor-style labor-share arguments; treats them as downstream. Partly closed by the July-2026 statement and the Schmidt Sciences labor program — but his October-2025 "more jobs than destroyed" and July-2026 "large-scale displacement" have not been reconciled by him, only by his signature
• "Democracies must lead" coalition rhetoric underspecified — which democracies, on what terms, with what compute access, is largely missing; still missing after a full window, hence the unfalsifiable mark
• Theory of x-risk is hawkish-strategic rather than technical — engages less with mechanistic alignment than with strategic stability; less robust if alignment turns out to be the binding constraint rather than geopolitics. Mitigated in behavior (his foundation now funds the technical work) but not in his own writing
• Track record on his own former company (Google's AI delay) is awkward — Schmidt now critiques the slow-incumbent posture he partly built
• NEW: MAIM's mechanism is unaudited. "Preventive sabotage" of a rival's data centers — cyber or kinetic — is proposed as a stabilizer; the paper does not show why sabotage stays below the threshold of war, and no state has endorsed the frame. Check: any official US or PRC document adopting deterrence-of-AI-projects language by end-2027
• NEW: the 99%-of-electricity line has no derivation he has published; it travelled as a headline and now underwrites an orbital-data-center business he owns. Treat as rhetoric until he shows the model
• NEW: Swift Beat's "at cost" pricing to Ukraine plus "hundreds of thousands" of units implies a funding source he has not disclosed (his own capital? Western government buyers?); the economics of the at-cost promise are the open question that decides whether the drone doctrine is a policy or a pitch
• NEW: undisclosed conflicts in op-eds — WaPo (2026-09-09) ran his drone argument without a Swift Beat line; Fortune (2026-03-06) ran his energy argument without a Relativity line. Editorial failures, but a pattern; assume every Schmidt byline has a company behind it until shown otherwise

Rich's take

• (your synthesis here)

Delta log

2026-09-21 — v1→v2 migration + validation (bulk batch)

• Grades: 5 Held / 7 Drifted / 4 Wrong (2 unfalsifiable marked for rewrite; 6 new in-window predictions logged live with checks, not graded).
• Wrong: PRC "~2 years behind" (Epoch: 4–14 months, 7 average, Epoch AI; retired by him in WaPo, 2025-01-28 and NYT, 2025-05-05) · open weights as proliferation risk, valence (he calls for US open models: WaPo 2025-01-28; Harvard, 2025-12-01; mechanism Held) · recursive self-improvement "within a year or two" of 2024-08 (he reset it to four years, Harvard 2025-12-01; MIT Technology Review, 2026-08-18) · "vast majority of programmers replaced within a year" of 2025-04-10 (SAN; BLS +15.8% developer projection, BLS OOH).
• Drifted: agentic/recursive timeline compression (agents yes, recursion no) · export-control expansion (kept in Superintelligence Strategy, doubted in NYT 2025-05; no statement found on the Aug-2025 H20 reversal) · x-risk secondary to geopolitics (flattened by the 2025 paper and Harvard remarks) · labor not his primary lens (signed "We Must Act Now," 2026-07-13, after "more jobs than destroyed," 2025-10-30) · decisive 2–5-year window (replaced by a deterrence-equilibrium frame) · NSCAI $40B/NAIRR (pilot still; ~$3B/yr nondefense AI R&D, NITRD) · nuclear as dominant new data-center power (~10 GW of deals, zero SMR megawatts delivered; Crane restart H2-2027, WITF, 2026-07-28).
• Held: nation-scale energy build-out, mechanism (House testimony, 2025-04-09; CNBC capex, 2026-02-06; valence shifted to ratepayer protection, Fortune, 2026-03-06) · government should fund/direct research, his side (Schmidt Sciences, 2026-01; Trustworthy AI RFP, 2026) · AI agents do real work within a year (Gartner 2026) · $300B+ capex (undershot 2x) · AI reshapes warfare within a decade (Noema, 2026-06-16).
• Unfalsifiable — rewrite or retire: "democracies need a coordinated AI bloc"; the diplomacy and epistemology branches of the 2021 decade claim (per-branch checks written above).
• Most surprising delta: the NSCAI chair who launched the US–China AI race in 2021 co-authored the paper arguing a US superintelligence sprint would get "MAIM'd" — and then spent 2025–26 becoming a rocket CEO and a wartime drone supplier (TechCrunch, 2025-03-05; TechCrunch, 2025-03-10; UAS Vision, 2025-07-07). The v1 line placing him "closer to Aschenbrenner" inverted 14 months before the May-2026 review that left it standing.
• Employer/role check (mandatory): the May-2026 v1 pass missed a 14-month-old CEO job, a renamed defense company with a head-of-state MoU, a co-founded medical-AI startup and a $5B+ AI portfolio. Hats counted before: 5 (SCSP, NSCAI-historical, ex-Google, Schmidt Futures/Sciences, White Stork). After: 11 (see Who he is). Post-review dated sources in-window (2026-05-22 → 09-21): NASA Aeolus selection (Scientific American, 2026-06-18); General Intuition Series A (2026-06-25); Gradium seed extension (TechCrunch, 2026-07-09); "We Must Act Now" (2026-07-13); Revolving Door Project conflict inventory (2026-08-17); Prop 40 dispute (Fortune, 2026-08-21); WaPo drone op-ed (2026-09-09) and its disclosure critique (DroneXL, 2026-09-11); NBC interview (2026-09-14). Incentive discount raised from "skeptical" to maximal on drones, energy and launch.
• Numbers re-verified at migration: v1's "~2 years" (his figure was "two or three," 2024); v1's "$300B+" (his Stanford figure; actuals $410B/$725B); Relativity "~$800M" is a press figure (IndexBox) not confirmed by the company — flagged; Swift Beat 90%/70% figures are unconfirmed and he declined to confirm them — flagged.
• Tier: quarterly — Terran R's maiden flight, the Prop 40 vote (Nov 2026), the AI-mathematician check (end-2026) and the Ukraine drone war all resolve inside two quarters; wake triggers cover launch, contract, vote, appointment, raise and restatement events.
• Confidence unchanged at medium: the record is now much thicker than v1's (which carried nine [TBD] tags), but his highest-stakes numbers are self-reported and undisclosed-interest bylines are the norm.

Sources

• 2025–26 primary — strategy and policy: Hendrycks, Schmidt, Wang — Superintelligence Strategy: Expert Version (arXiv 2503.05628, 2025-03-07; site nationalsecurity.ai) · TechCrunch on the paper (2025-03-05) · Written testimony, House Energy & Commerce (2025-04-09) · Hearing transcript (TechPolicy.Press) · WaPo op-ed — DeepSeek challenges US supremacy (2025-01-28) · TechCrunch recap · NYT with Selina Xu — "DeepSeek. Temu. TikTok. China Tech Is Starting to Pull Ahead" (2025-05-05, via SCSP) · NYT with Xu — "Silicon Valley Needs to Stop Obsessing Over Superhuman A.I." (2025-08-19) · syndicated text (Business Standard) · Fortune commentary — tech should power its own data centers (2026-03-06) · "We Must Act Now" signature + NYT essay (2026-07-14) · Spokesman on the statement · WaPo op-ed — "How Russia is erasing Ukraine's battlefield advantage" (2026-09-09)
• 2025–26 primary — talks and interviews: AI+Biotech Summit remarks — programmers, AGI 3–5y, SI 6y (2025-04-10) · "San Francisco consensus" / SI by 2031 (Benzinga, 2025-04) · Digitalist Papers — "The San Francisco Consensus" · TED — "The AI revolution is underhyped" (2025-05) · FII Riyadh — more jobs than destroyed (2025-10-30) · Moonshots podcast — Chinese open models will dominate adoption (reported 2025-11-12) · Digitimes with Bloomberg download data (2025-11-17) · Harvard Kennedy School forum — RSI in 4 years, AI mathematicians next year, no US open-source leadership (Crimson, 2025-12-02) · same forum — "under-hyped, not a bubble" · Noema — AI & drones, "the biggest revolution in the history of warfare" (2026-06-16) · Hvylya interview — declines to confirm Swift Beat figures (2026-09-09) · NBC News — war could get "much worse" (2026-09-14) · NBC/Engel — "rate of innovation is faster than I've ever seen" (2026-09-16) · SCSP AI+ Expo 2026 keynote listing — "Delivering AI Advantage" (content not retrieved)
• Role / company / money: TechCrunch — Schmidt becomes Relativity CEO (2025-03-10) · SpaceNews — CEO + Terran R update · IndexBox — ~$800M figure (press, unconfirmed) · SpaceNews — late-2026 debut, industry expects 2027 · Via Satellite — SES multi-launch deal (2025-11-12) · Scientific American — NASA Aeolus selection (2026-06-18) · TechSpot — data centers in orbit (2025-05) · Office of the President of Ukraine — Swift Beat memorandum (2025-07-03; page returned 403 at review, cited via UAS Vision) · UAS Vision (2025-07-07) · Kyiv Post — 90% Shahed-kill claim (unconfirmed) · Kyiv Post — on the WaPo op-ed (2026-09) · DroneXL — "He sells the fix" disclosure critique (2026-09-11) · STAT — Hologen $150M Series A (2026-01-28) · CNBC — Hillspire's 22 AI startups incl. Anthropic (2025-01-21) · General Intuition $320M Series A, Hillspire participating (2026-06-25) · Gradium seed (2025-12-02) · Gradium extension (2026-07-09) · Revolving Door Project — conflict inventory (2026-08-17; advocacy source, facts checked where possible) · Fortune — Prop 40, $3M+, dispute with Brin (2026-08-21) · Schmidt Sciences — AI at Work awards (2026-01) · Schmidt Sciences — Trustworthy AI RFP (2026) · SCSP AI+ Expo 2026
• Market-level data used for grading: Epoch AI — Chinese models lag the US frontier by 7 months on average · CNBC — China gaining, US advantage (2026-08-07) · Chinese open models ~4.4 months behind (2026-09) · CNBC — ~$700B 2026 AI capex (2026-02-06) · Utility Dive — DOE $1B Crane loan · WITF — Crane restart status (2026-07-28) · SMR data-center deal tracker (2026) · Gartner — 2026 Hype Cycle for Agentic AI · Gartner/IDC/McKinsey agent-adoption summary (2026) · MIT Technology Review — RSI "might not come so quickly" (2026-08-18) · BLS OOH — software developers · Tech job market 2026 (Indeed/Stanford figures) · NITRD — federal nondefense AI R&D · NSF 25-546 — NAIRR Operations Center solicitation · GovCIO — NAIRR transition
• Books: The Age of AI: And Our Human Future (2021, with Henry Kissinger & Daniel Huttenlocher); Genesis: Artificial Intelligence, Hope, and the Human Spirit (2024-11, with Kissinger & Craig Mundie); The New Digital Age (2013, with Jared Cohen)
• Reports: NSCAI Final Report (2021-03); SCSP reports (2022–) — v1's "[TBD: link specific reports]" remains open; the 2026 Expo keynote is the nearest in-window SCSP primary
• Earlier essays / talks carried from v1: Foreign Affairs essays (2022–2025) on AI and great-power competition [v1 TBD: specific dates and titles — not resolved this cycle]; Stanford ECON 295 guest lecture (2024-08; recording removed; secondary summary); Aspen Security Forum and Council on Foreign Relations appearances [v1 TBD: dates]; Noema interview (2024); The Diary of a CEO (2024) [v1 TBD: verify]; Bari Weiss / Honestly [v1 TBD: date]; NSCAI-era congressional testimony 2019–2021 and 2023-05-17 House testimony; WSJ / NYT / The Atlantic op-eds (2022–2025) [v1 TBD: source-tier links]; ericschmidt.com thought-leadership index

migrated v1→v2 2026-09-21 · backup: _archives/eric-schmidt.html.bak-20260921 · converts-from: people/eric-schmidt.md · AI & Society domain