Kai-Fu Lee

Who he is

Affiliation: Chairman & CEO, Sinovation Ventures (Beijing VC, founded 2009; US$3B under management as of 2022) and Founder & CEO, 01.AI (Beijing, founded 2023). Role check run 2026-09-21: both titles are unchanged since v1 and confirmed on the firms' own pages (Sinovation team page; 01.AI About) — but the company under the second title changed entirely and v1 never noticed. 01.AI stopped pre-training its own models in March 2025 and "no longer positions itself as a maker of frontier large language models" (Bloomberg, 2026-07-20, syndicated); it now fine-tunes DeepSeek, Alibaba's Qwen and Z.AI's GLM and sells enterprise data-and-agent software — Lee's own phrase: "the Palantir of China." About 240 employees; roughly half the business outside China; a pre-IPO round open; Hong Kong listing targeted for 2027 ("We're aiming for 2027, after the fiscal year-end"). Author: AI Superpowers (2018), AI 2041 (2021, with Chen Qiufan) and AI Native: The Mandate to Transform Your Company (Crown Currency; ebook and audio 2026-09-15, US hardcover 2026-11-03, 208 pp — publisher page). Earlier: President of Google China (2005–09); Microsoft (founding director of Microsoft Research China, 1998; corporate VP 2000–05); SGI (1996–97); Apple (1990–96); Carnegie Mellon faculty (1988–90). Born 1961-12-03 in Taipei; relinquished US citizenship in 2011; diagnosed with Stage IV lymphoma in 2013 (Wikipedia, biographical facts only).
One-line position (v1, carried): The AI race is not one contest but many — China and the US are each winning different races — and whoever masters AI deployment at scale (not just research) will shape the next economic order.
One-line position (2026 restatement, his): the race is now "iPhone vs Android" — US labs sell the closed, premium model to enterprises and take the profit; Chinese labs ship open, near-free models about six months behind and take the reach — and every company must become "AI native" or be run by one that is (Bloomberg Weekend interview, 2026-09; paywalled — reconstructed from two dated write-ups, see Sources).

Discipline & technical bet

Computer scientist by training (Carnegie Mellon faculty, 1988–90) who has not done frontier research since the 1990s; for 17 years a venture investor and, since 2023, a founder-CEO. The concrete wager, read from where 01.AI's money went rather than from his adjectives: not the model. In January 2025 he told KrASIA "It's clear that only tech giants can [bear the costs of training super-large models]" and described moving his large-model trainers toward an Alibaba joint lab — "Alibaba takes on the training of these massive models, while our nimble and efficient teams concentrate on developing smaller, cost-effective models" (KrASIA, 2025-01-09). By March 2025 the company had stopped pre-training altogether (Financial Times, 2025-03-22, as cited by Wikipedia; the FT primary was not fetchable this cycle). By November 2025 the model was a commodity input: "Whoever has the best model, we'll use it. If it's DeepSeek or Qwen, we are agnostic" (Bloomberg interview transcript, 2025-11). The bet is the application and data layer: TrueNorth, an "enterprise AI decision hub platform" released July 2026 with three applications — Boss AI, Investor AI, TopSales AI (01.AI About) — built on a "truth engine" that "records, organizes, and analyzes every internal meeting, sales call, and customer service request," and a 19-agent "decision assistant" he says he uses to run 01.AI himself (Semafor, 2026-09-18). His second bet, as an investor, is physical: "I'm betting on hardware" rather than large language models (Bloomberg transcript, 2025-11), which matches his TED AI line that "The VCs in the US don't fund robotics the way the VCs do in China" (VentureBeat, 2025-10-22). Net: the bench's "China wins deployment, not the frontier" theorist is now, personally, a deployment company built on someone else's frontier — thesis and product are the same object.

Key claims (Says)

• Empirical (2025): China leads in AI deployment, not just research — Chinese companies move faster in consumer applications, manufacturing robotics and cost-efficient inference. Held on robotics and on open-weight reach; the consumer half is live, not won. Market-level data, not his adjectives: China installed 295,000 industrial robots in 2024, "54% of global deployments," with an operational stock that "exceeded the 2 million mark," and Chinese makers' domestic share "climbed to 57%" (IFR World Robotics, 2025-09-25); Stanford's 2026 index says "China leads in publication volume, citations, patent output, and industrial robot installations" (AI Index 2026). Consumer: on a16z's mobile list "22 of the 50 apps were developed in China, while only three are primarily used in China" — reach, yes — but "ChatGPT still leads" on both web and mobile (a16z Top 100, 2025-08-27). His own carve-out stands: "The enterprise adoption will clearly be led by the United States" (VentureBeat, 2025-10-22).
• Empirical: The AI race is multi-dimensional — hardware, models, applications and robotics are separate contests; neither country dominates all. Held — he restated it as separate races (Bloomberg video, 2025-09-11, title: "The U.S. and China Are Winning Different AI Races"), then as consumer + robotics to China, enterprise + research to the US (TED AI, 2025-10-21), then as reach vs profit (2026-09). The Stanford split (US: more top-tier models and higher-impact patents; China: volume, citations, robots) is the same picture from a neutral source (AI Index 2026).
• Empirical (AI Superpowers, 2018): "Data is the new oil" — China's scale advantage in data generation (1.4B users, mobile-first economy) is a structural edge for AI training. Drifted — not restated since 2018–19, and his own newer causal story omits it. Last found in his mouth: "Data is the raw material on which AI runs. It is like the role of oil in powering an industrial economy" (Noema, 2018-09-24) and "China is now to data what Saudi Arabia is to oil" (NYU Tandon, 2018-10-09). His 2025–26 explanation of China's catch-up is cost and engineering, not data: "Chinese companies are spending less than ten percent of what the top American companies spend on training, yet are producing models that are ninety to ninety-five percent as capable" (Capgemini conversation, undated); Chinese labs closed most of the gap with "perhaps 1–3% of the GPU power" through "maybe ten times the engineering effort" (2026-09 interview, via write-ups). Foreign Affairs had objected to the book "overgeneralizing the usefulness of Chinese data sets" (Wikipedia, reception); the author has, in effect, conceded by silence. Check: any restatement of the data-advantage mechanism in the AI Native press cycle by 2027-03-31 reverses this to Held.
• Empirical (2018): Four waves of AI — internet AI → business AI → perception AI → autonomous AI; China competitive in all but autonomous AI. Drifted — superseded by its author. Generative AI, the wave that actually arrived, is not in the taxonomy, and his 2025–26 framings ("different races"; "iPhone vs Android") replace it. The book's per-wave China-vs-US forecasts could not be re-read this cycle (the book is not fetchable and its site carries none of the numbers — aisuperpowers.com), so they stay unverified here rather than restated from memory.
• Empirical (2025): China is pulling ahead in open-source AI (DeepSeek, Qwen, etc.). Held on the observable — with his own negative valence logged. "The 10 highest rated open source [models] are from China" (VentureBeat, 2025-10-22); Stanford: "In February 2025, DeepSeek-R1 briefly matched the top U.S. model" and the lead has since traded hands (AI Index 2026). But he does not think this wins the race: "I don't think open source models will win" (same TED AI talk) and "Open source may win on reach and adoption, while closed source wins on margin and ecosystem control" (Capgemini). Mechanism held; valence, in his own telling, is a consolation prize — read alongside Andreessen's file, which logged Chinese open-weight leadership as the "check on state control" pointing the other way.
• Empirical (2025): US will monetize first — American companies lead enterprise AI revenue, Chinese companies follow. Held — restated in November 2025 (US first, because most of OpenAI's spending "is spent for future revenues" — Bloomberg transcript) and at TED AI ("The Chinese companies have not yet developed a habit of paying for software on a subscription"). Market-level: US private AI investment "reached $285.9 billion in 2025, more than 23 times the $12.4 billion invested in China" (AI Index 2026). Revenue-side data will arrive with the 2027 Hong Kong prospectuses (Moonshot, 01.AI) — see the check under predictions.
• Empirical (2018–19): AI will displace ~40% of jobs within 15 years — stronger than most economists. Drifted — re-dated by its author, then softened. The primary: "all together in 15 years, that's going to displace about 40 percent of the jobs in the world" (60 Minutes, aired 2019-01-13). But he keeps a second, harder version: in 2017 he said AI "will probably replace 50% of human jobs" "in the next 10 years," and in March 2024 called it "actually uncannily accurate" (Fortune, 2024-05-25) — that is 2027. By September 2026 the language is reskilling and wages: "Most people will not be laid off if they're willing to learn new AI skills … 'but they might be paid three or five times more'" and "The winners will not measure success by payroll cuts" (Semafor, 2026-09-18). Market-level grade: unemployment "unchanged at 4.1 percent," payrolls "+162,000 in August" (BLS, 2026-09-04); the only displacement signal is at the hiring margin — 22–25-year-olds in AI-exposed occupations "19% below where it would be" relative to less-exposed peers, "primarily through reduced hiring of young workers rather than increased separations" (Brynjolfsson et al., Canaries, revised 2026-08-12). Mechanism (white-collar, entry-level first) is showing; magnitude is not. Both branches carry checks below.
• Empirical (2026-09-18, NEW): "The export controls do not work" and China "will be 'completely self-sufficient' in chips within two to three years" (Semafor); at WAIC he called the controls an "imaginative but unsuccessful effort to contain China" (StartupHub, 2026-07-20). Live — check under predictions. Note this is both his analysis and the line Beijing prefers said in English; the incentive map applies in full.
• Empirical (2026-07, NEW): AI models are now "smarter than all humans," so applying them to simple tasks is "a pity" (StartupHub, 2026-07-20). Unfalsifiable as stated — rewrite or retire: no benchmark, no task class, no date. Contrast Sutskever, who reads the same leaderboards as evidence of a generalization gap.
• Normative (2026, NEW): AI transformation is a CEO's job — "Only the CEO can define the ambition, concentrate resources, redesign the operating model, remove barriers, and remain accountable for the result"; "Middle managers often hope that the CEO will forget. And the CEO does forget, but the CEO's AI does not" (Semafor, 2026-09-18). Not gradeable; logged because it is the book's thesis and Boss AI's sales pitch in one sentence.
• Mixed (2018, carried): AI "is going to change the world more than anything in the history of mankind. More than electricity" (60 Minutes, 2019-01-13). Unfalsifiable — noted, not graded.

Notable predictions — with falsifiable checks

• (2018–19) ~40% of jobs displaced within 15 years (by ~2033–34). Graded under Key claims (Drifted). Check: BLS employment-population ratio and the Canaries exposed-occupation series at 2033-12-31; a sustained fall in prime-age employment of 10+ points in exposed occupations → Held; under 5 points → Wrong. The "capable of displacing" reading he sometimes retreats to is unfalsifiable and is not what he said on air.
• (2017, reaffirmed 2024-03) 50% of human jobs replaced by 2027. Live — trending Wrong on the date. Formal grade at 2027-12-31: BLS unemployment above 8% or a Canaries-style relative employment gap above 30% in exposed occupations → some credit; 4–5% unemployment with a hiring-margin-only signal → Wrong. As of 2026-09: 4.1% and a 19% entry-level gap (BLS 2026-09-04; Canaries 2026-08-12). Under the serial-re-dating rule this is his first dated miss if it fails; a second re-dating gets graded Wrong on the date by default.
• (2018) China will match or exceed the US in AI applications by the mid-2020s. Held on consumer reach and robotics (22 of 50 China-developed mobile apps, a16z 2025-08-27; 54% of robot installs, IFR 2025-09-25) and on model parity (top-model gap "just 2.7%" as of March 2026, AI Index 2026); the enterprise carve-out is his own. v1's "hardware gap persists" note now competes with his 2026 claim that the chip gap closes in 2–3 years — see below.
• (2025) US companies monetize AI first, China follows. Graded under Key claims (Held on investment data and his restatements). Revenue check: when the first Hong Kong prospectus from a Chinese frontier lab (Moonshot has confidentially filed — TechStartups citing Reuters, 2026-09-03) discloses audited revenue, compare with OpenAI/Anthropic disclosed run-rates at the same date; Chinese lab revenue under one-fifth of the US leader's → Held; parity → Wrong. Date: 2027-12-31.
• (2025) China pulling ahead in open-source AI. Graded under Key claims (Held on observable; his valence negative). Forward check: is the top open-weight model on an independent leaderboard Chinese at 2027-06-30? Yes → Held again; a US or European open-weight leader for a full quarter → Drifted. (Same check lives in Andreessen's file with the opposite hoped-for outcome — grade them as a pair.)
• (2025-03-20/21, NEW) China's model sector consolidates to three: DeepSeek, Alibaba, ByteDance, "with DeepSeek currently leading in momentum"; four in the US: xAI, OpenAI, Google, Anthropic (TechNode, 2025-03-21; Bloomberg 2025-03-20 paywalled). Drifted — consolidation is happening, but the survivor list is wrong so far. 01.AI itself left the frontier (his own company was the first casualty); yet "Z.AI, formerly known as Zhipu, and MiniMax have already listed in Hong Kong this year," Moonshot "released Kimi K3 in July" and "is seeking to raise about $3 billion" at "a $50 billion" valuation (TechStartups/Reuters, 2026-09-03). Three names were never a forecast of who quits — it was a forecast of who survives, and at least three non-named labs raised public or late-stage money in 2026. Check at 2027-12-31: count Chinese labs that shipped a frontier-class model in the prior 12 months; three or fewer, matching his list → Held; five or more → Wrong. No timeframe was given, so the check date is the file's, not his.
• (2025-10-21, NEW) China dominates consumer AI applications and robotics manufacturing "within years"; the US keeps enterprise adoption and cutting-edge research (VentureBeat, 2025-10-22). Live. Check at 2027-12-31: a Chinese-developed assistant in the global top three by monthly active users on a16z's list (consumer) and Chinese share of global industrial-robot installs still above 50% (IFR) → Held; ChatGPT/Gemini top three with no Chinese entrant, or Chinese robot share under 45% → Wrong on that half.
• (2025-11, NEW) Chinese models "trail by three to six months" (Bloomberg transcript, 2025-11); "roughly six months" in 2026-09 (via write-ups). Held — the top-model gap was "just 2.7%" in March 2026 after the lead "traded … multiple times since early 2025" (AI Index 2026). Check at 2027-06-30: gap on an independent index under 9 months → Held; over 12 → Wrong. Same number, same check, as Naam's "perhaps three months behind."
• (2026-09-18, NEW) China "completely self-sufficient" in chips within two to three years (Semafor) → 2028-09 to 2029-09. Live. Check at 2029-09-30: a Chinese flagship model trained end-to-end on domestic accelerators and Nvidia's China data-center revenue below 5% of its total → Held; flagship Chinese training runs still on Nvidia silicon → Wrong. Companion observable for "export controls do not work": the same two datapoints, read at 2027-12-31.
• (2026-07, NEW) 01.AI profitable "sometime next year" (2027) and Hong Kong IPO in 2027 "after the fiscal year-end"; bookings up "three to four times from last year" (StartupHub 2026-07-20; Bloomberg 2026-07-20). Live — self-reported until a prospectus exists. Check: audited FY2027 results show positive net income and a listing is priced by 2028-06-30 → Held; IPO withdrawn or delayed past 2028 → Wrong.
• (2026-09, NEW) "Within five years" companies restructure around units like "20 people and 100 AIs"; CEOs far from the frontier have "a three-to-five-year 'buffer'" (write-ups of the Bloomberg interview; Semafor). Live. Check at 2031-09-30: Census BTOS firm-level AI production-use above 50% and a documented headcount-per-revenue drop of 3x or more at listed software/services firms → Held; BTOS under 30% → Wrong.

Revealed behavior (Does)

• Conceded the frontier in three steps. 2025-01-02: Alibaba Cloud and 01.AI "will link arms on setting up an Industrial Foundation Model Joint Laboratory"; Lee denied selling the pre-training team (reported at about 60 staff) to Alibaba Cloud while confirming the trainers were the "natural fit" to move (Yicai, 2025-01-07; KrASIA, 2025-01-09). 2025-03: pre-training stopped (FT via Wikipedia); WorldWise enterprise platform released (01.AI About). 2025-11: model-agnostic — "If it's DeepSeek or Qwen, we are agnostic." He gave the reason himself: "Only a handful of companies with essentially bottomless balance sheets could still justify the cost of building models from scratch. Everyone else needed a new business model" (TNW, 2026-07-20).
• Rebuilt as an enterprise-software vendor and pointed it at an exchange. 2024 revenue "more than CNY100 million (USD13.6 million)" with 70% from enterprise clients and 20–30% from overseas productivity tools such as PopAi (Yicai; KrASIA). TrueNorth / Boss AI / Investor AI / TopSales AI released July 2026. About 240 employees; "Roughly half of 01.AI's business now comes from outside China"; "in the process of unwinding its offshore holding structure, a prerequisite for a smoother overseas listing"; pre-IPO round to close "around the time 01.ai unveils annual results for the first time" (Bloomberg, 2026-07-20). Kazakh-language model AlemLLM and a Kazakhstan partnership on the About page — the "outside China" half is partly sovereign-AI work.
• Put the fund's money on hardware. "I'm betting on hardware" (2025-11) — consistent with his claim that Chinese VCs fund robotics and US VCs fund generative AI (TED AI, 2025-10-21). Sinovation still lists him Chairman & CEO with an AI Institute he presides over; no new fund found in this cycle's searches (last on record: Fund V, US$203M, 2022 — Wikipedia). Flag: the fund's current dry powder is unknown.
• Wrote the sales book and toured it the quarter the product shipped. TrueNorth/Boss AI (July 2026) → AI Native ebook (2026-09-15) → Bloomberg Weekend (2026-09-04) → Semafor (2026-09-18). Per the newest-vehicle rule: the thesis ("companies need an AI boss") is the product.
• Uses his own product — the 19-agent decision assistant and the "truth engine" over every meeting; one detail he volunteered is that "93.6% of his meeting comments were detail-seeking questions" (2026-09 interview, via write-up). Eat-your-own-cooking is a real signal; it is also the demo.
• Personal stakes: Beijing-based, Taiwan-born, gave up US citizenship in 2011 (Wikipedia). Whatever he thinks about PRC policy is not something he can say from where he sits.

Feels

A survivor's urgency (Stage IV lymphoma, 2013) fused with a bridge-builder's wound: he has spent a career explaining each superpower to the other and wants both to be right about different things. He wants "good enough and cheap" to count as a virtue rather than a consolation prize, and reskilled workers to be "paid three or five times more." The fear underneath: being read as Beijing's spokesman abroad — and being the AI pioneer whose own model company lost the race he wrote the book about.

Hears

Chinese founders and LPs through Sinovation; Alibaba Cloud as investor and lab partner; enterprise CEOs buying transformation; DeepSeek/Qwen/GLM release cadence; a Western press circuit (Bloomberg, TED, Semafor, Capgemini) that books him as the China explainer. Almost nothing from labor economists — he cites no payroll data — and little from the safety community. Downstream of no single thinker; upstream of Washington's "application-first China" turn (see Schmidt).

Sees

Both ecosystems from the inside — the only bench member who ran Google China and then a Chinese "AI tiger." He knows what Chinese enterprise buyers will and won't pay for, the real cost curve of fine-tuning open weights ("a few million dollars rather than a few hundred million" — Capgemini), and which Chinese labs have a business model. His blind spot is the mirror image: he has watched US enterprise from the outside since 2009, and his "top 1,000 vs top 100,000" talent comparison (Semafor) is an outsider's aggregate.

Incentive map

Re-derived this cycle from income sources, not carried. (1) 01.AI — founder-CEO of a company in a pre-IPO round with a 2027 Hong Kong target, ~240 staff, self-reported bookings and profitability guidance, and a product line built on DeepSeek/Qwen/GLM. He cannot say "Chinese open weights are inferior or unsafe," "enterprises don't need AI transformation," or "the Palantir model doesn't travel." Every "China wins reach" claim raises the multiple on a company whose inputs are Chinese open weights. (2) Sinovation Ventures — Chairman & CEO of a Beijing VC selling Chinese-tech upside to LPs (US$3B AUM as of 2022; no newer fund found). He cannot say "Chinese AI is a bad bet" or "the robotics wave is overfunded." (3) The book — AI Native (Crown Currency, 2026-09/11): "they're mostly doing it wrong" sells both the book and Boss AI; anything said on the 2026-09 press tour is marketing-dated. (4) Geography — Beijing-resident, no US passport since 2011; "the export controls do not work" is simultaneously an operator's observation and the line the PRC wants said in Bloomberg. (5) Investors — Alibaba Cloud is both a 01.AI backer and the lab that absorbed his trainers (Bloomberg 2026-07-20; Yicai 2025-01-07); Qwen is one of his three input models. Rule-25 discount: high on every China-vs-US comparative claim, on "CEOs are doing AI wrong," on export-control efficacy and on 01.AI's own numbers until a prospectus exists; moderate on labor claims (Boss AI's pitch is a company's best people "amplified 10 times," so "you won't be laid off" is also a pitch); lowest on his 2005–09 Google-era and 2018 book-era claims, which predate the current vehicles. v1's single line ("VC + AI company CEO … commercial interest per rule 25") understated this by three vehicles.

Theories aligned with

• US–China AI competition as the structural feature of the 21st century (v1) — now his "iPhone vs Android" duopoly; cf. Noema's 2018 title for the same idea, "The Great AI Duopoly"
• Data-network effects as competitive moat (v1) — weakened; he no longer argues it (see Key claims)
• Adjacent to post-labor economics — displacement is central to his frame (v1); his 2026 line that "society must rethink dependence on jobs" (via write-up) and his 2018 preference for "universal basic capital instead of universal basic income" (Noema) put him in the redistribution camp
• Automation displacement and white-collar labor end — "AI will increasingly replace repetitive jobs, not just for blue collar work, but a lot of white collar work" (60 Minutes, 2019); the Canaries hiring-margin data is the first market-level evidence on his side of the mechanism
• Techno-optimism, deployment variant — the win is diffusion and cost, not the frontier; a deliberate inversion of the Sutskever / frontier-lab worldview

What he's reacting against

• Western-centric narratives that frame China as merely copying US AI research — argues China leads in deployment, applications and manufacturing robotics (v1; still his line, now with "study group of smart kids" as the mechanism: "A learns from B, B learns from C" — Bloomberg transcript, 2025-11)
• The assumption that the US has a durable AI lead — claims the competition is far more contested than Washington believes (v1; now quantified as a six-month gap)
• Pure research fetishism — deployment, data access and entrepreneurial speed matter as much as frontier models (v1)
• Both US and Chinese zero-sum framing — advocates mutual recognition of complementary strengths (v1)
• NEW (2025–26): US export controls — "do not work," an "imaginative but unsuccessful effort to contain China"
• NEW (2026-03): American labs that "overcharge" — he called China's OpenClaw agent craze a "rallying cry" against them on Bloomberg's China Show (Bloomberg video, 2026-03-31; page fetch blocked — the quoted phrase is from Bloomberg's own listing text and is unverified against the video)
• NEW (2026): CEOs who delegate AI transformation to middle managers; applying frontier models to trivial tasks ("a pity")

Where he overlaps / splits (with Rich)

• Overlaps with Schmidt on US–China competition as the defining AI geopolitical frame; splits on perspective — Lee sees from inside Chinese tech, Schmidt from the US defense establishment (v1). Now also: Schmidt's Aug-2025 "application-first China" is Lee's 2018 thesis, and Schmidt's file already says so; they split on export controls (Schmidt softened, Lee says they fail outright).
• Overlaps with Acemoglu — v1 said "on large-scale labor displacement"; corrected this cycle: Acemoglu's own validated file carries a small macro effect (~0.55%/decade TFP) and "so-so automation." The real overlap is on mechanism (substitutive AI hits hiring first — Canaries fact 5); the split is on magnitude, with Lee at 40–50% and Acemoglu near zero. v1's emphasis line (Lee on the race, Acemoglu on institutional response) still holds.
• Overlaps with Andreessen on speed-of-deployment mattering; splits on politics — Lee is more concerned about social costs (v1). NEW split on open-source valence: Lee reads Chinese open weights as positive-sum sharing; Andreessen's file logs the same fact as the "check on state control" pointing the other way. Author-list check: no co-authorship or shared vehicle found.
• Splits with Cowen on displacement magnitude — Lee predicts ~40% (or 50%), Cowen expects slower, more uneven diffusion (v1). The 2026 payroll data is on Cowen's side of the aggregate and on Lee's side of the white-collar-first mechanism.
• Overlaps with Graylin — the sibling lead is confirmed: both now say China's edge is deployment/reach, not the frontier model (Graylin adds energy and robot volumes; Lee adds cost and "good enough"). Splits on whether that edge wins: Graylin's file carries "open-source models could overtake closed-source in performance longer-term" (2025-12-11); Lee says "I don't think open source models will win" and "closed source wins on margin." Same fact, opposite forecast — grade the pair at 2027-06-30 with the shared open-weight-leaderboard check.
• Overlaps with Naam on the size of the gap (Naam "perhaps three months," Lee "three to six") and on positive-sum open weights. The suggested energy intersection does not hold: Lee's only "electricity" line is a 2018 metaphor for AI's diffusion ("It's not nuclear; it's electricity," NYU 2018-10-09), not a grid claim; no Lee statement on China's power advantage was found.
• Overlaps with Brynjolfsson and Frey on white-collar exposure; their Canaries data is the first market-level evidence for his mechanism and the first against his magnitude — the same dataset that regraded Brynjolfsson Wrong and Frey Held grades Lee Drifted.
• Splits with Sutskever on what the leaderboards mean — Lee: "smarter than all humans"; Sutskever: benchmarks without generalization. Lee's company depends on the models being good enough; Sutskever's depends on their not being.
• Cf. Huang on export controls — a likely overlap, but Huang's own statements were not re-read this cycle; check his file for wording and date before treating it as agreement.

Track record

• AI Superpowers (2018) was directionally correct on China's AI ambitions — DeepSeek validated the "China as real competitor" thesis (v1). Held and strengthened: the top-model gap was 2.7% by March 2026 (AI Index 2026), and Washington's "application-first China" framing (Schmidt, 2025) is his book seven years late.
• Founded 01.AI (2023) — "building a Chinese frontier model company; execution track record puts skin in the game" (v1). Wrong — the skin-in-the-game bet failed on its original thesis. Pre-training stopped in March 2025; the company "no longer positions itself as a maker of frontier large language models" (Bloomberg, 2026-07-20). What remains is a real business (CNY100M+ 2024 revenue, an IPO track) built on other labs' models. The pivot may succeed commercially; the frontier bet did not. Founding date conflict for the record: 01.AI's About page says May 2023; Wikipedia, citing Bloomberg (2023-11-05), says March 2023 — this file uses the company's own date and flags the difference. Unicorn status came within the first year either way.
• Google China presidency gives genuine insider knowledge of both ecosystems — unusual dual fluency (v1). Held (biographical).
• ~40% displacement claim is bold and unfalsified but also unconfirmed; most labor economists are more conservative (v1). Regraded Drifted above; the 2027 branch is his first dated test.
• NEW: the March-2025 "three survivors" call — right that consolidation was coming (he was the first to quit), wrong on the names so far (Z.AI, MiniMax listed; Moonshot at $50B).
• NEW: the "three to six months" gap call (2025-11) — the most precise number he has put on the record, and the one that held best.
• Pattern: strong on direction, loose on dates and magnitudes; his best calls are structural (two ecosystems, deployment vs research), his worst are quantified timelines (jobs) and his own company's positioning.

Empirical vs normative

• Empirical: US–China competitive dynamics, deployment patterns, labor displacement forecasts (v1); now also the six-month gap, the cost ratio (under 10% of US training spend for 90–95% capability), chip self-sufficiency timing, and the "only a handful can pre-train" economics
• Normative: both nations should invest in social safety nets; AI competition should be managed, not zero-sum; human connection and purpose matter beyond economics (v1); now also: CEOs must own transformation personally; "Move fast, but make sure safety keeps pace" (Semafor, 2026-09-18); "universal basic capital" over UBI (Noema, 2018)
• His commercial position (VC + AI company CEO) means competitive-landscape claims carry commercial interest per CONVENTIONS rule 25 (v1) — re-derived above as five vehicles, not two; the discount went up, not down, with the pivot

Weak spots / open questions

• Dual role as analyst and competitor (01.AI, Sinovation) creates structural bias — he benefits from a narrative of Chinese AI strength (v1). Held, and sharper: after the pivot he benefits specifically from "Chinese open weights are good enough," which is the input cost of his product.
• AI Superpowers was written pre-GPT era; the landscape has changed significantly since 2018 (v1). Held — by his own admission the frontier moved so far that his company left it.
• 40% displacement in 15 years is aggressive — economic history suggests slower absorption even with powerful technologies (v1). Held on the data through 2026-09 (4.1% unemployment; hiring-margin effect only).
• Less engaged with alignment/safety questions — the frame is economic and geopolitical, not existential (v1). Held — dated search found two one-liners in the window: "I'm less afraid of AI becoming self-aware and causing danger for humans in the short term" (TED AI, 2025-10-21) and "Move fast, but make sure safety keeps pace" (Semafor, 2026-09-18). Both are the stance that spares deployers; per the Karp rule, recorded as position-with-incentive, not belief.
• "Data is the new oil" metaphor may be dated — synthetic data and small-data techniques reduce the raw-data advantage (v1). Held — and the author has stopped saying it (see Key claims).
• NEW: the jobs number exists in two versions (40%/15 years; 50%/10 years) and he has affirmed both; the file now carries both with dates, and the nearer one is tested first.
• NEW: the 2026 Bloomberg Weekend interview — his most complete current statement — is paywalled; the six-month gap, 1–3% GPU, "20 people and 100 AIs" and "within five years" figures come from two dated write-ups, not the primary. Treat those numbers as provisional until the primary or a transcript is read.
• NEW: every 01.AI number (240 staff, half outside China, bookings 3–4x, profitability 2027) is self-reported; the prospectus will be the first audit.
• NEW: the Capgemini conversation carries no date on the page; its content (fine-tuning economics, "a few million dollars") places it after the March-2025 pivot — that placement is this file's inference.
• NEW: the "export controls do not work" claim has no observable attached by him; the file supplies one (domestic-chip flagship training; Nvidia China revenue).

Rich's take

• (your synthesis here)

Delta log

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

• Grades: 9 Held / 4 Drifted / 1 Wrong (plus 2 unfalsifiable items noted, not graded, and 5 new live predictions with checks). The nine Held are the four Key claims above, two predictions, one Track-record entry (the 2018 book's direction) and two Weak spots re-graded as still true (dual role; thin safety engagement).
• Wrong: v1 Track record "founded 01.AI (2023) — building a Chinese frontier model company" — pre-training stopped March 2025 (FT 2025-03-22 via Wikipedia); "no longer positions itself as a maker of frontier large language models" (Bloomberg, 2026-07-20); the June-2026 v1 review carried the claim 15 months after it stopped being true.
• Drifted: "data is the new oil" (not restated since 2018–19; his 2025–26 mechanism is cost and engineering — Capgemini; NYU 2018-10); four waves (superseded by his own framings); 40% jobs in 15 years (re-dated to 50% by 2027 — Fortune 2024-05-25 — then softened to reskilling — Semafor 2026-09-18; market data: BLS 2026-09-04, Canaries rev. 2026-08-12); three-survivors call (TechNode 2025-03-21 vs Reuters via TechStartups 2026-09-03).
• Held: China leads deployment (IFR 2025-09-25; a16z 2025-08-27); multi-dimensional race (AI Index 2026); China ahead in open source (VentureBeat 2025-10-22; AI Index 2026) — valence logged as negative in his own words; US monetizes first (AI Index 2026: $285.9B vs $12.4B); 2018 applications-parity prediction; three-to-six-month gap (Bloomberg transcript 2025-11; AI Index 2026: 2.7%); weak spot "less engaged with safety" (two one-liners in window).
• Most surprising delta: the "China can compete at the frontier" author's own frontier company quit the frontier — and he then wrote the book on why everyone else should build on top of someone else's model. Says (China's labs are six months behind and closing) and Does (01.AI is model-agnostic and sells the layer above) now point the same way, which is coherent — but it means the bench's "skin in the game" credit was attached to a bet he had already closed.
• Employer/role check (mandatory): titles unchanged (Chairman & CEO Sinovation; Founder & CEO 01.AI — Sinovation, 01.AI); entity check: 01.AI still exists but changed thesis, headcount (~240), product line (TrueNorth, July 2026) and capital path (pre-IPO round; HK 2027); new hat: author of AI Native (Crown Currency, 2026-09-15 / 2026-11-03). Post-review sources found: 2025-10-22, 2025-11, 2026-03-31, 2026-07-20, 2026-09-04/05, 2026-09-18 — the June-2026 v1 pass had none of them.
• Sibling leads: (a) Graylin convergence on deployment-not-frontier — confirmed, with a new split on open-source valence; (b) Naam energy/grid intersection — not confirmed: no Lee statement on China's electricity advantage exists in the fetched record; the only "electricity" quote is a 2018 metaphor. Dropped.
• Corrections to v1 cross-references: the Acemoglu overlap was mis-stated (he does not predict large-scale displacement); the Sources entry "VentureBeat (2025)" is now dated 2025-10-22 and tied to TED AI (2025-10-21); "Bloomberg interview (2025-11-18)" confirmed as a Bloomberg TV segment of that date (page fetch blocked; transcript via third party).
• Incentive map re-derived: two vehicles → five (01.AI pre-IPO, Sinovation, book, geography, Alibaba Cloud as investor-and-lab-partner). Discount raised.
• Tier: quarterly — a pre-IPO round timed to first annual results (early 2027), a book launch (2026-11-03), a chip-self-sufficiency clock and a consolidation call being tested by Moonshot's filing all land inside two quarters. Wake triggers cover each.
• Confidence unchanged at medium: the central 2026 interview is paywalled and reconstructed; company numbers are self-reported; but the grades rest on neutral market data (IFR, Stanford, BLS, Canaries, a16z) wherever a market-level claim was made.
• Search budget note: the session-wide search cap was reached late in this run; four planned market-data searches were replaced by direct fetches of the primaries (BLS, Stanford Canaries, IFR, a16z), which succeeded.

Sources

• 2026 primary / near-primary: Semafor — "Why Kai-Fu Lee thinks companies need an AI boss" (2026-09-18, Andrew Edgecliffe-Johnson) · Penguin Random House — AI Native: The Mandate to Transform Your Company (Crown Currency; 2026-09-15 ebook/audio, 2026-11-03 hardcover) · Bloomberg Weekend — "China Will Win the AI Race for Reach" (2026-09; paywalled) and Bloomberg video (2026-09-04), reconstructed from StartupHub (2026-09-04) and ejshin AI Notebook (2026-09-05) · Bloomberg — "AI pioneer Kai-Fu Lee's startup targets Hong Kong IPO next year" (2026-07-20, syndicated via Yahoo SG); Bloomberg original (paywalled) · TNW — 01.ai heading for a Hong Kong IPO (2026-07-20) · StartupHub — WAIC remarks (2026-07-20) · Bloomberg China Show — on OpenClaw (2026-03-31; fetch blocked) · 01.AI — About (TrueNorth July 2026; WorldWise March 2025; Yi models) · Sinovation Ventures — Team
• 2025 primary / near-primary: Bloomberg TV — Sinovation Ventures' Lee on China-US AI race (2025-11-18; fetch blocked), transcript via mvark (2025-11) · VentureBeat — TED AI 2025 remarks (2025-10-22; event 2025-10-21) · Bloomberg — "The U.S. and China Are Winning Different AI Races" (2025-09-11) · TechNode — endgame of China's AI models (2025-03-21); Bloomberg original (2025-03-20; paywalled) · KrASIA — Lee on 01.AI's pivot (2025-01-09) · Yicai — denial of pre-training team sale (2025-01-07) · Capgemini — "Open vs. closed AI models: a conversation with Kai-Fu Lee" (undated)
• Market-level data used for grades: Stanford HAI — 2026 AI Index Report · IFR — World Robotics 2025 (2025-09-25) · a16z — Top 100 Gen AI Consumer Apps, 5th ed. (2025-08-27) · BLS — Employment Situation, August 2026 (2026-09-04) · Brynjolfsson, Chandar, Chen — Canaries in the Coal Mine? (revised 2026-08-12) · TechStartups citing Reuters — Moonshot IPO filing; Z.AI and MiniMax listed (2026-09-03)
• Jobs number primaries: CBS 60 Minutes release (2019-01-09; aired 2019-01-13) — "40 percent … in 15 years" · Fortune — "50% by 2027 … uncannily accurate" (2024-05-25)
• Books: AI Superpowers: China, Silicon Valley, and the New World Order (Houghton Mifflin Harcourt, 2018-09-25 — Wikipedia; book site); AI 2041: Ten Visions for Our Future (2021, with Chen Qiufan); AI Native (2026) — v1 listed the first two
• 2018-era primaries (for "last restated" dating): Noema — "The Great AI Duopoly" interview (2018-09-24) · NYU Tandon Dean's Lecture (2018-10-09) · NCUSCR podcast (undated) — "The U.S. is stronger in research; China is stronger in implementation"
• Biography / company reference: Wikipedia — Kai-Fu Lee · Wikipedia — 01.AI · Wikipedia — Sinovation Ventures · 01.AI (v1) · Asia Society Australia — "The Bigger Picture" event (fetch blocked; date unverified)
• Talks: TED, Stanford, World Economic Forum; frequent media appearances (v1)

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