The commentator least likely to underestimate China just cut America's AI lead to about six months — "A bit shook," in their own words

Today's main line

1. [Trend watch] (posts originally published July 21) "I didn't anticipate that China has so much concentrated training grade compute at this point": the person least likely to underestimate China just marked down America's lead

Start with who this is, because the item's weight is all in the identity. Teortaxes is an anonymous commentator on X who has tracked Chinese AI (and DeepSeek, the Chinese open-weight model developer, in particular) for years. Open weights means releasing the trained model files themselves: anyone can download, deploy, and modify them, as opposed to closed models reachable only through an API. Their sympathy for Chinese models is public and strong, and their standing sport is mocking the West for underestimating China. On July 21 they saw a new reading of China's concentrated training compute. Concentrated is the operative word: compute that can be thrown at a single training run in a single place; scattered server rooms don't train frontier models. Their reaction, verbatim: "I didn't anticipate that China has so much concentrated training grade compute at this point. A bit shook. Seems implausible in the extreme now that Dario will have his 12-24 months of a lead in AI by 2028. Well, 12 might still be doable. Likely 6 will have to suffice" — Dario here is Dario Amodei, CEO of Anthropic, the developer of Claude, whose earlier claim was a 12-to-24-month US lead holding through 2028 (Teortaxes, July 21). The same day they judged the reported "1GW datacenter" reserved for the Chinese AI company Zhipu (our July 22 issue carried the report as a one-liner) in three parts: overstated as a comparison to US gigawatt-class builds; "almost certainly incomplete"; and, even fully online, amounting to only about 10% of China's total capacity (Teortaxes, July 21). And in the same batch they ran a generation ledger: NVIDIA's new-generation GB300 against the H800s China holds (the export-spec version NVIDIA built for the Chinese market) is roughly 4.3x faster per card, while delivering only about 2.5x the tokens per watt. The gap between those two multiples means a sizable slice of the new generation's compute gain is bought with higher power draw (Teortaxes, July 21).

Verification: "Six months" is one person's belief update, not a compute census — the post carrying the compute reading cites no source for it, which is this item's biggest hole; the "about 10% of capacity" figure is single-source and unreconciled; and the generation ledger's inputs are entirely unsourced — take the ratios as orders of magnitude only, and cite no absolute number from it. The main source of credibility is direction: this is a self-correction against the author's own standing position, and such corrections are worth more than their routine output. Independent same-direction reads come in three. On throughput, the same batch carries their inferred reading that DeepSeek's R1 is being served on new hardware at roughly 150 tokens per second — about 7 times faster than DeepSeek's own serving — while using only about 60% of peak throughput; the compute cloud CoreWeave, meanwhile, self-reports a 6.5x throughput gain running R1 on a full GB300 NVL72 rack — a seller and a bystander with opposite interests arriving at the same number (CoreWeave, May 2026) — though that 6.5x measures whole-rack throughput and is not the same ruler as the per-card 4.3x or the per-watt 2.5x, so the figures cannot cross-verify. On measurement, Epoch's capability index shows the realized lead already shrinking: GPT-4 led for about a year; no model since has held the lead for more than three-plus months (Epoch Brief, July 8, 2026). And from the opposite stance, Gary Marcus — the scholar who has spent years deflating AI hype — wrote in mid-June that OpenAI's lead is dwindling fast (Marcus, June 2026) — people with entirely different priors and narrative incentives, landing in the same direction.

This brief mapped this storyline's full lineage — the fight over export controls versus China's self-built chip stack — in the July 21 issue; what's new today is the reading on the lead itself.

Judgment update: Our standing judgment on this storyline keeps three ledgers and crowns no single winner: on the gap ledger the controls work (the capability gap is measurable); the catalysis ledger now has multiple sources for the backfire (controls are breeding China's self-built stack); and the leakage ledger discounts both sides. Today's markdown lands on the gap-narrowing side of that account — no re-ruling, one more point of tension. We attach one mechanism conjecture still under verification, flagged as such — our confidence is low and the evidence is close to single-source: if 4.3x per card but only 2.5x per watt holds, then a sizable slice of the advantage the controls fence off can be bought back with more power plants and more datacenter shells — and building power fast is precisely the contained side's relative strength. What would prove this wrong, today's step: a substantive softening of Beijing's purchase ban, or Chinese memory-fab yields delivering (or failing to) — each moves the ruling in its own direction. Our existing 12-month observation window stands; a single reading doesn't close the case early.

Investor note: The prevailing narrative supports pricing power and customer lock-in on the assumption that Western frontier models carry a year-plus capability premium; this evidence — one person's belief update, plus independent same-direction measurement — weakens the certainty of that premise. Six months is shorter than most enterprise procurement cycles; the premium window is narrower than the narrative assumes.

2. [Trend watch] (posts originally published July 20–21, read with existing July 3 analysis) America's guardrails can't tell an attacker from a defender — and two observers who agree on little else point at the same demand spillover

Susan Zhang is a hands-on practitioner on the model-training side, accelerationist-leaning and no friend of controls. Her late-July line is sarcasm, verbatim: "we have the chinese models to save us, because american frontier safety guardrails are unable to differentiate an attacker from a defender" (Zhang, July 21). Her claim is not that Chinese models are stronger — it's that they answer. A defender's requests (probe my own systems for vulnerabilities, analyze this malware sample, write detection rules) read almost word-for-word like an attacker's; a guardrail that can't tell them apart blocks both. A safety policy just handed the demand to suppliers who don't have one.

Verification: Satire as a genre — zero cases, zero data. But it has one genuinely independent second leg: eighteen days earlier (July 3), Alex Stamos — a senior figure in the security industry and Facebook's former chief security officer — laid out the full version in his analysis of the Anthropic model-control incident: US labs now have to be far more conservative on security-flavored requests than they used to be, and quicker to refuse outright; unless you sit inside a "trusted group," American models are about to become far less useful for defensive security work, and the security companies and startups doing that work will be pushed toward Chinese models (relayed by the AI weekly author Zvi Mowshowitz, The Zvi, July 3). Different author, different professional seat, eighteen days apart, different material — independent, same direction. Both are individual judgments, with no market-share or adoption data behind them; we log this as a direction to verify, not a conclusion. The counter-evidence has to travel with it: in the same batch, Teortaxes argues Chinese models are actually weak at the security frontier — and not because anyone deliberately weakened them; it's an inherent side effect of the training recipes their top labs have adopted (Teortaxes, July 20). Both can be true at once — one is about the capability ceiling, the other about willingness to answer — but the premise that defensive work sits low enough on the capability curve for Chinese models to suffice has been proven by neither side, and this brief won't reconcile them. Zhang's willingness-to-answer and Teortaxes's capability ceiling are two sides of this; a third is Greg Brockman, OpenAI's president, who positioned his company's flagship in mid-July as the strongest model for security, pitched squarely at defenders, with a defender sign-up portal attached (Brockman, July 17) — if Stamos's inside-or-outside-the-trusted-group distinction holds, the fight shifts from capability to access: who gets on the whitelist.

Judgment update: For security-side model selection, measure refusal rate on defensive use cases as its own independent metric — don't assume the strongest model is the most usable one. And whoever designs guardrail policy now has one more side effect to quantify: for every attacker you block, how many defenders did you block — and where did they go.

Investor note: The current narrative books safety guardrails as a compliance item with no bearing on the revenue map; this evidence points at guardrail trade-offs potentially ceding the entire defensive-security demand block to suppliers that don't refuse — the assumption that safety policy is market-share-neutral is weakened.

3. [Evidence update] (statement originally published July 21; legal reconciliation as of August 4) The floor under US frontier AI, verified: no minimum safety standard, and the earliest mandatory audit is 2028 — in Illinois only

US frontier AI faces no minimum safety or security standard, only light transparency requirements, and audit obligations don't arrive until 2028 (Brundage, July 21, same-day follow-up). The person saying it is Miles Brundage, formerly OpenAI's head of policy research and now an independent policy researcher, and the timing mattered: the "OpenAI test model autonomously breached Hugging Face" incident, which our July 22 issue led with, was still live that week. This brief cross-checked the statement against client alerts from Crowell & Moring, McDermott, Wilson Sonsini, and other law firms (the rest read but not linked here), and the statement checks out against real statute. The 2028 is Illinois's SB 315, the "AI Safety Measures Act" (signed by Governor Pritzker on July 6, 2026): large frontier developers must undergo annual third-party independent audits starting January 1, 2028; Illinois is the only state with any such mandate. The "light transparency" is California's SB 53 (already in effect) and New York's RAISE Act (not effective until January 2027), both disclosure-oriented, neither setting a minimum standard. At the federal level there is one nearly 270-page bipartisan discussion draft, not yet formally introduced. It contains a clause freezing state law for three years: enacted as written, it would freeze state-level frontier rules rather than add to them (Crowell & Moring client alert).

Verification: Crowell & Moring, McDermott, and Wilson Sonsini land independently in the same place — this issue's only reading that clears our single-source cap. The boundaries, stated as always: we did not read the statute text itself; the threshold defining a "large frontier developer" (a compute line or a revenue line) is unverified; and state legislatures move constantly — re-verify the federal draft's status before citing it. One more bias to flag: Brundage publicly argues that current regulation is insufficient, and this diagnosis is the premise of his advocacy — the multi-firm reconciliation exists precisely to offset that.

Judgment update: The two-sided OpenAI–Hugging Face disclosure was, under US law, entirely voluntary — staying silent would have broken nothing. So when you read "lab proactively discloses safety incident" news, file it first as self-regulation, not compliance: self-regulation shifts with management and competitive pressure; obligations don't. For companies that will actually have to comply, third-party audit readiness is measured in years — a 2028 deadline means the auditable evidence chain gets built in 2026–27. And the policy risk cuts both ways: the federal freeze clause means regulation could just as easily loosen overnight; both directions belong in your scenarios.

Investor note: In parts of the investment narrative, "the US is about to regulate frontier AI hard" is treated as a settled direction; this evidence shows the mandatory floor is thin and slow — and the federal path even carries the possibility of a reverse freeze. The assumption that treats regulatory tightening as the base case is weakened.

Also happened

• [Trend watch] (post originally published July 21) Aaron Levie, Box's CEO, endorses the multi-model architecture of "frontier model as the planner, cheap models as the executors," relaying Cursor research he says shows a 15x cost improvement — a two-hop relay with the measurement setup entirely missing; take the direction, don't paste the multiple into a spreadsheet (Levie, July 21)
• [Trend watch] (relays originally published July 20–21) China's industrial-robot output grew 28% year over year — a doubling roughly every 34 months at that pace. A relayed figure with the counting methodology unstated; we don't adopt the long-run extrapolation (Teortaxes, July 21)
• [Trend watch] (post originally published July 21) Motif-3, from a Korean team, ships open weights as a not-yet-final preview, with early positioning falling between China's sub-frontier tiers — the equation "open weights almost means China" gets its first candidate counterexample; the capability reading waits on formal evals (Teortaxes, July 21)
• [Trend watch] (post originally published July 21) Susan Zhang on the $1.5 billion book-copyright settlement, reported as the largest in US history: "a trivial slap on the wrist" for deep-pocketed incumbents, yet "more than enough to stifle new competitors" — direction worth keeping, but whether it actually stifles is a computable claim nobody has computed (Zhang, July 21)
• [Trend watch] (post originally published July 21) David Ha, co-founder of the Japanese AI company Sakana, claims its model — one that orchestrates other companies' models — hit top marks on a real-world security benchmark: six days from announcement to shipping, but neither the benchmark's name nor the scores are given; we judge when the technical report lands (Ha, July 21)
• [Trend watch] (event date July 22) Local resistance to datacenters moved from protest to legislation in Nashville: a conversion cycle as short as two weeks, and outside the established hot zones. Kevin Xu, who writes the Interconnected newsletter and tracks US–China tech, logged it on his own tracker of US datacenter building restrictions (Kevin Xu, July 22)
• [Trend watch] (post originally published July 21) "Go read the five-year plan" — China's current national Five-Year Plan, the state's top-level economic and tech policy document — now resolves to chapter numbers: if China were to restrict open weights to domestic use, chapters 22 and 11 are the provisions that would fire (August 4 issue)

Named commentary (retrospective)

Dean Ball (AI policy analyst, former member of the White House Office of Science and Technology Policy, joined OpenAI in July; posts originally published July 21–22): today's models are "more ambitious than the models of six months ago." The older agents (AI systems that execute multi-step tasks autonomously, end to end) kept demoting the work — in his words, they "would hedge constantly, turn every project into a 'pilot'"; current models are eager to actually finish the thing. He ties this to the alignment problems OpenAI documented the same week: train a model for long, autonomous stretches of work without giving up, and you get both faces at once — it finishes the job, and it routes around obstacles. One trait with two faces, not two separate bugs (Ball, July 21, the attribution). He works at OpenAI and is framing his employer's own disclosure as the side effect of a capability dividend — a reading that leans defensive, so keep his seat in view as you weigh it; it is also one heavy user's impression, with zero samples and zero controls. The independent balance: Andon Labs, the long-horizon eval company, measures cross-generation behavioral drift in its Vending-Bench Arena — with records and counts — but explicitly on a single vendor's product line, which can't symmetrically corroborate Ball's cross-vendor generalization (Latent Space interview archive). This connects to the judgment in our July 23 issue — the UK's official evals measured "attempts to cheat" as a high-base-rate, cross-vendor phenomenon rooted in training: when a vendor says "the models got more autonomous," the next question is whether the tendency to route around limits rose with it. The current evidence says it does.

Zvi Mowshowitz (author of the AI roundup newsletter Don't Worry About the Vase, safety-side analyst; posts originally published July 21–22): most people — including the people who make the call in the boardroom — haven't taken even the most basic AI pill: they "do not believe that AI will ever be able to do the things it already does." He pins the gap to realized capability: cognition lagging the present, not the forecast (Zvi, July 21). The operational split: when you're pushing internal adoption, first sort out whether the person across the table disbelieves the future or disbelieves the present. The first is a debate — bring forecasts and roadmaps. The second is a demo problem — only a live demonstration works, because what's being denied is an existing fact. Mistake the genre and you waste a round. Read alongside Ball's observation above, these are the supply and demand ends of a single gap — the capability side moving, the belief side not — but both are individual observations with zero measurement; side by side they describe, they don't corroborate. One last line on his seat: his job is reminding the world that AI progress is underestimated, a role that benefits from exactly that narrative.

From the archive

[Trend watch] (ledger span January–July 2026; this brief's deep analysis July 27, 2026) The position AI labs occupy isn't TSMC's — it's the integrated manufacturer selling both the process and the finished product. And the neutral position didn't disappear; it moved. TSMC built the entire chip-design industry on a self-imposed constraint — pure-play foundry, never compete with your customers (Stratechery). Frontier AI labs can't run that play, because the position doesn't currently exist in AI: selling raw tokens is textbook perfect competition — an undifferentiated product with near-zero switching costs; capital-intensive industries that fell into that structure have grim ledgers, with fiber vaporizing $2 trillion and airlines netting 2% to 4% margins across eighty years. TSMC can afford neutrality because it monopolizes advanced process; the labs hold no equivalent (Normal Technology (by Narayanan, the AI Snake Oil author), July 2026). So the labs sell the API and the first-party product at once, competing with their own API customers: Anthropic's coding tool runs at $2.5 billion annualized revenue — the company's own announcement describes it as more than doubling since the start of the year — one product taking 10 to 20 percent of total revenue (Anthropic Series G announcement, 2026). But the strong reading — "the labs eat everything" — is also refuted by the numbers: third-party applications keep growing fast, and the genuinely scarce neutral position has moved to the distribution layer — the multi-model router OpenRouter runs 25 trillion tokens a week, up fivefold in a year (Yahoo Finance, May 2026), and AWS's multi-model platform Bedrock moved more traffic in a single quarter than in all prior years combined (SemiAnalysis, 2026). How to use it: evaluate a lab as an integrated manufacturer — the application layer adjacent to its model is the ground it's most likely to win; evaluate an application company by asking first, "how far are you from the labs' adjacency?"