On the morning of September 14, at the All-In Summit in Los Angeles, Jensen Huang was on stage discussing “We Must Pace the Frontier,” the essay Dario Amodei had published two days earlier, when his phone rang. The caller was Trump. According to TechCrunch’s report from the room, Trump said robots are not going to take over and called the slowdown worry a hoax: “They’re just playing right in the hands of a lot of people that don’t want to see it happen. That could be political people. It could also be China. And we’re not going to let that happen. It’s a hoax.” He added: “We have to be a little bit careful … we have to do things and we have to do them prudently, but that doesn’t mean we’re going to stop an industry.” Huang’s answer: “You’re right. We’re not going to let that happen, sir.” The room applauded.

The call put on the record what the preceding 48 hours had already sketched. Amodei’s pacing proposal drew responses in an unusually clean formation: the other lab chiefs who responded, Altman, Musk, and Hassabis, all agreed, and the noes came from the company selling the compute, from the president, and from one of his advisors.

The 48-hour scorecard

First, the proposal itself. On September 12, Amodei published the essay on his personal site (TechCrunch’s same-day report). His plan for slowing the growth of frontier model capability has three steps. Step one: each frontier company gives a team of embedded third-party evaluators ongoing, employee-like access (desks, badge access, the freedom to publish what they find), with Anthropic committing unilaterally and immediately. Step two: frontier companies in democratic countries coordinate common safety standards and rate limits, shielded by a government antitrust exemption. Step three: democratic governments negotiate with authoritarian ones, starting with bans on biological-weapons uses and extending, in the most ambitious version, to a “speed limit” on recursive self-improvement. I broke down the mechanics in that day’s briefing, so I won’t repeat them here.

Then the responses, in order:

  • September 12, same day. Sam Altman posted: “I agree with Dario that we need to pace the frontier … we will do the same” (quoted here via Zvi Mowshowitz’s roundup, as are the next few). Note the second half. That is OpenAI matching step one, a commitment with specific content that can be checked later.
  • Elon Musk’s first response ran three words: “Dario is right.” Demis Hassabis said the direction is correct. Bernie Sanders said pacing is too little and demanded a full pause: “When you are racing towards a cliff, you don’t just ease up on the gas pedal. You hit the brakes.” One line of criticism, in other words, came from the too-mild side.
  • September 13. David Sacks, formerly the White House AI czar and now co-chair of the President’s Council of Advisors on Science and Technology (his March move, per TechCrunch), responded on X: “go ahead. You guys are the frontier.” Slow yourselves down if you want, he argued, but “stop pretending you need anyone else’s permission,” and don’t ask for antitrust law to be suspended “so you can form a cartel.” As I wrote yesterday, this is not a formal government decision, but it turns away the parts of the plan that need government cooperation.
  • September 14. The scene at the top. Per TechCrunch, Huang took the opposite side from Amodei in the on-stage discussion; the statement that got fully recorded is the one he gave Trump: “We’re not going to let that happen.”

As of publication, I have not found any public response from Meta; none appears in Zvi Mowshowitz’s roundup of reactions either. The three other lab chiefs who spoke, Altman, Musk, and Hassabis, all accepted the premise of pacing. The noes on record came from a compute supplier and from Washington.

Huang’s quarrel is with the order book, not the safety essay

The formation needs no conspiracy theory. Look at what each company sells.

Model labs sell access to capability. A symmetric rate cap, applied to everyone, freezes relative position: if you slow down and I slow down, nobody gives up their place in line. That is the design premise Amodei states repeatedly in the essay: pacing must keep democratic countries’ lead over authoritarian ones as large as possible. Meanwhile the costs of racing, the compute burn, the liability and reputational damage when something goes wrong, sit on the labs’ own books. Follow that structure (this is my inference; nobody would say it out loud) and a cap that binds all your rivals is not obviously bad business: burn rate down, ranking unchanged, tail risk smaller.

Nvidia’s books point the other way. It sells the race’s input. Amodei’s essay proposes pacing by inputs and names three things to limit: training compute, the nature of training runs, and the internal use of AI to improve AI. The first item is Nvidia’s product. If the whole industry coordinates a slowdown, the first thing squeezed is Nvidia’s order growth. And the exposure runs beyond future orders. Nvidia has been putting its own money behind its customers since at least April 2023, when it joined CoreWeave’s Series B round (TechCrunch), and the commitments have grown with the boom: a pledge to backstop up to $6.3 billion of CoreWeave’s unsold compute through 2032 (SEC 8-K), and an announced plan to invest up to $100 billion in OpenAI, paid in progressively as 10 gigawatts of Nvidia systems get deployed (official announcement). The Economist called Nvidia the central bank of AI earlier this month (my breakdown): the capital it has committed is a bet that its customers keep expanding. For a balance sheet like that, “pacing” reads as a systemic-risk word. That, at least, is how I read it.

There is a second front. Amodei’s essay restates his export-control position: “Do not sell powerful AI chips or semiconductor manufacturing equipment to China, and crack down on chip smuggling operations …” Huang’s position on that question is equally public. At Computex in May 2025 he said “export control was a failure,” and gave his own estimate that Nvidia’s share of China’s AI chip market had fallen from roughly 95% in 2021 to 50% (CNBC).

So the September 14 line was not improvised. In June 2025, at VivaTech in Paris, Huang said he “pretty much disagree[s] with almost everything” Amodei says, and compressed Amodei’s position into three claims: AI is so scary that only they should build it, so expensive that nobody else can, and so powerful that everyone will lose their jobs, “so only their company should do it.” His own counter: “If you want things to be done safely and responsibly, you do it in the open. Don’t do it in a dark room and tell me it’s safe” (Fortune; Anthropic denied at the time that Amodei ever claimed only Anthropic should build AI). Fifteen months on, judging by the public statements, the disagreement sits exactly where it was. What changed is that this time a president phoned in to pick a side.

The incentive map fails on Musk

Reducing every position to a sales relationship has one obvious counterexample: Musk. xAI is a heavy chip buyer. Its Colossus complex in Memphis already runs on hundreds of thousands of Nvidia GPUs and is building toward a footprint that one industry tracker puts at over 555,000 (Introl’s tally, an estimate assembled from public reporting rather than an official count). If order books decided positions, Musk belongs on Huang’s side. And when the Pacing the Frontier open letter circulated in July, the official signatory list included employees of OpenAI, Anthropic, Google DeepMind, and Meta AI, with xAI absent (my analysis at the time, in Chinese). This time he said “Dario is right.”

My reading of the counterexample: it prices the statements. “Dario is right” is three words, commits xAI to nothing, and is fully compatible with the AI-risk rhetoric Musk has used since 2014 (the Washington Post’s report on his “summoning the demon” remark). Saying it costs almost nothing. The only part of Altman’s statement that carries weight is the half-sentence with content: “we will do the same.” In July I wrote about why that open letter could gather over a thousand signatures: it downgraded the ask to “preserve the option to slow down,” which made signing nearly free. What Amodei’s essay does, precisely by moving one step further, is attach a price to agreement. Embedded evaluators take up desks, get badges, and keep the right to publish. Once a price exists, verbal agreement and paid agreement separate on their own.

What is left of the plan

Score the three steps against these 48 hours. Step three, international coordination, runs through the White House, and the president just called this class of slowdown worry a hoax on a live phone call. Step two depends on an antitrust exemption that the PCAST co-chair has publicly spoken against. Neither statement is a formal government decision, and neither step is formally dead, but the wind is not hard to read. What remains fully in the labs’ own hands is step one, the step that requires nobody’s permission: Anthropic’s unilateral commitment, now joined by OpenAI’s promise to match it.

That is the field-test answer to the question of what a unilateral initiative can do without enforcement power. It stops none of its opponents. But within two days it got Altman, Musk, Hassabis, Sacks, Huang, and finally the president to show, in public and in sequence, where each of them stands relative to “speed.” The July letter registered who is worried. This round put on the table whose business is staked on continued expansion. Both records are permanent.

The checkpoints worth watching have narrowed to three. Do evaluators from organizations like METR actually get desks and employee-level access inside Anthropic, and when does the first independent report appear without company editing? Does OpenAI’s “we will do the same” turn into access at the same level? And does any of the verbal yeses, xAI first among them, follow up with a single action that costs something? If the brake ever gets built, the first press has to happen inside the labs’ own buildings. Huang and Trump spent these 48 hours saying, on the record, that it will not happen in theirs.

References