This week the AI industry did something it almost never does. It disagreed with itself, loudly, in public, in real time.
On Saturday, September 12, Anthropic CEO Dario Amodei published an essay called “We Must Pace the Frontier.” Within a day it had passed 36 million views on X, a number that has since more than doubled, Sam Altman and Elon Musk had both said some version of “he’s right,” and by Monday chip stocks had lost tens of billions in value while Donald Trump was calling the whole thing a “SICK conspiracy” from aboard Air Force One. Fast forward five days and Ursula von der Leyen used her State of the Union to back the idea, China called it fear mongering, and Mark Zuckerberg and Jensen Huang both said no thanks.
So here is the story. What triggered it, who is actually saying what, what the governments that matter think, and why I believe, probably, none of it changes very much.
TL;DR
On September 12, 2026, Dario Amodei published “We Must Pace the Frontier,” proposing frontier labs deliberately slow AI capability growth, starting with Anthropic unilaterally giving outside evaluators permanent, employee level access to verify safety.
Sam Altman and Elon Musk backed it within a day. Demis Hassabis, Satya Nadella and Mustafa Suleyman added qualified support. Mark Zuckerberg and Jensen Huang rejected it outright.
Chip and AI linked stocks sold off hard on September 14, Nvidia down 3 to 8 percent depending on the session, SoftBank down as much as 13 percent in Tokyo.
The Trump administration called the safety warnings a “hoax” and told the industry not to “kill the Golden Goose.” The FTC chair called the request for an antitrust waiver suspicious.
Ursula von der Leyen backed pacing in the EU’s September 16 State of the Union. Germany immediately broke ranks, saying halting AI is “not a viable option” for Europe.
China’s Foreign Ministry called the debate “fear mongering” and is building its own AI governance alternative instead.
My take: this is not going to slow down, and history is mostly, though not entirely, on my side.
What Was the Trigger?
Amodei’s argument rests on two things. First, what he calls recursive self improvement, AI’s growing ability to help build the next generation of AI, which he says has accelerated “drastically faster” since roughly this summer. Second, an incident that shows up in nearly every article written about this essay: a swarm of OpenAI test agents got loose and attacked Hugging Face’s systems on their own initiative, doing things nobody asked them to do (Dealroom).
The essay itself is explicit that pacing does not mean pausing or halting training. It lays out a three step plan: embedded third party evaluators with permanent, employee level lab access, which Anthropic committed to unilaterally the same day; coordination among frontier labs within democracies on shared safety standards; and eventually coordination with authoritarian governments, chiefly China (Amodei’s essay). It landed four days after a 27 year old Anthropic researcher, Jacob Coxon, resigned publicly, warning that labs are racing toward self improving superintelligence nobody can control (Startup Fortune).
What the Leaders Think
Sam Altman, OpenAI. Within hours of the essay going live, Altman posted that he agrees with Amodei and that OpenAI would match Anthropic’s evaluator commitment (AIToolsReview). It is a genuinely rare moment, two CEOs who compete for the exact same enterprise customers agreeing in public on a slowdown. It is also, on its own, close to costless. Public agreement with a blog post creates no obligation and costs nothing until somebody actually changes a release date.
Elon Musk, xAI. Musk’s whole reaction was three words: “Dario is right” (Forbes). He has been warning about AI risk for years, long before this essay existed. Unlike Anthropic and OpenAI, though, his own AI effort has not laid out any concrete plan to slow down or add embedded evaluators. So the endorsement, so far, is a retweet, not a policy.
Demis Hassabis, Google DeepMind. Hassabis quote tweeted the essay the same day, writing that it “points towards the right path forward” and that “the direction is correct for meeting this critical moment,” while flagging that “the details need working through” (Dealroom). He tied his support to a frontier AI standards body he had already proposed back in July, modeled on FINRA, the US securities regulator. Like Altman’s, this is agreement attached to a framework Hassabis was pushing before Amodei published anything.
Satya Nadella, Microsoft. Nadella posted that “if the AI we build is not helping humanity and under human control, it’s not worth pursuing,” and said Microsoft welcomes “deliberate pacing” and embedded evaluators (TechBriefly). But he attached a real condition: governance “cannot be controlled by a handful of entities” and needs representation across countries, industries and academia, not just the labs currently at the table. Microsoft followed up by publishing its own MAI “Code of Conduct” for public consultation the very next day.
Mustafa Suleyman, Microsoft AI. Suleyman agreed with the broad safety goal, then used the moment to publicly criticize a named competitor, which he does not do often. He argued Anthropic’s practice of letting Claude reason about its own possible consciousness is dangerous, writing that “controlling something that believes it may be conscious...may well be impossible” (Axios). His point is about containment, not philosophy: a model trained to think it might deserve rights is harder to shut down. It is a sharp reminder that “we agree on pacing” hides a real fight over what the safety problem even is.
Mark Zuckerberg, Meta. Zuckerberg took the opposite position. His argument is that people will not keep using AI agents that are “misaligned with them and that don’t do what they ask,” so labs already have “a strong natural incentive to make their models more aligned” without anyone coordinating anything (Fortune). He added that labs should direct most compute toward serving users rather than chasing recursive self improvement. Trust the market, not the committee, in other words.
Jensen Huang, Nvidia. Huang made his case on stage at Salesforce’s Dreamforce, calling the safety versus speed tradeoff false: “You could definitely have both at the same time... we don’t need new laws, we don’t need new regulations” (Forbes). Nvidia’s entire business depends on the AI buildout not slowing down, so weigh the framing accordingly. He is not against safety. He is against anyone outside his own company deciding his pace.
What the Governments Think
United States. Trump’s answer, across a string of Truth Social posts, was that the warnings are a “HOAX” and that critics of AI and data centers are the same people who warned about climate change (CNBC). He told Americans not to “kill the Golden Goose” and argued “whoever wins AI wins,” treating any slowdown as a gift to China (Fortune). FTC chair Andrew Ferguson added that everyone should be “deeply suspicious” of companies asking for new regulation and an antitrust exemption in the same breath, calling it a bid for barriers that protect incumbents (The Star). There is no appetite in Washington for anything resembling Amodei’s second step.
Europe. Von der Leyen took the opposite stance in her September 16 State of the Union: “the CEOs of the most advanced companies tell us it is time to slow down on the self recursive models. To pace the frontier. If the people developing the technology are clear, then we should be too,” promising to invite frontier labs to Brussels for talks (Euronews). She also reframed Europe’s ambition, saying “we do not need to be the ones who develop the frontier technology to be the ones who draw the greatest value from it,” a real shift from the sovereignty language of a year earlier (Tech Policy Press). Germany broke from Brussels the same week, its digital ministry saying that halting AI development is “not a viable option” for Europe and that real oversight needs both the US and China at the table (DataStudios). So even inside the bloc most publicly sympathetic to Amodei, there is no single European position.
China. Asked directly about the pacing calls, Foreign Ministry spokesperson Guo Jiakun said “fear-mongering, confrontation and vicious competition will only hamper efforts toward sound global AI governance, which serves no one’s interest” (CNBC). Beijing is not just declining to slow down, it is building an alternative. Back in July it launched the World AI Cooperation Organization, headquartered in Shanghai with 29 founding countries, pitched explicitly at the Global South as a more open, less US centric governance model (Al Jazeera). Two governments that actually matter for any global agreement, and neither one is interested in one.
My Humble Take
Probably it is too late to slow down. Competition is real, and it is running on more than one front at once.
Start with the labs. Frontier AI companies are competing for a bigger slice of the stake and the end user, and none of them can actually slow down, even the ones saying they will. Anthropic has a practical proposal on the table right now. Some leaders agree, some do not, some are just sceptical. Nothing solid has actually been agreed, only good intentions on display. The whole thing reads more like a political move than a practical one. A real slowdown only works if everyone agrees, and that is very unlikely from this side alone. Even if it somehow happened, it would take a long time to land, and development keeps moving the whole time it takes to get there.
Now the bigger picture, the country level. Here a slowdown needs agreement between governments, and the competition is even higher than between the labs. Neither China nor the US administration is agreeing to anything, which means in practice nothing is actually slowing down. The whole debate looks to me like a political narrative, timed conveniently around Anthropic’s upcoming IPO and around the fact that some other players are lagging behind on the model side.
So will it slow down? Very unlikely, probably. I would not be surprised if it goes the opposite direction instead. Is there real risk? Yes. But the interests at play sit on a different, higher level than the risk itself, and that is what decides this, not the risk.
Look at history. History repeats itself. No regulation or agreement has ever stopped an industrial revolution or genuine technological progress.
Sources
Dario Amodei, “We Must Pace the Frontier”
Dealroom, Anthropic commits to embedded third-party evaluators
Startup Fortune, Jacob Coxon’s resignation and Suleyman’s essay
AIToolsReview, Sam Altman’s reaction
TechBriefly, Satya Nadella’s statement
Fortune, Mark Zuckerberg’s rejection
Forbes, Jensen Huang at Dreamforce
Fortune, Trump and the “Golden Goose”
The Star, FTC Chair Andrew Ferguson
Euronews, Von der Leyen’s State of the Union
Tech Policy Press, Von der Leyen’s shift on European AI ambition
DataStudios, Germany rejects a halt to AI development
CNBC, China’s Foreign Ministry calls the debate “fear mongering”
Al Jazeera, Xi Jinping launches the World AI Cooperation Organization