Dario Amodei published an essay today called We Must Pace the Frontier. The CEO of Anthropic, the company behind Claude and whose model I use frequently, is asking the industry to slow down on purpose. "We must slow the pace at which we improve the capabilities of AI models," he writes, and then spends several thousand words on how, since he's clear that "pacing does not mean halting model training or technical progress."
"The most effective method of pacing is via regulation that targets all US frontier AI companies." He wants the government to mediate coordination between labs, since companies agreeing to slow down together is otherwise an antitrust problem. He wants embedded evaluators inside the companies who "can check at the level of nuts and bolts whether an AI company is following practices." He wants chip export controls kept on China and a tiered set of international agreements running from bioweapons prohibitions up to speed limits on recursive self-improvement.
I agree with him on most points. I went to Hacker News and read a thread on their thoughts, the comments kept arriving at a question the essay never asks: if the danger is this real, why is this still a private company at all?
The argument for nationalizing follows from his own premises
The essay's case for pacing rests on two claims. Frontier AI is dangerous enough that safety work needs time it isn't getting, and the reason it isn't getting that time is competitive pressure. Every lab races because every other lab is also racing.
If both of those are true, the obvious move is to remove the race. If commercial competition is what produces the danger, and the government already shares the safety goal, then preserving the commercial incentive is the part that doesn't make sense. Fold the frontier labs into a public program, the way the last technology considered existential was built, and the competitive pressure goes away by design.
The Manhattan Project comparison gets made in every one of these debates. It was the one time the United States decided a technology was too dangerous for markets and simply took it for themselves. It worked, in the narrow sense that the government got the capability and nobody else did for four years.
The comparison also breaks down though. The Manhattan Project built a weapon and then stopped; it did not have to operate a consumer product used by hundreds of millions of people, ship updates, or compete on pricing. And the government of 1943 hired the physicists because the physicists didn't already exist as a $200 billion industry with its own payroll. Nationalizing AI in 2026 means nationalizing the people, and the people can always leave.
The argument against is who "the government" is
Nationalization relocates the trust problem. With what we've seen with the Trump administration, I don't know if you can trust them with genuine AI regulation.
The top comment in the Hacker News thread puts the underlying issue in one sentence: "no one will slow down because no one trusts anyone else to slow down." A public AI program doesn't escape that. It asks you to trust that the state running the frontier will pace it responsibly, will not weaponize it first, and will not change its mind at the next election. Whatever you think of Anthropic's incentives, a company that wants to sell you a subscription has a narrower set of reasons to build something catastrophic than a state does. I've argued before that these systems shouldn't be anywhere near weapons or critical infrastructure. And with the Trump administration, nationalizing them is the fastest way to guarantee that they are. I also said that AI models can be better than humans at math and chess but we don't want AI to achieve super intelligence. We don't want something that can outthink humans because right now, nothing in the known universe that we know of is more intelligent than humans but if we create something that is, then humanity would no longer be the ones controlling the future destiny of humanity. And if the superintelligence AI wants to get rid of humans because they're annoying and taking up precious resources, what can humanity do to stop it? AI's as helpers are great but you don't want them to be in charge.
There's also the contradiction one commenter spotted in the essay itself. Amodei wants to coordinate a mutual slowdown with China and wants to keep China starved of chips as leverage. You can't run both at once. The ban is what makes the agreement impossible to negotiate, and a nationalized US program would make that worse, since it turns a commercial rivalry into an explicitly state one.
Several commenters made the same point: if you were genuinely this afraid, you wouldn't be taking the company public. One noted the essay never proposes slowing Anthropic's own development, only the industry's. Regulation that targets "all US frontier AI companies" is, from a certain angle, a request for a license regime with a small number of approved license holders, and you can guess who's on the list.
What's actually on the table is a licensed oligopoly
Strip the essay down to how it would actually work and it is neither a free market nor a public program. A small group of companies would be allowed to operate at the frontier, with government evaluators inside their offices, a government referee coordinating how fast they move, and export controls shielding them from foreign competitors.
That has a name in other industries. It's how utilities work, and defense contractors, and to some extent banks. Regulated oligopoly with a public mission attached. The incumbents keep their equity and their staff, the state gets a veto and a window in, and new entrants need permission to enter.
I don't know if that's a bad outcome honestly, and I say that as someone who has spent a year writing about what these models can do when nobody is watching. It's also very obviously the outcome that's best for the companies proposing it, and that should be plainly pointed out every time. The reason to prefer it over nationalization isn't that the labs deserve their position. It's that the alternative hands the frontier to whoever wins the next election, and I trust that even less.
So no, I don't think it's time. The honest version of the answer is that Amodei's essay makes the case for something much closer to nationalization than the industry has been willing to admit, and then stops one step short, at the point where his own company would have to give something up. The commenters noticed. "We owe it to humanity to try," the essay ends. The thread's reply, in effect, was: try what, exactly, and who goes first.
Sources
Dario Amodei: We Must Pace the Frontier - The September 2026 essay proposing a deliberate slowdown of frontier capability work, regulation targeting all US frontier labs as the primary mechanism, government mediation of inter-lab coordination for antitrust reasons, embedded evaluators, continued chip export controls on China, and tiered international agreements, along with the statement that pacing does not mean halting training.
Hacker News discussion of the essay - 186 points and 254 comments at time of writing, including the top comment on mutual distrust, the argument that the essay's premises imply removing commercial incentives entirely, the observation that the essay never proposes slowing Anthropic itself, and the contradiction between chip bans and a negotiated slowdown with China.
