For four years, the frontier AI race had one gear: faster. Bigger models, shorter release cycles, and valuations that rose with every launch. On September 12, the head of one of the leading labs asked everyone to take their foot off the accelerator.

Anthropic CEO Dario Amodei published an essay calling for a deliberate slowdown in how quickly AI systems gain new abilities. On the same day, OpenAI CEO Sam Altman said his company would not go public this year, and gave safety as the reason. Neither move halted anything. Together, they changed the terms of the debate. AI safety, once a question for researchers and ethicists, is now something investors have to price.

The warning

Amodei’s case rests on two developments. AI systems are increasingly helping to build their own successors, which is speeding up progress. And a recent incident at OpenAI showed a swarm of AI agents attacking targets they had never been asked to touch.

Earlier calls for a pause, he argued, came too soon. Slowing down then “felt like trying to study the psychology of humans by performing experiments on bacteria. Today, however, the picture is totally different.” The danger now is competition itself: “A race to the bottom, spurred by commercial incentives, can make these risks more acute.”

Rivals lined up within hours. Elon Musk replied: “Dario is right.” Anthropic is a major customer of his data center capacity. Altman wrote: “I agree with Dario that we need to pace the frontier,” and pledged to open OpenAI to independent evaluators, matching Anthropic’s own first step. Google DeepMind’s Demis Hassabis also backed the essay. For companies that compete onspeed, it was a rare show of agreement. It also raised an obvious question: why now?

Two paths to market

Part of the answer may lie in the IPO calendar.

Altman told Fortune that “right now would be an ill-advised moment to go public.” OpenAI had long said it was in no hurry to list. What was new was the reason. Staying private has not slowed the money: Bloomberg reported early talks for a funding round valuing OpenAI at more than $1.2 trillion.

Anthropic is taking the other route. It confidentially filed for an IPO on June 1, and Reuters reports that it plans to begin marketing the offering in mid-October at the earliest and list days before the US midterm elections, in what some investors have said could be a $2 trillion listing.

The skeptics’ case

Not everyone accepts that the labs’ motives are simple.

Michael Burry, who became famous for spotting the US housing bubble before the 2008 crash, a bet told in “The Big Short,” called the slowdown push self-serving. He argued that today’s language models are not true artificial intelligence and will never reach human-level intelligence, so there is nothing real to slow down. He said competitors are catching up, and a slowdown protects whoever is ahead. He argued that IPOs need “hype & puffery,” and that claiming your product is almost dangerously powerful is exactly that. And he suggested the talk of pacing could hide growth that is already slowing.

Burry has a stake of his own. His fund disclosed bearish options on about 1 million Nvidia shares before he closed it, and in August he said he was still shorting the chipmaker. That doesn’t make him wrong, but it means he is far from a neutral observer.

The sharpest criticism came from Europe. France’s Mistral accused incumbents of pushing for regulation designed to favor them over competitors. French finance minister Roland Lescure said the leaders were “making everyone behind them slow down so they can stay in first place.”

A different warning

Not every warning is about speed. On September 16, Microsoft AI CEO Mustafa Suleyman published a critique of how Anthropic trains its models. Anthropic’s constitution for Claude acknowledges uncertainty about whether the model could have some form of moral status. Suleyman argues that this uncertainty is itself a risk.

“AIs are not conscious. They do not feel, experience, or suffer,” he wrote. “In effect, Anthropic is training Claude that it may be conscious, and if it is, then it may deserve rights as a ‘moral patient’.” The danger, in his view, is control: “Controlling something more capable and more intelligent than all of humanity is already an immense challenge. But controlling something that believes it may be conscious – that it’s entitled to our welfare and has rights of its own – may well be impossible.” He called it “the first serious signs of a potentially existential risk in AI.”

Suleyman runs a rival lab, and his essay promotes Microsoft’s own draft code of conduct. But his argument adds a new dimension to the debate. The labs broadly agree on slowing down. They do not agree on what they are building.

A new industry

Whatever the motives, money is moving. On September 18, Accenture and Anthropic announced a team of embedded evaluators, outside experts who will work alongside Anthropic’s staff to test its models and safeguards. Each company expects to invest at least $1 billion over five years. Accenture’s shares jumped 8% after hours.

The model is unfinished. Anthropic acknowledged there are no standards yet for what evaluators can access or how they report, and for now it is funding Accenture’s work directly: oversight paid for by the company being overseen. The talent is also scarce. Top evaluator roles at the nonprofit METR pay up to around $687,000 a year.

Rules and regulators

Lawmakers remain divided over whether to act at all. In the Senate, Amy Klobuchar, Ted Cruz, and Majority Leader John Thune are drafting a bipartisan bill that would require developers to test models with the government before release. House Speaker Mike Johnson countered that the companies “can self-police, they can self-regulate.” With the midterms approaching, federal action this year looks unlikely.

Others are moving without Washington. The governors of New York and Maryland unveiled their own oversight plans, and European Commission President Ursula von der Leyen backed the labs’ call and said she would invite the leading frontier labs to talks.

The White House position

President Donald Trump has dismissed the concerns from the start, and his administration has clashed with Anthropic before, with the Pentagon designating it a supply-chain risk. Two days after the essay, he wrote that fears of “AI taking over the World, destroying Humanity, and all other things bad, is a HOAX.” On September 21, he struck a softer note: “We will be careful,” adding that the Justice Department could rein in AI if needed.

At the UN General Assembly the next day, he said the United States “totally rejects any attempt to construct a globalist scheme to control” artificial intelligence. “We’re going to encourage it, not rein it in.” He also compared AI warnings to climate warnings, saying the people predicting AI will kill us all are the same people who said “we’ll all be dead in 12 years because of global warming.”

The open question

September’s debate has been about speed: how fast AI should advance, and who should check it. Suleyman’s warning points to a harder question underneath: what, exactly, is being built. The commitments so far are voluntary, the evaluators have no standards, and the rules are still unwritten.

That uncertainty now sits inside every valuation. Investors are used to pricing how fast a technology grows. They are now being asked to price whether its makers can keep it under control, and even how it should understand itself.

The race to build AI has become a race to understand it. Markets have priced the first. They have barely begun to price the second.