"We must slow the pace at which we improve the capabilities of AI models," Anthropic CEO Dario Amodei wrote on September 12, 2026, in a roughly 3,800-word essay titled "We Must Pace the Frontier," published on his personal site. "Progress will still seem fast, and we must make wise use of the time we gain."[1] It is an unusual thing for the head of a company that has spent the last two years racing to ship ever more capable models to say in public. Stranger still is what happened next: within roughly a day, the industry's other most powerful executives were saying it too.
Amodei's proposal was explicit and three-part. First, Anthropic would unilaterally give outside evaluators -- naming the AI safety group METR specifically -- "desks in our offices, access badges, and company laptops," with standing rights to publish what they find without Anthropic's editorial control.[1] Second, frontier labs within democracies would agree to common, industry-wide pacing standards applied equally to every company actually at the frontier. Third, democratic governments would eventually try to extend that pacing coordination to authoritarian governments as well.[1]
OpenAI's Sam Altman told staff the company was open to slowing its own frontier work too, and followed with a public post the next day: "I agree with Dario that we need to pace the frontier. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon."[2] Elon Musk, who runs a third frontier lab in xAI, added two words of his own on X: "Dario is right."[2] Three competitors who have spent two years locked in a capital and capability arms race publicly agreeing on anything, let alone agreeing to voluntarily slow down, is not normal industry behavior -- which is exactly what made the pushback that followed notable too.
Late that same Saturday night, David Sacks posted a public rebuttal on X daring Amodei and Altman to slow down on their own, without asking government for anything in return. "Stop pretending you need anyone else's permission," Sacks wrote, adding that the labs should "stop pretending you need to suspend antitrust law in order to create a cartel," and that they should "stop pretending METR is independent of Anthropic when they are literally intertwined with Anthropic's investors and staff" and "stop pretending you need these same people to police your competitors who are not even at the frontier of AI."[3] His underlying argument: model safety is ordinary commercial risk management -- a company that ships a dangerous product already faces real product-liability exposure -- and a leading lab that genuinely believes its next model is too risky to release does not need a government-blessed standards body to justify holding it back.[3]
Sacks' title is worth getting exactly right, because the loose version of it is already circulating. He was Trump's White House AI and crypto czar from the start of the administration, but he stepped down from that specific role in March 2026 after exhausting the 130-day limit on his status as a special government employee, moving to co-chair the President's Council of Advisors on Science and Technology alongside White House technology adviser Michael Kratsios.[4] He is not currently the government's AI czar in name. He remains, by every available account, one of the administration's loudest and most closely watched voices on AI policy -- which is a different, narrower claim, and the one this piece is making.
The administration's actual position arrived a day later and more bluntly, from the president himself. Asked about the pacing calls from Amodei, Altman, and Musk while at his golf course in Ireland on Sunday, September 14, Trump rejected the idea outright: "We're leading China in AI. We're the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI wins." He added: "We can put guardrails. We can do this and that. But I think you have a lot of negative forces that are bringing it up that shouldn't be bringing it up."[5]
The pacing call did not pause Anthropic's own buildout. Since October 2025, the company has signed contracts for 14.8 gigawatts of compute capacity -- roughly ten to fifteen times what it was running as of mid-2025 -- adding up to as much as $517 billion in compute commitments over the coming decade, according to reporting from The Information. Google and Amazon combined account for more than $300 billion of that across roughly 11 GW; Microsoft has separately pledged $30 billion in Azure capacity; SpaceX and other providers make up the rest. Most of it is structured as take-or-pay: Anthropic owes the minimum spend whether or not it uses the capacity.[6] That is not a contradiction of Amodei's essay -- pacing capability growth and building out infrastructure are different levers -- but it is the plain fact sitting underneath a company simultaneously arguing the industry should deliberately slow down.
Altman, separately, gave Fortune a related answer the same week about why OpenAI is staying private a while longer. Asked directly whether an IPO was off the table for 2026, he said: "I would say not 2026, yeah. We got a lot of stuff to do." Pressed on the reasoning, he was more specific: "I actually think that, given everything happening with safety, right now would be an ill-advised moment to go public."[7] Both OpenAI and Anthropic have nonetheless filed confidential IPO paperwork and are reportedly targeting valuations approaching $1 trillion -- so "not yet" and "not never" are two different claims, and only one of them is the one being made out loud this week.
Four days into this, Fitch Ratings put a number on what happens if the spending behind all of it turns out to be wrong. In a report dated September 10, 2026, Fitch modeled a scenario combining an AI-investment retreat with a broader selloff: US stocks falling 35% over six months, a 15% decline in overseas equities, US GDP contracting 0.6% in 2027 with year-over-year growth turning negative in the second quarter, private capital spending dropping more than 6%, and global growth falling below 1% -- a level Fitch itself characterized as outright stagnation.[8]
Fitch was explicit, in its own language, that this is "a severe downside scenario, not its forecast" -- a model built to stress-test what a bust would look like, not a prediction that one is coming.[8] That distinction did not survive first contact with wider circulation. A newsletter forwarded to this site's author this week cited the same Fitch report as proof that "the bubble can no longer be stopped" -- treating a modeled downside as a foregone conclusion. That is not what the report says, and the difference is not a technicality. A ratings agency stress-testing a bad outcome is doing exactly what ratings agencies are supposed to do; a newsletter converting that stress test into a certainty is doing something else, and it is the kind of gap between what a source actually said and what got repeated that is worth naming plainly rather than letting stand.
Why does this matter? Three people who run or heavily influence the three companies building the most capable AI systems on earth spent a week in open, public disagreement over whether their own industry should regulate itself or be regulated -- and the fight is not academic, because the same week produced a $517 billion compute bet from one of those companies and a modeled 35% stock-market downside tied to the same industry from an independent ratings agency. Amodei's proposal, Altman's alignment, and Sacks' rebuttal are all arguments about who gets to decide when AI development has gone too fast: the labs policing themselves, a coalition of labs setting shared standards, or a government that just said, through its president, it has no interest in slowing down at all. None of that argument requires the Fitch scenario to be a prediction to matter -- it only requires the underlying bet, the $517 billion and the 14.8 gigawatts, to be real, which it is. What doesn't hold up is skipping the word "scenario" on the way to a headline.
Representatives from Anthropic, OpenAI, and Google have been meeting regularly since July 2026 to discuss forming a joint AI industry standards body -- for technical testing and auditing frameworks, led on Anthropic's side by Amodei himself -- which means the private coordination effort predates his public "Pace the Frontier" essay by roughly two months.[9] That reframes the sequence: the essay was not the opening move. It was the public argument for something three competitors had already been building toward privately.
Cohere CEO Aidan Gomez, who runs a fourth frontier lab outside the Anthropic/OpenAI/Google/xAI cluster, published his own rebuttal the same weekend as Sacks -- from a different angle. Where Sacks argued the labs don't need government cover to slow down on their own, Gomez argued the real risk is who gets to write the rules at all: "Should a handful of select, market-dominant AI companies from Silicon Valley get to define the rules and safety standards of a generational technology for the entire world?" -- calling the proposed coordination "a cartel by any other name," one that would entrench the largest labs' power at smaller competitors' expense.[10] Two competitors, two separate cartel accusations, for two different reasons -- one about regulatory overreach, one about who is in the room.
The word "cartel" is not just rhetoric here. Competitors publicly agreeing to jointly limit the pace of their own industry's output is close to the textbook definition of an antitrust violation, which is exactly why Amodei's own essay does not propose the labs simply do this on their own -- it calls for pacing coordination "with government mediation or waivers of antitrust restrictions." A voluntary agreement among competitors to slow down carries real legal exposure without that cover; a government-sanctioned standards body does not. Read against the July meetings, the public essay reads less like a pure safety argument and more like the opening bid in getting the legal clearance the private talks needed to keep going.
The scale of what's being negotiated is also worth stating plainly. Anthropic is reportedly pitching IPO investors a total addressable market above $30 trillion -- topping SpaceX's own $28.5 trillion claim in its May 2026 prospectus, of which SpaceX had attributed $26.5 trillion to AI specifically. Tech analyst Fred Hickey's comparison point: annual US GDP is currently about $32.5 trillion, meaning Anthropic's claimed opportunity approaches the size of the entire American economy.[11] A company pitching a market that large has an obvious interest in how that market gets shaped by whoever writes the industry's safety and competition rules -- which is the tension sitting underneath the pacing debate that a simple "safety-minded CEO versus deregulation-minded critic" framing leaves out.
The antitrust-waiver ask documented above -- Amodei's footnote calling for "government mediation or waivers of antitrust restrictions" so frontier labs can coordinate on pacing without collusion exposure -- drew a direct rebuttal from outside the industry this same week. Former FTC Chair Lina Khan posted publicly on September 14, 2026 that no such waiver is necessary: "there's no AI exemption from laws already on the books," and that "law enforcers already have authority to charge companies and their CEOs for creating and releasing dangerous, unvetted, or defective products."[12] The same week, Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez's March 2026 federal moratorium bill, S.4214 -- which would freeze all new AI data-center construction until Congress passes AI safety and labor legislation -- became a useful second data point: two objections, from two different institutional positions (a former antitrust regulator and two sitting federal legislators), arguing against the industry getting more legal room at the exact moment it is asking for it.[13] A fuller treatment of both threads, including an independent antitrust analysis from Truth on the Market reaching a similar conclusion by a different route, is at "Bernie Sanders and AOC Introduced a Bill to Freeze Every New AI Data Center in the Country Until Congress Regulates AI. Five Months Later, Lina Khan Said the Industry Doesn't Need the Antitrust Waiver It's Asking For."