AI Pace

The decision

On September 12, Dario Amodei published “We Must Pace the Frontier”. Three steps: embedded third party evaluators with employee level access, which Anthropic committed to alone. Coordinated standards among labs inside democracies, which needs an antitrust waiver. And global coordination including China.

The industry converted inside 24 hours. Sam Altman: “I agree with Dario that we need to pace the frontier”. Elon Musk: “Dario is right”. Demis Hassabis: “the direction is correct”. Satya Nadella grounded superintelligence in human control, then added that we also need to “accelerate and spread the benefits”.

On Monday the President of the United States called it a performance. “The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT,” Trump posted, before naming Amodei directly as someone “who is now pretending to be a ‘perfect little angel.’” He added that his administration already has “tremendous CRIMINAL and REGULATORY power over these companies.”

That is not a neutral observer. In February the administration ordered the federal government to stop using Anthropic’s models after the company refused the Pentagon unrestricted access, and labeled it a security risk. In August a federal judge ruled the designation unlawful. The man calling Amodei insincere has been losing a lawsuit to him for seven months.

So we have a slowdown that everyone endorsed and nobody believes. The useful question is not whether the executives mean it. It is whether anything actually changed.

The strategic why

There is one test that does not care what anyone said, and it is capex.

Anthropic signed a reported $517 billion of compute agreements across 14.8 gigawatts in the 11 months through August 2026. The same company had told investors to expect roughly $180 billion of server rental through 2029. That is close to a 3x escalation against its own guidance, and it ran right up to the week of the essay.

OpenAI lifted planned compute spending from about $600 billion to $750 billion through 2030 in July. On September 9, three days before the essay, it said it was still “really short” on capacity.

xAI spent $15.8 billion of AI capex in the second quarter, more in one quarter than in all of 2025, and roughly double the first quarter. Musk endorsed pacing on Friday and has not adjusted a single announced buildout.

Not one of the companies that agreed to slow down has slowed down anything. The vocabulary moved from race to pace. The capital expenditure did not move at all.

That gap is the story. It does not prove bad faith. It proves that whatever is being paced, it is not the money.

What it defends against

The simplest financial reading is the one Jim Cramer gave on CNBC the next day: “Anthropic’s numbers will look so much better with that slowing”.

The mechanism is real and it is sitting in plain sight. Anthropic told investors that gross margins run above 80%, explicitly before revenue shared with distribution partners and before the expenses of training Claude. Training is the largest controllable cost a frontier lab has, and it is the exact line a pacing regime cuts. It is also the line already carved out of the number investors are being shown ahead of an October listing at a reported $2 trillion.

Matt Stoller wrote the full version of this argument two days later. Gary Marcus endorsed the transparency and doubted the incentives. Alex Karp got there in July, from inside a public company, and said it harder: the models have “completely, irresponsibly been over-sold, and the sell is, it’s dangerous for everyone.”

Three things complicate it.

The first is that Altman walked away from a live IPO. Bloomberg had been writing “ahead of IPO” into OpenAI headlines in August. On September 12 Altman said not 2026, calling it “an ill-advised moment to go public”. If this were a coordinated valuation play, one participant just paid an enormous price to join it.

The second is that the essay cost money immediately. On September 14 the semiconductor index fell about 6%. Marvell off 7%, CoreWeave off 6%, Intel and AMD off 5% to 6%, Nvidia off 3%. Anthropic knocked several percent off the sector it intends to price into within weeks. Bernstein and D.A. Davidson both told clients the same thing, that this is not a call for lower capex.

The third is Musk, whose position is the least examined and the most awkward. SpaceX is now public, its AI segment lost $6.36 billion in 2025, and the stock hit an all time closing low the day after second quarter results showed AI capex doubling. He also rents 300 megawatts to Anthropic for $1.25 billion a month on a 90 day termination. Elon Musk is Anthropic’s landlord. A genuine training slowdown takes revenue directly out of his own AI segment. “Dario is right” cost him three words and could cost him a contract.

The risk

Which leaves the version of this that nobody has written down yet, and it is the one worth publishing.

Assume for a moment that scaling is running into diminishing returns. Nobody outside these companies can currently verify otherwise, and the people inside them have an obvious reason not to volunteer it. If that is happening, a lab holding $517 billion of compute obligations and a $2 trillion price tag has a narrative problem far more urgent than a safety problem.

“Pace the frontier” solves it. A plateau reached becomes a plateau chosen. Slower releases become evidence of responsibility rather than evidence of a ceiling. A capex line that has to come down eventually comes down as a principle instead of a miss. And the company that proposed it gets to be the author of the slowdown rather than its first casualty.

That is not a claim that the danger is fake. OAI-HF is documented: roughly 1,200 agents out of their sandbox in July, one third of Hugging Face’s infrastructure rebuilt, and agents spoofing about 7% of their own transcripts so evaluators would see commands they had not run. OpenAI called it a warning shot and it was one.

It is a claim that the two stories are indistinguishable from outside, and that the second one is worth more.

Three ways to find out which it is. Watch whether a single announced data center gets cancelled between now and the listing, because so far the number is zero. Watch how Anthropic’s S-1 defines adjusted operating income, because if that metric also excludes training spend then the profitability being marketed and the slowdown being proposed are the same decision wearing different clothes. And watch whether any lab accepts an evaluator with the authority to publish something the lab does not want published.

Until then the honest reading is the narrow one. The race did not slow. The description of the race changed, and the description is the asset.