Rivals Backed Anthropic's Call to Slow AI, the White House Did Not

Anthropic co-founder and chief executive Dario Amodei published an essay on his own site on Saturday, September 12. It was called "We Must Pace the Frontier."

Its thesis was a single sentence: "We must slow the pace at which we improve the capabilities of AI models."

Within hours OpenAI's Sam Altman, xAI's Elon Musk and executives at Google DeepMind posted that they agreed. The next day Microsoft CEO Satya Nadella backed it too, using the phrase "deliberate pacing."

Over one weekend, the executives whose companies carry the sector's valuation asked for their own products to improve more slowly.

Amodei Listed Three Steps and Took the First Alone

The plan has three rungs: every frontier lab embeds independent evaluators inside the company, companies in democracies converge on a common standard, and governments then reach an international agreement that includes China.

Amodei took the first rung without waiting. Anthropic will give outside evaluators such as METR desks, company laptops and permissions comparable to its internal risk teams. Anthropic will hold no editorial control over what they publish.

The essay gives two clocks. Six to twelve months before a misaligned agent swarm could do serious damage across the internet; one to two years for meaningful progress in interpretability and evaluation techniques.

The White House Refused the Next Day

Trump spoke to reporters on Sunday in Ireland, at his golf resort in Doonbeg, and rejected the call.

"Whoever wins AI wins."

In the same remarks he allowed that guardrails were possible, but framed the question as competition with China rather than safety.

Asia Priced It on Monday Morning

The market did not read this call as a regulatory story.

It read it as an order-schedule story.

The Associated Press reported the Kospi down 3.3% at 6,684.37 and the Nikkei 225 down 0.8% at 63,492.99. SK Hynix fell 6.4% and Samsung Electronics 4.1%, while Taiwan's Taiex lost 0.7%.

In US futures, the Nasdaq was down 1.64%, the S&P 500 0.68% and the Dow 0.23%.

The size of each decline tracked how close the company sits to the AI order book.

Why SoftBank Fell the Most

Altman spoke to Fortune on Saturday and said of a listing that "right now would be an ill-advised moment to go public." Pressed on timing, his answer was short: not 2026.

For SoftBank that sentence is not a safety debate. It is a payment-schedule problem.

The value of a stake bought with borrowed money depends on when that stake turns into cash. If the listing slips a year, the interest keeps accruing through that year while the shares behind it stay private. More: What Is Leverage?

The memory makers fell for a different reason: cost structure. For companies like Micron and SK Hynix, most of the cost base is fab depreciation, which does not move when demand does.

A 10% decline in revenue turns into a decline several times that size in operating profit. Because frontier training runs are the marginal buyer of HBM, or high-bandwidth memory, every step taken on that side comes back through this multiplier.

There Is No Binding Instrument

Here is the part the headline leaves out.

No rule has taken effect. There is no Federal Register notice, no enacted statute, no court order. Four executives stated a preference, and that is all.

Turning that preference into an instrument is itself a legal problem. Competitors jointly restricting output can fall under the Sherman Act.

OpenAI put exactly that question to members of Congress on September 11: would an industry-wide slowdown breach antitrust law? A narrow exemption bill introduced in July has not been enacted.

So the only body that could make the call binding is the government, and the government said no a day later.

The Part That Does Not Fit: Hong Kong

The same morning, the Hang Seng rose 0.4% to 24,904.46.

The pacing call aims at companies inside democracies. Amodei's third step, an international agreement including China, is still only an idea. Korean and Japanese shares are tied to the Western order book; Hong Kong's AI names are not.

If one side slows alone, the other side's relative position improves. That is what was priced on Monday morning.

What Comes Next

Three dates are worth watching. Microsoft opens its model code of conduct to public consultation today and SoftBank's third installment is due October 1. The Fed's rate decision follows on September 16.

None of those is the real test. Will the next large training cluster be ordered, and if so, when? Until that answer arrives, there is no way to tell whether the weekend's words left a lasting mark on price.

Amodei's essay is taken from the primary text on his own site. The Asian closing figures and US futures come from the Associated Press market summary filed Monday morning; Trump's remarks from reports of his comments to reporters in Ireland on Sunday; Altman's from his Fortune interview. SoftBank's installment schedule and bridge-facility figures are drawn from the company's own disclosures. The ratios given for memory makers' cost structure are a general illustration, not any single company's quarterly results.