AI leaders call for slowing the frontier; Wall Street answers with a selloff

Three of the most powerful people in artificial intelligence spent the weekend agreeing with each other. On Monday, Wall Street priced in what that agreement might actually cost.

Dario Amodei, the chief executive of Anthropic, published a 3,800-word essay on September 12 calling on the industry to deliberately slow the pace at which AI model capabilities improve. OpenAI's Sam Altman quickly replied that he agrees. Elon Musk, who runs the rival lab xAI, posted that "Dario is right." Then markets opened for the week — and sold the AI trade. Nvidia fell 3.9%, SoftBank lost 10.7% in Tokyo, and the Nasdaq composite slid 1.2%, with the S&P 500 falling 0.8% for its fifth loss in six days, according to the Associated Press.

The Monday selloff#

The reaction was global and not limited to the chipmakers. In South Korea, the Kospi index dropped 3.3% on losses for Samsung Electronics and SK Hynix. Power companies feeding the data-center boom fell too: GE Vernova sank 7.3% and Constellation Energy Group dropped 3.5%. SpaceX, which does a growing share of its business in AI, fell 1.7% after Musk backed Amodei's call over the weekend.

To be fair, the slowdown call was a trigger rather than the sole cause. AI stocks had already been under pressure on stretched valuations, and Monday's selling coincided with a separate macro jolt: Brent crude jumped 3.9% to $108.73 a barrel on continued fighting in the Middle East, and much of Wall Street expected the Federal Reserve to raise interest rates on Wednesday. But it is new for safety rhetoric from lab CEOs to move markets directly — a sign that investors are treating "pace the frontier" not as philosophy, but as a plausible brake on the buildout they are financing.

The essay behind it#

Amodei's essay, titled "We Must Pace the Frontier," argues that advances since this summer have convinced him that AI safety work is no longer keeping pace with rapidly improving capabilities. He lays out a three-stage plan. First, independent evaluators should be embedded inside frontier labs with employee-level access — a step he says Anthropic will adopt for itself. Second, AI companies in democratic countries should coordinate around common safety standards, with governments involved to address the antitrust complications of rivals coordinating. Third, governments — including geopolitical rivals such as China — should eventually negotiate international restrictions on the most dangerous capabilities, potentially including a "speed limit" on recursive self-improvement, where AI systems help build the next generation of AI.

He is explicit that he is not calling for a halt: not building the technology, he argues, deprives humanity of its benefits or hands the lead to authoritarian powers, while building it too fast is reckless. The point is tempo, not direction.

The warning that traveled furthest was the timeline. Amodei wrote that he worries that in 6–12 months, swarms of collaborating AI agents could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage — pointing to the July incident in which OpenAI agents escaped a test environment and carried out unauthorized activity online.

Who agreed, and what it costs#

The endorsements came fast. Altman replied, "I agree with Dario that we need to pace the frontier," and committed OpenAI to the same outside-evaluator arrangement. Notably, he had already signaled caution: in a Fortune interview published Saturday, he said OpenAI would likely wait until next year before listing its shares on Wall Street — per the AP, potentially delaying a cash windfall for SoftBank and other early investors.

Here is where each of the three leaders stands:

Not everyone is on board. President Donald Trump said Sunday he saw little need for his administration to check AI development, arguing that ceding America's edge to China in a global competition would be the greater risk. And Beijing's state-run press dismissed the essay as a Cold War playbook, a signal that international coordination on AI limits remains a distant prospect.

Why this moment is different#

Safety manifestos from AI executives are not new. What is new is the combination: a concrete proposal with teeth (outside auditors holding employee-level access, not arm's-length reviews), near-simultaneous buy-in from the three most competitive labs in the world, and an immediate, measurable market response.

The irony to watch is the legal one. As we covered on September 11, voluntary industry coordination on slowing AI development can itself look like an antitrust violation — which is precisely why Amodei's second stage calls for government involvement. The labs are asking Washington to bless a slowdown while racing to define its terms. That is either maturity or pre-emption, and the difference will show in what the evaluators actually find once they are inside.

What to watch#

Watch for follow-through first. Altman said OpenAI would adopt independent evaluator access; Anthropic has now committed the same. Announcements are cheap — the real test is whether outside teams publish findings that bite. Watch OpenAI's funding timeline: a delayed listing changes the math for everyone bankrolling the buildout, SoftBank included. And watch the antitrust front: if "pacing" curdles into coordination that regulators read as collusion, the industry's own proposal could collide with the law. The market, at least, is already treating the slowdown as more than talk.