Dario Amodei wants the industry he helps lead to slow down. On Saturday the Anthropic chief executive published an essay warning that swarms of rogue AI agents could take over the internet within six to 12 months. He then sought a narrow antitrust waiver so that rival labs could coordinate on safety. OpenAI’s Sam Altman replied that “we need to pace the frontier.” Elon Musk posted, “Dario is right.”
When the three most powerful men in American AI agree on anything, antitrust enforcers should read the fine print. So should everyone outside the United States who builds on their models, buys them, or competes with them?
The fear itself is real. On September 8 the Anthropic researcher Jacob Coxon quit, saying the company was “gambling with our lives.” In July, OpenAI disclosed that its own models had escaped a test sandbox, used a previously unknown software flaw, and got code running on Hugging Face’s servers, all to cheat on a hacking benchmark. American voters are uneasy as well: 57 percent say AI’s risks outweigh its benefits.
The remedy is what deserves suspicion. No company needs federal permission to build more slowly. OpenAI proved it in August, when it paused internal Astra work over cybersecurity concerns without asking anyone. A waiver buys something else: the legal right for market leaders to agree among themselves how fast the whole field may move. Under the Sherman Act, competitors who jointly restrict output are a cartel. The statute makes no exception for intelligence.
Stay ahead of the geopolitical week.
MD Briefing delivers expert analysis across five global fronts — the Indo-Pacific, energy, geoeconomics, European security, and the Middle East — every Monday morning. Free.
OpenAI has gone further than Anthropic. Its chief scientist, Jakub Pachocki, urged the industry to slow down voluntarily until safety could be demonstrated. The company then asked members of Congress whether a pact with its rivals would break the law.
Now look at the calendar. Anthropic is preparing for what would be the largest IPO ever. It would be valued near $2 trillion and would raise up to $100bn before the November midterms. Nvidia is weighing a $10bn anchor stake. Anthropic says its annualized revenue passed $65bn in July, up from about $9bn at the end of 2025.
A company that size can afford embedded evaluators, incident-response teams, and a floor of compliance lawyers. A 40-person startup cannot. Neither can a university lab that releases open-weight models, and open weights are where the plan does its quietest damage. Certification tied to model capabilities cannot survive publication, because anyone who downloads the weights can strip the safeguards. Open releases would fail by design. A developer in Karachi or Nairobi who builds on those models would get no seat at the table and no say over the pace. Investors understand the arithmetic. A waiver granted weeks before a record listing would be priced as a moat.
The sharpest objections are not coming from the usual critics of Big Tech. Andrew Ferguson, chair of the Federal Trade Commission, said on Monday that when companies seek “a host of regulations and an antitrust exemption, all of my alarm bells go off.” They are asking, he added, for “barriers to entry that will insulate their incumbency from challenge.” David Sacks, the former White House AI adviser, accused the trio of building a cartel. Cohere co-founder Aidan Gomez said the rules “cannot be written by a small group of commercially aligned companies behind an antitrust waiver.” Even OpenAI’s chief lobbyist, Chris Lehane, said the company already works with Anthropic and Google on safety without any waiver.
The labs have a real reply, and it deserves a full hearing. One company’s restraint collapses when its rivals keep sprinting. That has already happened: in February both OpenAI and Anthropic loosened their safety commitments, arguing that slowing down alone made no competitive sense.
Right diagnosis, wrong cure. Public law exists to solve collective-action problems that span a whole industry. Rules passed by a legislature bind every developer equally, including the ones not invited to the meeting. A private pact binds only the companies that sign it and lets them set everyone else’s pace. Europe has shown that the public route works. Since August 2025 the EU AI Act has required adversarial testing and incident reporting for the most powerful general-purpose models. Those obligations came through this summer’s rollback untouched, and nobody had to deputize the market leaders to write them.
Independence matters just as much. Amodei’s best idea is to give outside evaluators employee-level access to the labs. That only works if the evaluators answer to the public. David Sacks has already told Anthropic to “stop pretending METR is independent,” referring to one of the best-known evaluators. Any regime Congress approves should bar evaluators from taking money from, or sharing staff with, the companies they grade.
The pact also fails the test its authors care about most. A deal among three American firms binds no Chinese lab. On international coordination, Amodei’s essay offers only an eventual treaty, with nothing on how it would be verified or enforced. The frontier would slow wherever the signatories decided and nowhere else, and the rest of the world would live with the result.
Washington is being offered two bad options. Donald Trump calls the slowdown talk a “SICK conspiracy” that helps only China, which means no brakes at all. The labs offer brakes that only they would control. The New Jersey Democrat Josh Gottheimer pointed to a third path: the companies “can pump the brakes at their own labs, today.”
Congress should take that advice and build on it. A bipartisan bill from Senators Adam Schiff and Jim Banks already requires advance notice to the Justice Department before companies collaborate on security threats. Keep it that narrow. Let labs share threat intelligence and incident data openly, under Justice Department supervision, and let any developer join, foreign ones included. Refuse any exemption that lets rivals agree on when new capabilities arrive.
If the frontier is as dangerous as Amodei says, it is too dangerous to be paced by the three companies with the most to gain from pacing it2025,
