The most consequential line in Dario Amodei’s “We Must Pace the Frontier” is a footnote. The essay, posted on 12 September, lays out a three-step plan. The first step is one Anthropic is “unilaterally committing to.” The third “requires global coordination.” The second, the essay says, “requires industry-wide coordination”, and footnote 1 explains what that means in eight words: “With government mediation or waivers of antitrust restrictions.”
The body text is more specific. Labs should “voluntarily work together to set standards,” and “for antitrust reasons, it’s helpful for the US government to mediate or at least enable these discussions — they don’t need to participate, but do need to issue a narrow waiver for certain kinds of safety conversations.”
That is a request with an addressee. Over the following day the addressees answered, and I think their answers matter more than the agreement from rival lab heads. This piece works through who each step needs, what those people have now said, and what a waiver would take under the law as written.
Step one is embedded evaluators: an outside team with “desks in our offices, access badges, and company laptops,” and a right to publish findings “without editorial control by Anthropic.” No one’s permission is needed for this. A company can invite auditors in on its own. Sam Altman’s reply, relayed by Yahoo News began “I agree with Dario that we need to pace the frontier,” and Elon Musk wrote “Dario is right.” Altman’s post, as embedded by ThePrint, went further: “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.” No terms have followed yet, so for now it is a promise. Kept, it would be a second lab matching a unilateral step, and it needs nothing from Washington.
Step two is where the essay splits into two routes. The preferred one is “regulation that targets all US frontier AI companies,” which needs Congress and a presidential signature. The faster one is voluntary standards among the labs, and that one needs the waiver. Step three needs governments, including China’s, and the essay is frank that its most ambitious version, a full pause, is “unlikely to actually happen any time soon.”
At the Irish Open, held at his golf resort in Doonbeg, Ireland, on 13 September, President Trump said: “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 allowed that “we can put guardrails,” then said “a lot of negative forces” were raising risks that “won’t happen.”
The same day, House Speaker Mike Johnson told Jake Tapper on CNN’s State of the Union, as quoted by TheWrap, “We cannot put a moratorium on this because China will overlap us.” He did not close the door completely. The companies, he said, “need to come together in a meeting with us and sit down and figure out what the right balance is.” Tapper’s own line was the driest of the weekend: “I promise you nobody is going to accuse Congress of racing to do anything about AI.”
David Sacks, formerly the White House AI and crypto czar and now co-chair of the President’s Council of Advisors on Science and Technology, answered the antitrust footnote directly. His post on X, quoted by Tech Policy Press, said: “If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible. But stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel.” In the same post, as republished from Zero Hedge, he wrote that “the easiest way not to build superintelligence is for you to agree not to build it.”
Outside those three, Demis Hassabis also weighed in. He wrote that “the details need working through, but the direction is correct for meeting this critical moment,” and pointed back to his own July proposal for a standards body “modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA).” Amodei’s essay links to that same proposal as one possible venue for step two.
Sacks’s post contains two claims, and they should be separated.
The first is that a company that wants to slow down needs nobody’s permission. That is correct. Section 1 of the Sherman Act reaches “every contract, combination … or conspiracy, in restraint of trade.” A firm that decides on its own to train more slowly has not entered into a contract with anyone. Step one, and any purely unilateral slowdown, sits outside it.
The second claim is the instruction to “agree not to build it.” If “you” means each company separately, that is the first claim again. If it means the companies together, which is how the post addresses them, then the verb is the problem. An agreement among competitors to hold back development is close to the textbook case of what Section 1 was written to stop. Nicholas Felstead made this argument in Lawfare in March, before any of this week’s statements. A coordinated pause among frontier labs “could also readily be construed as an output restriction,” he wrote, and output restrictions are among the “paradigmatic examples of restraints of trade.” The obvious existing shelter, the National Cooperative Research and Production Act, “expressly excludes agreements that restrict output,” as he notes.
A safety motive is unlikely to rescue such an agreement. The Supreme Court rejected that defense in 1978, in National Society of Professional Engineers v. United States. The engineers’ society had banned competitive bidding on the theory that price competition would push engineers toward unsafe designs. The Court called that defense “nothing less than a frontal assault on the basic policy of the Sherman Act,” and added: “Exceptions to the Sherman Act for potentially dangerous goods and services would be tantamount to a repeal of the statute.”
Nor is there current agency guidance to consult. The FTC and the Justice Department withdrew their 2000 Antitrust Guidelines for Collaborations Among Competitors in December 2024. On 23 February 2026 the agencies opened a public inquiry into what should replace them. I could not find replacement guidance issued since.
So the part of Sacks’s post that reads as a gotcha is legally the weak part. A company deciding alone to slow down needs no waiver. Several companies agreeing to slow down together plausibly do. And a second passage of the same post, calling Anthropic and OpenAI “the frontier” by “market share, revenue growth, model capability,” describes the market in the terms that would make a joint slowdown look worst to an antitrust court.
Amodei asks for a “narrow waiver,” but one has been on the books since 1950. Section 708 of the Defense Production Act, codified at 50 U.S.C. § 4558, lets competitors make “voluntary agreements and plans of action” with an antitrust defense. It is also unusually durable. Much of the Act is set to terminate on 30 September 2026, but section 4558 is named among the exceptions.
The statute also says who has to sign. The power belongs to the President, and it begins with a “finding that conditions exist which may pose a direct threat to the national defense.” The President can delegate it only to Senate-confirmed officials, and a designee then has to consult the Attorney General and the FTC “not less than ten days” before talking to industry. Rules go into the Federal Register at least thirty days before they take effect. The agreement itself has to be worked out in meetings chaired by the President’s designee, with the Attorney General and the FTC chairman or their delegates in the room, noticed in the Federal Register at least seven days ahead, open to interested persons unless an exemption applies, and recorded in “a full and verbatim transcript” available for public inspection. Before anything takes effect, the Attorney General must find in writing that its purpose “may not reasonably be achieved through a voluntary agreement or plan of action having less anticompetitive effects …” An agreement expires after five years unless that finding is made again.
Read against the essay, that procedure shows two gaps. The first is that the government’s role is not optional. Amodei wrote that officials “don’t need to participate.” Under section 708 they chair the meetings. The second is the first gate. The statute starts with a finding, by the President or an official he has delegated it to, that conditions threaten the national defense, and on 13 September the President described the threat that concerns him as losing to China.
The provision has carried industry coordination through a crisis before. During the Suez crisis, American oil companies coordinated emergency supply to Europe through a Foreign Petroleum Supply Committee established under section 708. In a June 1967 letter now in the State Department’s Foreign Relations of the United States series, Luke W. Finlay of Standard Oil of New Jersey reminded the Interior Department that the plan had been “drawn up on August 10, 1956” and approved by the Attorney General on 3 December 1956. Writing during the Six-Day War, he asked that the same route be used, because “our lawyers frankly tell us that there is no other means, short of new Congressional legislation, that would provide the necessary antitrust clearance for effective American industry collaboration …” The two routes he named, section 708 or an act of Congress, are essentially the two in front of the AI labs now.
Laid out that way, the ledger after two days looks like this. Step one needed no one and has one company committed and a second saying it will follow. The regulation route through Congress has the Speaker ruling out a moratorium while inviting a meeting. The waiver route through the executive branch starts with a finding that belongs to a President who framed the issue as a race, and a senior adviser calls coordination a cartel. The global route depends on China, and the essay’s own author doubts its strongest form.
None of that makes the proposal empty. Johnson’s invitation to meet is an opening, and a statute could create a safe harbor narrower than section 708, with less procedure. Felstead’s suggestion of amending the cooperative-research law to allow safety-triggered, time-limited output restrictions is one example. Hassabis’s FINRA model would most likely need Congress too. FINRA works because it is registered with, and supervised by, the SEC under federal securities law, and I know of no comparable statute for AI developers. As I argued about the Sanders–Casar bill, each of these routes eventually needs someone to write down the capability that triggers the restraint, and that remains unwritten.
What this week settled is narrower. The parts of pacing that a company can do by itself are legal today and need no one’s approval, which is Sacks’s point. The coordinated slowdown the essay treats as the heart of step two is the part that turns a company decision into a joint one, and the routes the essay names for making it lawful run through the White House and Congress. On the evidence of these two days, neither is in a hurry.
References