On 12 September, Dario Amodei published “We Must Pace the Frontier”. The essay asks the industry he helped build to do something it has never done: slow down on purpose. Not pause — he is explicit that progress will still feel fast — but deliberately throttle the rate at which frontier models gain capability, so that alignment, interpretability and plain operational discipline can catch up.
Calls for overarching regulation are not new from the labs. What is new is the object of the request, and the sequencing. The object is pace itself, not deployment. And the sequencing is that Anthropic has committed unilaterally to a first step rather than waiting for everyone else. The plan has three stages. First, embedded external evaluators with employee-like access to the labs — desks, badges, access to internal tooling, and the right to publish findings without editorial control. Second, common safety standards among frontier labs in democratic countries, with government backing. Third, global coordination, including with governments that are not democracies. Sam Altman (OpenAI), Elon Musk (xAI) and Demis Hassabis (Google) endorsed the direction within hours; OpenAI matched the evaluator commitment. 1,386 employees of the frontier labs had asked for something similar in the “Pacing the Frontier” open letter in July.
Amodei’s own precedent for the first stage is banking supervision, where examiners sit inside the institutions they oversee. That comparison will land differently in Europe than in the United States, and it is the reason this essay exists.
What Washington is and is not offering
The reaction in Washington has been instructive. The Speaker of the House wants the lab chief executives in a room at the White House, but has ruled out legislation before the midterms. The administration reads artificial intelligence primarily as a race with China; the message is that guardrails are fine, a statutory regime is not, and anyone raising the pace of progress as a risk is bringing the wrong kind of news. The Treasury Secretary put it bluntly a few days before the essay: nothing else matters if China wins. Amodei, notably, agrees with that sentence — his argument is that pacing and export controls are how the democracies keep their lead, not how they lose it.
The gap this leaves is precise. Amodei says in the essay that government regulation of the frontier labs would be the first-best solution, because only law is mandatory. Coordination among the labs is his second-best, chosen because he does not expect the law to arrive in time. But the second stage, as he frames it, still needs the state: agreements between competitors to hold back capability require either a legal basis or a government willing to convene and cover them. In Washington today, neither is on offer. What exists instead is a frontier-model framework that has been discussed with the labs but not published, and a Justice Department that has shown no appetite to facilitate an industry-wide pact.
What Brussels already has
In the European Union, most of what Amodei proposes is not a proposal. It is the law, and it has been in force for a year.
Since 2 August 2025, Article 55 of the AI Act has obliged providers of general-purpose AI models with systemic risk to evaluate those models including through adversarial testing, to assess and mitigate systemic risks — the recitals name loss of control over the model as one of them — to track and report serious incidents to the AI Office, and to maintain adequate cybersecurity for the model and its weights. Article 56 gives providers a Code of Practice as the route to demonstrating compliance; the GPAI Code was published in July 2025 and signed by Anthropic, OpenAI and Google, among others, with xAI signing the safety and security chapter. Since 2 August 2026, Article 92 has given the Commission, acting through the AI Office, the power to conduct its own evaluations of these models, Article 93 the power to demand mitigation measures or restrict a model’s availability, and Article 101 the power to fine.
Henna Virkkunen, the Commission Executive Vice-President responsible for the AI Act, reminded the labs this week that these obligations apply to them. That is the correct observation, and it is also the polite version of a sharper point: the first stage of Amodei’s plan — verification by an actor without a commercial stake — is a thing the EU can already do under its own authority, without waiting for a lab to volunteer.
The AI Office is still being built. Its staffing is thin relative to the mandate, its evaluation capability is young, and its credibility with the labs will be earned model by model rather than granted by the regulation. None of that changes the legal position. Mandate does not depend on headcount.
The role Europe could play
Amodei’s plan is missing an institution. The second stage needs a body with the standing to bring the frontier labs to a table, keep them there, and give what they agree the force of an obligation rather than a press release. In the United States that body does not yet exist, and the political conditions for creating it are not in place. In the EU it exists, has a legal basis in Articles 56 and 92 to 94, and is already in a working relationship with every lab that Amodei would need in the room.
The reason this is plausible is not that Europe builds frontier models. It does not, and the honest version of this argument has to say so. The reason is market access. No frontier lab is going to leave the European market, and every one of them has therefore already accepted that a European authority may test its models, demand changes and, in the limiting case, restrict deployment. The Brussels effect in this domain is not a theory about regulatory diffusion; it is a description of who has signed what. If the labs want a convening authority for common standards, the one that can actually compel attendance is in Brussels, not Washington.
That gives the EU a role it has not claimed and, so far, has not been offered. Amodei’s essay does not mention Europe, the AI Act or the AI Office. That silence is worth reading as an opening rather than as a verdict.
Two limits worth stating
The first limit is reach. The AI Act attaches to models placed on the EU market. The frontier that Amodei is worried about is increasingly not the released model but the internal one — training runs and internal deployments in San Francisco that never touch a European customer and that Article 55 reaches only indirectly, through the provider’s obligations for what it does eventually release. A European convening role would give Brussels leverage over the labs; it would not give it visibility into their internal pace.
The second is sovereignty. Slowing American labs without building European alternatives organises Europe’s own dependency more carefully; it does not reduce it. Pacing the frontier and building a European frontier are not the same project, and a Commission that pursues the first while neglecting the second will have chosen supervision over capability. That may be the right choice. It should be a conscious one.
What this means for those of us who deploy these models
For a European insurer, the practical consequence is modest and immediate. The frontier models we deploy are already subject to Article 55, and from this year to Article 92. The documentation providers must hand to deployers under Article 53, the incident reporting they owe the AI Office, and the evaluations the AI Office may now run itself are becoming the reference points against which our own model governance will be assessed — by our supervisors, and eventually by our auditors. Whether or not Amodei’s plan is adopted, the verification layer it describes is arriving in Europe through the front door of the regulation. Governance frameworks that already treat the provider’s systemic-risk obligations as an input rather than a footnote will find the next few years easier.
Whoever wants to pace the frontier needs someone to hold the stopwatch. In Washington they are still looking for him. In Brussels he has had a mandate since August 2025.