Trump Pulled the Federal AI Review Order at the Last Minute. The Rules Now Come From Sacramento and Brussels.

The Trump administration spent Thursday, May 21, getting ready to sign an executive order that would have given the federal government its first structured look at frontier AI models before they ship. Hours before the ceremony, the President pulled it.

The order was not stopped by a hearing or a court. It was stopped by phone calls. Reporting across the Washington Post, Axios, and NBC has the former White House AI and crypto czar David Sacks, along with Meta chief Mark Zuckerberg and xAI chief Elon Musk, reaching the President between Wednesday night and Thursday morning and talking him out of it.

What got pulled matters less for what it contained than for what its absence sets in motion. A federal pre-release review of frontier models was the one mechanism that could have produced a single national standard. Without it, the rules that actually bind a model before it ships now come from Sacramento, Brussels, and the legal departments of the labs themselves. That is the story.

What the order would have done

The draft, framed around cybersecurity and AI, would have set up a voluntary arrangement. Leading AI companies share their most advanced models with the federal government before public release, and federal agencies get a window of up to 90 days to run a security review. Several agencies would have been directed to define which models fall under the regime and to build the process for evaluating not-yet-released systems alongside the companies that make them.

Read the word doing the work there: voluntary. This was not a licensing regime or a hard gate. It was an opt-in pre-release check, the federal version of the safety-testing agreements that bodies like the US AI Safety Institute had already signed with individual labs. OpenAI and Anthropic were both at the table negotiating it.

So the order mostly codified what the frontier labs were already doing on a handshake. That is what makes the reversal interesting. The thing that became too hot to sign was the formalization, not the substance.

Who killed it, and why

The stated reason and the real reason are both on the record, and they do not match. Trump told reporters he pulled the signing because he did not like certain aspects of the order and did not want anything to become a "blocker" on an AI race the US is currently leading against China. That is the competitiveness argument, and it is the one the administration will keep repeating.

The version from people close to it is blunter. One source told reporters the President "just hates regulation," and that Sacks hated this order in particular.

The three names that moved him are the tell. Sacks is the administration insider on AI. Musk runs a frontier lab that would have been subject to the review. Zuckerberg runs the largest open-weight model program in the country, the one a pre-release federal review would complicate most. The opposition here was not the safety camp arguing the order was too weak. It was the industry worried it was a precedent. A voluntary review you negotiate today is the template for the mandatory one a future administration writes.

The binding rules are coming from everywhere else

Here is what the competitiveness framing obscures. Pulling the federal order does not leave frontier AI unregulated. It leaves it regulated by everyone except Washington.

The California Transparency in Frontier AI Act, SB 53, is already law, with a training-compute threshold near 10 to the 26th FLOP, a 500 million dollar revenue line, pre-deployment transparency reports, and incident reporting. The EU AI Act Code of Practice for general-purpose models is live and carries real penalties. Both bind any lab that wants to sell into those markets, which is all of them.

The labs have read the board. Each frontier company now runs a split my colleague mapped in our companion piece on the new frontier-governance frameworks: a voluntary best-practices policy it can edit at will, and a statute-facing compliance framework it cannot quietly walk back. OpenAI published exactly that this week.

A federal pre-release review would have been the one chance to fold that patchwork into a single national standard the whole industry answers to. Pulling it means the consolidation does not happen, and the patchwork hardens.

What it means for the model-release pipeline

For anyone shipping or building on frontier models, the near-term read is simple and the medium-term read is not.

Near term, there is no new federal gate between a finished model and its release. The release pipeline you can track on our models catalog does not get a 90-day review step bolted onto it. Models ship on the timelines the labs choose, the same as last week.

Medium term, the compliance surface fragments instead of consolidating. A lab now answers to California, to the EU, to whatever a New York bill becomes, and to its own published framework, with no federal layer harmonizing them. For the labs that is a legal cost. For the agent builders downstream it is a sourcing question, because the model you wire into an agent is governed by a different rulebook depending on where your users sit, and for now the provider absorbs that, until the day a state decides the deployer shares the obligation.

The GUARD Act fight in the Senate, which I wrote up when it cleared committee 22 to 0, is the reminder that the federal vacuum is at the model-oversight layer, not across the whole field. Congress is still moving on the consumer-protection side. It is the pre-release review of frontier capability specifically that just lost its federal champion.

Our take

The competitiveness argument has it backwards. A single, voluntary, federal pre-release review was the most industry-friendly shape that oversight of frontier models was ever going to take. It was opt-in, it was negotiated with the labs in the room, and it would have handed every company one rulebook instead of fifty. Pulling it does not make AI less regulated. It makes it regulated less coherently.

The winners are not the builders who wanted freedom. The winners are California and Brussels, which now set the binding rules by default, and the compliance departments inside the labs, which become the de facto standard-setters in the gap. The loser is the idea of a national AI policy, which just got pushed past the next model generation.

Three signposts over the next ninety days. First, whether the administration revives a narrower order or cedes the ground entirely. The Washington Post has reported it is already building alternative defenses, so this reads as a delay, not a funeral. Second, whether a second or third state passes an SB 53 variant before any federal action lands, hardening the patchwork into the de facto national rule. Third, whether the labs, with the federal formalization off the table, quietly loosen the voluntary commitments they made when an order looked imminent, or hold the line because Sacramento and Brussels are watching regardless. The order is gone. The rulemaking did not stop. It just changed addresses.