Three phone calls killed Trump's AI executive order
Trump pulled his AI executive order hours before signing, after calls from Musk, Zuckerberg and Sacks. What it signals.
A text drafted for months, killed overnight
On Wednesday evening, May 20, the White House was preparing to sign the first AI executive order of the Trump 2 era. The text had been circulating in federal agencies for weeks, and OpenAI and Anthropic were actively negotiating the wording. Twenty-four hours later, Trump pulled the order without signing. Three overnight phone calls did the job.
The draft, surfaced by Axios, held two main blocks. A "covered frontier models" section asked AI labs to submit advanced models to the federal government up to 90 days before public release. The entire setup was voluntary, with no penalty and no licensing.
Industry was lobbying for 14 days, not 90. A second section organized a cybersecurity "clearinghouse" run by Treasury to identify vulnerabilities in models before release.
For scale, the Biden executive order from October 2023, which Trump killed in January 2025, was substantially more restrictive. The May 20, 2026 text was already the watered-down version. And it was still too regulatory for the accelerationist camp.
Three phone calls, a signing that vanishes
According to Axios and Semafor, three calls reached Trump overnight on May 20 to 21. Elon Musk for xAI. Mark Zuckerberg for Meta.
David Sacks rounded out the trio. Sacks left his post as "White House AI and Crypto Czar" in late March 2026, but he now co-chairs PCAST, the President's Council of Advisors on Science and Technology. The core argument boiled down to a single word: China.
On May 21, Trump confirmed the withdrawal publicly with a blunt line: "I didn't like certain aspects. I postponed it. I think it gets in the way of, you know, we're leading China, we're leading everybody."
Sacks reportedly argued internally, according to a source cited by Axios, that the voluntary 90-day mechanism could drift into a de facto licensing regime. A curious argument for a text that contained no binding obligations and no penalties.
Journalistic caveat: the parties pushed back on the story. Musk denied it on X, saying he had spoken to Trump only after the decision and that he was unaware of the order's contents. Meta took a similar line for Zuckerberg, saying the exchange came after the withdrawal.
The Manila Times carried the denials on May 23.
The Washington Post, for its part, confirmed the Axios angle on May 22 with several additional White House sources.
The exact timing of the calls remains contested. On the direction of travel, however, sources converge: the accelerationist camp won the internal fight before signing was even possible.
A structural lock
The withdrawal is not a calendar accident, it is the trace of an underlying line. Three days before the planned signing, on May 18, 60+ MAGA figures published an open letter titled "Humans First."
Steve Bannon, Amy Kremer and several Tea Party lawmakers asked Trump to impose mandatory federal testing on advanced models. "Mandatory testing and government approval," wrote Bannon, who explicitly demanded a binding regime.
The May 20 draft already ignored that demand and stuck to voluntary measures. Even that compromise collapsed within hours.
The accelerationist camp holds the field: David Sacks and PCAST, Marc Andreessen and a16z, the founders of the largest labs, the National Economic Council, part of the Vice President's staff. On the other side, Bannon and the populist wing of MAGA see AI as a threat to jobs and national security.
The AI regulation rift now runs through the Trump coalition rather than separating Democrats from Republicans. The withdrawn order is less a failed compromise than a power test the accelerationists won.
The mechanical conclusion: through the two years left in the term, there will be no binding federal framework on AI models in the United States. As Transformer News puts it, the political window for federal AI regulation in Washington is closed until 2029.
Europe alone against the market
For France and Europe, the calculation shifts radically. The AI Act becomes the only binding regulatory safety net at the scale of a major Western bloc. And the concrete date is on the calendar: on August 2, 2026, ten weeks from now, the AI Office gains the power to impose fines of up to 3 percent of global revenue for violations on GPAI (general-purpose AI models).
The sanctions are live. The timing with the US executive order withdrawal is striking.
Mistral and European publishers enter a new configuration: a US market now without federal constraints, a European market with rules that Americans are no longer required to follow. Either Europe holds its line and exports its standards through extraterritoriality, the way it did with GDPR. Or European capitals fold under the industrial lobbying that is about to land. Paris and Berlin have already started arguing for a loosening.
The real question of the coming months shifts from Washington to Brussels. And the European window measures in tens of weeks, not years.
