Quick one. The expected AI executive order didn’t get signed this week. Trump pulled the ceremony at the last minute, telling reporters the text on the desk risked undermining America’s competitive edge. The order — leaked drafts of which had been circulating for ten days — would have tightened the existing pre-deployment evaluation regime, expanded CAISI’s mandate, and bolted on a set of compute-export controls that went measurably further than the equivalent Biden-era framework.
It didn’t get signed. That’s the news.
What was in the draft, roughly
From the drafts that circulated and what’s been reported on the Hill:
- Expanded CAISI authority to require pre-deployment evals for any model trained above a (lowered) compute threshold.
- Stricter compute-export controls, including new reporting requirements for cloud providers leasing capacity to entities tied to a longer list of adversary jurisdictions.
- A federal procurement preference for models with documented safety evaluations — effectively a soft mandate routed through the GSA.
- Liability language that civil-society groups had been pushing for and the labs had been quietly resisting.
The combination is not a regulatory hammer. It is a tightening of the existing administrative perimeter. Which is exactly why the labs, judging from their lobbying patterns over the last three weeks, did not want it.
Why I think it got pulled
Three readings, in descending order of how much I believe them:
1. The labs won the meeting. The thirty-day comment-and-consult window before the planned signing was, per multiple D.C. reporters, dominated by frontier-lab government-affairs teams arguing that anything that touches export controls or pre-deployment evals slows them relative to Chinese open-weights releases. The “competitive edge” language Trump used yesterday is, to a word, the language those teams have been using since March. The order was watered down twice during the comment window. It was apparently still not watered down enough.
2. The compute-export piece spooked the cloud providers. The reporting-requirement language would have hit AWS, Azure, GCP, and Oracle hardest. All four have spent the last quarter rebuilding their government-affairs benches. It’s plausible — and consistent with the lobbying disclosures filed last week — that the cloud lobby and the lab lobby converged on the same kill list.
3. The administration genuinely doesn’t know what it wants from AI policy yet. This is the read I find least flattering but probably the most accurate. The signals out of the White House on AI have been incoherent for six months — open-weights friendly one week, hawkish on compute the next, deeply pro-CAISI in one channel and dismissive of it in another. Pulling the signing at the last minute is consistent with an administration that hasn’t actually picked.
The three readings are not mutually exclusive.
What the absence of a signature means
For the labs: another quarter of operating under the existing voluntary-CAISI framework, plus whatever they signed bilaterally last quarter. That’s a meaningful concession.
For the cloud providers: the compute-export reporting requirements are deferred indefinitely. The provisional reads of the draft suggested the reporting burden was several hundred million dollars per provider in the first year. That money stays unspent.
For policy: the gap between the AI policy environment in the EU (binding, granular, ratcheting) and the one in the US (voluntary, bilateral, contested) widens by another notch. Companies will arbitrage that gap. They already are.
The interesting thing about the un-signed order is that it makes the next one harder to write, not easier. Once an administration pulls a draft, the next draft starts from a lower ceiling. The labs and the cloud providers have just learned the same lesson they learned in 2024: the comment window is the whole game, and the pen is the formality.
It is going to be a strange eighteen months for US AI policy. Yesterday was the cleanest data point we’ve had on which way the strange goes.