The EO that wasn't — what Trump's postponement tells us about the next two years of AI policy
Trump pulled the AI executive order hours before signing. The accelerationist camp won Thursday. But the structural pressure that produced the EO doesn't go away because the order didn't sign — it just routes through different channels. Here's the map of what those channels look like.
What actually happened
Trump postponed the EO citing 'I didn't like certain aspects of it' and 'I don't want to do anything that's going to get in the way of [the US AI lead].' The pulled order would have established a voluntary review process for advanced AI models, with federal agencies getting up to 90 days for pre-release security review. Major AI firms — OpenAI, Anthropic — had been negotiating with the administration.
The postponement is not an abandonment. White House reporting consistently describes the order as 'delayed,' not 'killed.' What changed Thursday is the assumption that the order would land in May at 90 days. The new assumption: the order may resurface in weeks or months, possibly at a softer 30-day window, possibly with carve-outs that the accelerationist camp finds acceptable.
The split inside the administration
Reporting describes two camps inside the administration:
- The accelerationists. Argue any disclosure framework slows release velocity, cedes ground to Chinese frontier labs, and represents the kind of preemptive regulation the administration explicitly rejected when it 'tore down' the Biden-era AI rules in early 2026. This camp won Thursday.
- The Mythos camp. Mostly national-security alumni and AISI-aligned staffers. Argue unsanctioned frontier deployment produces unmanaged cyber and bio risk — pointing to Anthropic's Mythos capability as proof of concept. This camp lost Thursday but did not concede.
Why the structural pressure doesn't disappear
The labs themselves are partly the driver of the regulatory pressure. Anthropic's principled refusal of Pentagon contract language produced the procurement-exclusion mechanism that's now the binding US-side regulatory reality. The Pentagon's testing of rival AI models is the operational consequence. The EO would have provided a more orderly framework; without it, procurement-exclusion remains the cudgel.
The administration's choice on Thursday was between formal voluntary disclosure (the EO) and informal procurement-driven coercion (the Pentagon mechanism). They chose the latter by default.
The Glasswing pattern as the industry's answer
The under-noticed development is that the industry is building its own governance structures faster than the federal government is. Anthropic's Project Glasswing — $100M in usage credits to a 10-partner consortium including AWS, Apple, Google, Microsoft, NVIDIA, JPMorgan, and the Linux Foundation — is the institutional pattern the Mythos camp inside the administration wanted to formalize. Anthropic is shipping it as a private arrangement instead.
If more high-capability holds follow the Glasswing pattern (capability access gated to a contractually-bound consortium), the EO becomes less relevant. The industry is creating its own version of the framework Trump was about to sign — without federal sanction, without standardized requirements, but with the actual operational effect the EO would have produced.
The forward read for the next 12 months
- The EO resurfaces in softer form. 30-day window instead of 90, possibly with carve-outs that make it commercially survivable for labs that were going to refuse the 90-day version. Probable.
- Procurement-exclusion remains the binding regime. Whatever the EO ends up looking like, the Pentagon's mechanism for disciplining labs that refuse contract language stays. The Anthropic case sets the precedent either way.
- Glasswing-shaped consortia multiply. Each high-capability lab that holds back a public release builds its own version. The patchwork becomes the de-facto governance regime.
The Q3 2026 watch is whether the AISI publishes the methodology updates that would let an eventual EO actually function. The Opus 4.5 evaluation showed the methodology is improving but not yet sufficient for circuit-level test-awareness checks at scale. Without that methodology, even a signed EO produces longer reports of the same shape — not better safety.
CNBC — Trump postpones AI EO → · Axios — why pulled → · Washington Post — White House AI defenses →