AIAcademy · AIAcademy · 2026-05-16
Framework for AI Diffusion (Federal Register, Jan 2025)
A four-month life. The Framework for Artificial Intelligence Diffusion was published in the Federal Register on 15 January 2025, five days before the Biden administration left office. It was the most ambitious export-control regime the United States had ever proposed for AI compute: a tiered system dividing the world into a small group of trusted allies, a large middle tier with per-country caps, and a denial list led by China.
What it tried to do. Tier 1 (roughly the Five Eyes plus a few NATO allies and Japan, Korea, Taiwan) had effectively unlimited access. Tier 2 — most of the world, including India, Israel, Saudi Arabia, the UAE — was capped at a national compute ceiling and required Validated End-User authorisations for large orders. Tier 3 (China, Russia, plus the existing denial list) was a near-blanket prohibition. Model weights themselves above a 10²⁶ FLOP training threshold also fell under the rule.
Why it died. The Trump Commerce Department rescinded the rule in May 2025, citing two arguments: the tiered structure penalised diplomatic partners, and the compute-cap mechanism would have required BIS to administer a global allocation system it lacked the staffing to run. The replacement framing was that AI export controls should be case-by-case and bilateral, not categorical.
What replaced it. Through 2025 the administration approved targeted Nvidia H200 and AMD MI325X export licences to specific mainland Chinese end-users under closely-monitored conditions — the precise opposite of the Diffusion Rule's logic. The CSIS analysis of the post-Diffusion landscape traces how DeepSeek's late-2024 success and Huawei's Ascend ramp-up changed the political calculus: blanket bans were producing domestic substitution faster than restraint.