AI Diffusion Rule
How it works
1) Scope: the control covered advanced AI accelerators and the weights of closed models above training-capability thresholds. 2) Three-tier country split: Tier 1 uncapped, Tier 2 with total-compute caps and a trusted-recipient (VEU) path, Tier 3 effectively prohibited. 3) Compute caps: for Tier 2 it set maximum purchase ceilings measured in compute-performance units, with higher ceilings for trusted-recipient entities. 4) VEU mechanism: firms could obtain Universal/National VEU status authorizing larger volumes after meeting security requirements. 5) Enforcement via the EAR: violations fell under the reexport and Foreign Direct Product regimes. The rule never took effect — it was rescinded before its compliance date.
Problem solved
It sought to solve the problem of fragmented, ad hoc AI-chip export controls: instead of reacting country by country, it set a single global framework limiting the accumulation of compute by actors that could use it to train strategically significant models, and the risk of leakage of the most capable model weights.
Components
Classification of states into Tier 1 (uncapped), Tier 2 (compute caps + trusted-recipient path) and Tier 3 (effective prohibition).
Quantitative purchase ceilings for Tier 2 countries, measured in compute-performance units (TPP / total accelerator count).
The Universal/National VEU mechanism allowing vetted entities larger volumes after meeting security requirements.
Bringing the export of the weights of the most capable closed AI models above capability thresholds under control.
Evolution
On 13 January 2025 BIS publishes the Framework for Artificial Intelligence Diffusion as an interim final rule.
In May 2025 BIS withdraws the rule before its 15 May 2025 compliance date, signaling replacement by bilateral agreements.