Robots Atlas>ROBOTS ATLAS
Ethics & Law

AI Diffusion Rule

2025HistoricalPublished: 17 November 2026Updated: 17 November 2026Published
Key innovation
For the first time it linked, in a single rule, the export of advanced AI chips with the export of the weights of the most capable models, and introduced a global, three-tier system of per-country compute caps.
Category
Ethics & Law
Abstraction level
Standard
Operation level
DeploymentLaw and governance
Use cases
Analysis of the geopolitics of AI compute access (compute governance)Assessing regulatory impact on the data-center GPU supply chainA case study of export controls on model weights (not just hardware)Context for later bilateral AI-chip access agreements

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 The rule never took effect — it was rescinded before its compliance date.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

Three-tier country systemThe core of the rule — it determines how much compute may flow to a given country.

Classification of states into Tier 1 (uncapped), Tier 2 (compute caps + trusted-recipient path) and Tier 3 (effective prohibition).

Compute capsOperationalizes the limit on compute accumulation outside a narrow group of allies.

Quantitative purchase ceilings for Tier 2 countries, measured in compute-performance units (TPP / total accelerator count).

Validated End User (VEU) statusProvides flexibility for trusted firms and data centers in Tier 2 countries.

The Universal/National VEU mechanism allowing vetted entities larger volumes after meeting security requirements.

Model-weight controlThe first systematic extension of export control from chips to the models themselves.

Bringing the export of the weights of the most capable closed AI models above capability thresholds under control.

Evolution

Original paper · 2025 · U.S. Bureau of Industry and Security (BIS), Federal Register
Framework for Artificial Intelligence Diffusion (interim final rule)
2025
Rule published
Inflection point

On 13 January 2025 BIS publishes the Framework for Artificial Intelligence Diffusion as an interim final rule.

2025
Rescinded before taking effect
Inflection point

In May 2025 BIS withdraws the rule before its 15 May 2025 compliance date, signaling replacement by bilateral agreements.