The Framework for Artificial Intelligence Diffusion was an Interim Final Rule issued by the Bureau of Industry and Security (BIS) (BIS) of the U.S. Department of Commerce in the final week of the Biden administration. It established a three-tier country framework, per-country and per-entity compute caps measured in Total Processing Performance (TPP), and a new export-control entry, ECCN 4E091, covering the weights of certain closed-weight AI models. The rule was signed January 13, 2025 and rescinded by Trump-era BIS on May 13, 2025, two days before its May 15, 2025 compliance date.
Superseded: Rescinded May 13, 2025 by Trump-era BIS (Under Secretary Jeffrey Kessler), two days before its May 15, 2025 compliance date. Retained for historical reference as the most detailed public articulation to date of a compute-threshold and country-tier AI export-control framework.
Provenance
- Agency: Bureau of Industry and Security (BIS) (BIS), U.S. Department of Commerce.
- Action: Interim Final Rule with request for public comments.
- Federal Register: 2025-00636, January 15, 2025.
- Signed: January 13, 2025 (final week of Biden administration).
- Effective: January 13, 2025; rescinded before its compliance date of May 15, 2025, on May 13, 2025.
Three-tier country framework
The rule sorted destinations into three tiers. Tier 1, trusted allies with no compute cap, comprised the United States plus 18 allies: Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the UK. Tier 2 covered most of the rest of the world under per-country and per-entity TPP allocations, with Validated End User (VEU) paths available. Tier 3 was effectively embargoed and covered China (including Hong Kong), Macau, and Country Group D:5 (Afghanistan, Belarus, Burma, Cuba, Iran, North Korea, Russia, Syria, Venezuela, and others).
Compute metric: Total Processing Performance (TPP)
The framework used Total Processing Performance as its unit of measurement. License Exception LPP permitted up to 26.9M TPP annually per Tier 2 entity without individual licensing. Per-country allocations for Tier 2 totaled 790B TPP across 2025–2027. National VEU (NVEU) allocations were calibrated so that NVEU clusters would trail frontier clusters by "approximately 12 months, or one generation."
License exceptions
The rule created three new license exceptions:
- UVEU (Universal Validated End User) — available only to Tier 1-headquartered entities. It limited AI compute outside Tier 1 to a maximum of 25%, capped any single Tier 2 country at 7%, and required US-headquartered providers to keep at least 50% of compute domestic.
- NVEU (National VEU) — per-country authorizations for Tier 2 data-center operators meeting extensive security compliance requirements.
- LPP (Low Processing Performance) — a streamlined exception for low-volume transactions.
ECCN 4E091: control on closed-weight AI model weights
The rule added a new control, ECCN 4E091, covering closed-weight dual-use AI models trained with at least 10²⁶ operations. Open-weight models were not controlled regardless of compute, and closed-weight models trained below 10²⁶ operations were also not controlled. Tier 1 entities could export controlled weights to Tier 2 under security conditions; export to Tier 3 was prohibited.
Predecessors and residual controls
The framework built on the advanced computing export rules of October 7, 2022 and October 17, 2023. A companion rule, "Implementation of Additional Due Diligence Measures for Advanced Computing ICs" (Federal Register 2025-00711, January 16, 2025), accompanied it.
After rescission, the pre-existing 2022 and 2023 China-focused controls remained in force. BIS issued three replacement guidance documents: an AI Model Training Policy, GP10 (a warning on Huawei Ascend and other PRC-origin advanced chips), and Diversion Guidance. See America's AI Action Plan for the Trump-era AI export policy.
Key claims and contested points
The rule advanced the position that compute, measured as TPP, is a workable export-control unit for AI. This concept was novel and untested in enforcement practice before rescission (confidence: medium).
It also held that closed-weight model weights can be controlled while leaving open-weight models uncontrolled. This was contested: critics argued the approach effectively subsidized open-weight diffusion to adversaries, while supporters argued open weights were already uncontrollable (confidence: contested).
A third premise was that a tiered country framework can balance allied access with adversary denial. Diplomatic backlash from Tier 2 countries including Israel, Poland, and Saudi Arabia was a factor in the rescission (confidence: medium).
As the most detailed public articulation of a compute-threshold export control, the rule remains a reference point in debates over US AI export-control structure, AI diffusion strategy, and whether the US should pursue ally-centered compute alliances.
Relationships
- supersedes: the October 2022 / October 2023 advanced computing rules (augmented, did not replace; those remain in force after BIS rescission)
- superseded-by: America's AI Action Plan — Trump successor AI export-control strategy
- depends-on: Bureau of Industry and Security (BIS) — issuing agency
- related: Executive Order 14110 — Safe, Secure, and Trustworthy AI — conceptual predecessor on compute reporting
- related: Export Controls (AI), Compute Governance, AI Diffusion, US-China AI Competition: Different Races, Different Metrics
- related: Nvidia & TSMC — AI Compute Infrastructure — most-affected companies
- contradicts: America's AI Action Plan — Trump policy reversed key diffusion provisions while retaining China-directed controls