The U.S. Department of Commerce’s Bureau of Industry and Security (BIS) has published an interim final rule dubbed the “Framework for Artificial Intelligence Diffusion” — a sweeping update to U.S. export controls that goes beyond China-focused restrictions and aims to shape how advanced AI computing capacity spreads globally.
At a high level, the rule does three big things:
-
Tightens and broadens controls on advanced AI chips and systems (including items classified under ECCNs 3A090.a and 4A090.a, plus related “.z” items).
-
Creates a new control on AI model weights for certain advanced closed-weight models under a new classification (ECCN 4E091) — a major step toward regulating not just hardware, but also the “crown jewels” of frontier AI.
-
Builds a tiered, quota-like approach for most countries, while giving close allies lighter-touch pathways — and offering major cloud providers a route to operate globally under strict compliance obligations.
What changes today — and when it starts to bite
Although the rule is effective January 13, 2025, BIS set a compliance date of May 15, 2025 for most of the new requirements (with certain data-center security guidelines delayed to January 15, 2026). BIS is also soliciting comments through May 15, 2025.
A new “AI weights” line in the sand
One of the most consequential additions is ECCN 4E091, which targets model weights of certain advanced closed-weight dual-use AI models. BIS also introduced a new Foreign Direct Product (FDP) rule for these model weights (new § 734.9(l)), designed to extend U.S. jurisdiction to certain foreign-produced outcomes when the criteria are met — similar in spirit to how chip-related FDP rules have been used in prior rounds of controls.
BIS explicitly says it is not imposing controls on open-weight models today, arguing that (for now) the benefits of open publication outweigh the risks — while still setting guardrails for when a “derived” model becomes materially different through additional training.
The “who gets what” logic: allies, caps, and denials
In practice, the framework splits the world into tiers:
-
A group of close allies is largely exempt from the new caps. Reuters reports about 18 allied destinations are effectively carved out, while most other countries face caps and stricter licensing conditions; embargoed or high-concern destinations remain heavily restricted.
-
The rule text also lists a set of destinations (including the UK, Japan, South Korea, Taiwan, and EU members such as Germany, France, and the Netherlands) where BIS says it is making “minimal changes” and where eligible end users can access advanced items under the new structure with compliance certifications.
For many “non-top-tier” destinations, BIS introduces a per-country allocation concept measured in Total Processing Power (TPP) — including a published default allocation of 790,000,000 TPP under a presumption-of-approval policy up to the cap (after which the review posture can shift).
Data centers and hyperscalers: special paths, strict guardrails
The framework also updates data center pathways through BIS’s validated end-user structures — aiming to make it possible for trusted operators to build and run AI infrastructure globally while limiting diversion risk.
Notably, the rule includes geographic concentration limits for certain globally operating entities: for those with the relevant validated status headquartered in the top group, BIS limits how much controlled compute can be installed outside those destinations — including no more than 25% outside top-tier destinations and no more than 7% in any single non-top-tier destination (with additional limits discussed in public reporting).
New license exceptions (and what they’re for)
BIS adds multiple new license exceptions designed to keep lower-risk trade moving while still constraining “frontier-scale” capability buildout:
-
AIA (Artificial Intelligence Authorization) — tied to exports/reexports of eligible advanced chips and certain model weights (including 4E091) to specified end users/destinations.
-
LPP (Low Processing Performance) — a mechanism for controlled items below certain thresholds, subject to volume limits and reporting.
-
ACM (Advanced Compute Manufacturing) — aimed at supporting certain manufacturing-related flows.
-
plus an expansion of ACA (Advanced Computing Authorized) destination scope.
Industry reaction: “overreach” claims vs. national security framing
The policy landed with immediate pushback from some of the most powerful stakeholders in the AI supply chain. Reuters reports Nvidia criticized the rules as “sweeping overreach,” while Oracle warned the approach could advantage Chinese competitors in the longer run.
BIS and the White House argue the goal is to preserve U.S. leadership and reduce the risk that advanced AI enables military, cyber, WMD, or mass-surveillance abuses.
