U.S. Export Controls and AI: A Practitioner's Guide

This content is provided for educational and informational purposes only and is not intended and should not be construed as legal advice.

For any company building or deploying advanced AI, the U.S. export controls administered by the Bureau of Industry and Security (BIS) under the Export Administration Regulations (EAR) are the main regulatory constraint on advanced hardware, large-scale compute, and the most capable semiconductors. They are distinct from OFAC sanctions. Export controls are item-based and destination-based licensing requirements. Sanctions are person-based and entity-based prohibitions. A company can satisfy one regime and still violate the other. This guide covers the export-control side: what triggers a license, how the rules changed in 2025 and 2026, and what that means in practice.

What the controls cover

The most important control applies to the advanced computing chip, classified under ECCN 3A090 and the related entries below.

ClassificationWhat it coversControl threshold
3A090.aAdvanced computing integrated circuitsTPP of 4800 or more, or TPP of 1600 or more with performance density of 5.92 or more
3A090.bSomewhat less capable datacenter chipsLower TPP and performance-density tier
3A090.cCertain high-bandwidth memoryPer the 3A090 technical notes
4A090Computers and assemblies containing 3A090 chipsContain ICs meeting or exceeding 3A090
".z" paragraphs (Categories 3, 4, 5)Other listed items incorporating controlled chipsMeet or exceed 3A090 parameters

The performance-density figure was added to prevent a buyer from spreading compute across many smaller chips to stay under the raw-performance threshold. Beyond the chips, the controls also reach the semiconductor manufacturing equipment used to make advanced-node devices, the software and technology behind controlled chips including source code and know-how, and the end-use and end-user "catch-all" controls in Part 744. The catch-all controls can require a license for an otherwise-uncontrolled item when the exporter knows it will be used for a military-intelligence or weapons-of-mass-destruction end use in a Country Group D:5 destination. The Foreign Direct Product Rule extends U.S. jurisdiction to certain foreign-made items that are the direct product of U.S.-origin technology or software, or are produced by a plant that uses U.S.-origin equipment, when they are destined for restricted end uses or parties. For advanced chips, that reach is close to global.

The controls do not reach everything, and over-compliance carries its own cost. Publicly released open-weight models, hardware below the performance thresholds, services that transfer no controlled technology, and exports to close allies generally fall outside the licensing requirement.

How the rules changed in 2025 and 2026

Two sets of controls form the baseline and remain in force, and the activity since then sits on top of them. The sequence is summarized below, with the legal detail in the paragraphs that follow.

DateActionCurrent status or effect
Oct. 13, 2022 (87 FR 62186)Initial controls on advanced computing chips and semiconductor manufacturing equipmentBaseline, in force
Oct. 25, 2023 (88 FR 73458; 88 FR 73424)Expanded chip controls (revised 3A090) and a companion equipment ruleIn force
Jan. 15, 2025 (90 FR 4544)AI Diffusion Rule: worldwide IC license requirement, country tiers, model-weights control (4E091)Effective Jan. 13, 2025; compliance was set for May 15, 2025
May 13, 2025BIS announces non-enforcement, initiates rescission, issues guidanceNot enforced, but not formally rescinded
Jan. 15, 2026 (91 FR 1684)China and Macau move to case-by-case review for H200 and MI325X-class chipsIn effect
May 12, 2026 (GAO B-337935)GAO holds the non-enforcement is itself a Congressional Review Act ruleDiffusion Rule remains in the Code of Federal Regulations

In January 2025, BIS issued the Framework for Artificial Intelligence Diffusion (90 FR 4544, Jan. 15, 2025), effective January 13, 2025, with certain provisions carrying a delayed compliance date of January 15, 2026. The rule is broader than the model-weights control it is sometimes reduced to. It added a control on the model weights of the most advanced AI models under ECCN 4E091, and it also imposed new worldwide license requirements on advanced computing chips and the assemblies containing them, created a tiered country framework, expanded license exceptions, added red-flag guidance, and updated the Data Center Validated End-User authorization.

The rule never reached enforcement, but it has not disappeared, and its current status is often misstated. Two days before the May 15, 2025 compliance date, BIS announced it had initiated rescission and instructed enforcement officials not to enforce the rule. BIS initiated rescission. It did not complete it. Commerce confirmed to the Government Accountability Office in January 2026 that the rulemaking is not finished and that the Framework remains in the Code of Federal Regulations. On May 12, 2026, the GAO held, in decision B-337935, that Commerce's non-enforcement announcement is itself a rule under the Congressional Review Act that Commerce did not submit to Congress. The result is that the AI Diffusion Rule remains on the books and legally in effect, suspended only by an unenforced and now legally contested non-enforcement posture. The prudent course is to continue assessing advanced-computing transactions against the rule, not to treat it as repealed.

In place of enforcing the rule, BIS issued a policy statement and industry guidance. The guidance treats the supply of advanced computing chips or related support, where the supplier knows the items will be used to train AI models for parties headquartered in D:5 countries, as triggering the Part 744 catch-all license requirement. Separate guidance treats dealings in 3A090-specification chips developed or produced in China, with the Huawei Ascend 910B, 910C, and 910D named as examples, as a presumptive General Prohibition 10 violation.

The most significant recent change took effect on January 15, 2026. BIS revised the license review policy for exports from the United States of certain advanced computing commodities to China and Macau (91 FR 1684, Jan. 15, 2026), moving from a presumption of denial to case-by-case review. The eligible chips are commercially available U.S.-market products, the Nvidia H200 and AMD MI325X and lesser equivalents, and only where the exporter certifies sufficient U.S. supply, no diversion of foundry capacity away from U.S. customers, adequate recipient security, and independent U.S. third-party testing. The change applies to exports from the United States only. Reexports, exports from abroad, in-country transfers to D:5 destinations or Macau, and dealings with D:5-headquartered entities remain under a presumption of denial. The rule implements the administration's December 2025 decision to permit these exports to approved Chinese customers in exchange for a share of the resulting revenue paid to the U.S. government.

Model weights

The unfinished status of the Diffusion Rule matters here. Because the rule was never formally rescinded, the model-weights control under ECCN 4E091 remains in the Code of Federal Regulations, even though BIS is not enforcing it and the GAO has questioned the legality of that non-enforcement. The accurate statement is not that model weights are uncontrolled. It is that a model-weights control exists on the books, is currently unenforced, and is under active legal and political dispute. The compute used to train and run models remains controlled regardless, and the catch-all framework still applies. Public release of open-weight models has generally been treated as outside the licensing requirement, because public release is not the same act as a controlled export. The distinction to watch is between publishing weights and providing controlled compute, hosted access, or technical assistance to a restricted party.

Deemed exports, cloud, and remote access

Sharing controlled source code or technology with a foreign national, including inside the United States, can be a deemed export that requires a license tied to that person's most recent country of citizenship or permanent residency. Cloud and remote-access arrangements that allow a person abroad to reach controlled technology raise the same issue. The standard controls are technology control plans, access segregation, nationality screening, and access logging. For a company that hires foreign-national engineers and runs workloads across borders, this is often a larger day-to-day exposure than the headline question of chip licensing.

What a company should do

StepWhat to do
ClassifyIdentify the ECCN of the hardware, software, and technology. For advanced compute, assume 3A090 or 4A090 exposure until classification shows otherwise.
ScreenCheck customers, investors, employees, and infrastructure providers against the Entity List, the Denied Persons List, and current BIS red-flag guidance.
Control accessUse geofencing, per-account compute limits, logging, and know-your-customer checks for high-capability models and APIs.
Foreign-national accessApply nationality screening and technology control plans before any foreign-national access to controlled source code.
DiligenceRun beneficial-ownership diligence on investors, and export, CFIUS, and outbound-investment screening on transactions.
Records and trainingKeep records for five years and train the teams that handle controlled items.
DisclosureIf a violation surfaces, a prompt voluntary self-disclosure to BIS can significantly reduce the penalty. Make that decision quickly and with counsel.

Bottom line

In 2025 and 2026 the visible direction was toward relaxation: an unenforced Diffusion Rule and a conditioned, case-by-case opening for H200-class chips to China. The legal position is more restrictive than the headlines suggest. The Diffusion Rule remains in the Code of Federal Regulations, the catch-all and diversion-knowledge controls have tightened through guidance, and the China opening is narrow and heavily conditioned. The practical exposure for a frontier company is the combination of knowledge-based catch-all controls, diversion risk, and the speed at which these rules change. Any written summary, including this one, is a snapshot, and the operative rule should be confirmed before acting.

Authorities

  • Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification, 87 FR 62186 (Oct. 13, 2022).
  • Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections, 88 FR 73458 (Oct. 25, 2023), and Export Controls on Semiconductor Manufacturing Items, 88 FR 73424 (Oct. 25, 2023).
  • ECCN 3A090, 15 CFR Part 774, Supplement No. 1 (3A090.a: TPP 4800, or TPP 1600 with performance density 5.92; 3A090.b datacenter tier; 3A090.c high-bandwidth memory).
  • Framework for Artificial Intelligence Diffusion, 90 FR 4544 (Jan. 15, 2025), effective Jan. 13, 2025.
  • BIS, May 2025 non-enforcement announcement and accompanying AI Model Training Policy Statement, General Prohibition 10 Guidance, and Diversion Guidance.
  • U.S. Government Accountability Office, Matter of U.S. Department of Commerce, Bureau of Industry and Security, B-337935 (May 12, 2026).
  • Revision to License Review Policy for Advanced Computing Commodities, 91 FR 1684 (Jan. 15, 2026).
Previous
Previous

U.S. Sanctions and AI: A Practitioner's Guide.