How do U.S. export licenses apply to AI chips and model weights?

What is an AI export license?

It is permission from the U.S. government to send a controlled item to a foreign destination or person.

The rules are the Export Administration Regulations (EAR), run by the Commerce Department’s Bureau of Industry and Security (BIS). Each controlled item has a code on the Commerce Control List. For AI, the key codes cover advanced computing chips (3A090), computers and servers that contain them (4A090), and, under the 2025 rule, certain model weights (4E091).

If an item is controlled for a destination, the exporter must apply for a license. BIS reviews the application under a stated policy, such as presumption of approval, case-by-case, or presumption of denial.

Unlike most AI rules, an export license is a true yes-or-no permit. That is why export controls belong in any discussion of AI licensing. See AI licensing explained.

How did chip export controls begin?

On October 7, 2022, BIS issued an interim final rule restricting exports of advanced computing chips and chipmaking equipment to China. BIS described the goal as limiting China’s ability to buy and make high-end chips for military uses. The advanced computing controls took effect October 21, 2022.

BIS expanded those controls with three rules on October 17, 2023, extending licensing requirements to more countries and closing loopholes, according to the agency’s announcement.

These controls are separate from the 2025 Diffusion Rule and remain in force.

What was the AI Diffusion Rule?

The AI Diffusion Rule was a January 2025 Commerce Department rule that sorted countries into three tiers for AI chip exports and required a license to export the weights of the largest closed models.

BIS published the Framework for Artificial Intelligence Diffusion as an interim final rule on January 15, 2025, days before the change of administration. Most provisions had a compliance date of May 15, 2025.

It sorted the world into three tiers, according to an Arnold & Porter summary:

• Tier 1: the United States and 18 allies and partners, including Japan, the United Kingdom, Germany and Taiwan. Most exports could go under a new license exception (AIA); license applications got a presumption of approval.
• Tier 2: most other countries. Approval up to a capped allocation of computing power, measured in “total processing performance,” then a policy of denial.
• Tier 3: arms-embargoed countries, including China, and Macau. Presumption of denial, as before.

The model-weight control was new. The rule created ECCN 4E091, a control on the weights of closed models trained with 10²⁶ or more operations. Licenses for those weights would be reviewed with a presumption of denial, with a license exception for Tier 1 destinations. Open-weight models were excluded.

That put a license requirement on model weights themselves, not just the hardware.

Was the AI Diffusion Rule rescinded?

Commerce stopped enforcing the AI Diffusion Rule in May 2025, but as of October 2026 it had not published a formal rescission, so the rule is unenforced rather than repealed.

On May 13, 2025, two days before the compliance date, BIS announced it was rescinding the rule, which it called burdensome. Under Secretary Jeffrey Kessler told enforcement officials not to enforce it. BIS said it “plans to publish a regulation formalizing the rescission” and would issue a replacement later.

The same announcement issued guidance on the risks of using certain Chinese advanced chips, including specific Huawei Ascend chips, and on preventing diversion of U.S. chips.

As of June 5, 2026, no replacement rule had been issued, according to Holland & Knight. On July 14, 2026, Under Secretary Kessler told Congress the administration no longer intends to replace the rule and is focusing on new rulemaking, according to Morgan Lewis. Earlier, on May 12, 2026, the Government Accountability Office concluded in decision B-337935 that the non-enforcement announcement was itself a “rule” under the Congressional Review Act. Pillsbury read the decision as leaving the original rule legally in effect, with companies making risk-based compliance choices.

The result is that the 4E091 model-weight control is still in the regulations but is not being enforced. A future administration could change that.

What are the current rules for AI chips to China?

The 2022 and 2023 controls still require a license for advanced chips going to China. What changed in 2025 and 2026 is how BIS treats those applications, and how chips designed to fall under the thresholds are handled.

The H20

Nvidia designed the H20 for the Chinese market below earlier thresholds. In April 2025, the U.S. government told Nvidia that exports of the H20 to China would require a license. Nvidia said it would take about $5.5 billion in charges tied to H20 inventory and commitments.

In August 2025, Nvidia and AMD agreed to pay the U.S. government 15% of revenue from sales of the H20 and MI308 chips to China in exchange for export licenses, according to the Financial Times and to Reuters, which cited a U.S. official. Neither company confirmed the terms, and we found no regulation setting out the arrangement.

The H200

On December 8, 2025, President Trump posted that Nvidia would be allowed to ship H200 chips to approved customers in China, with 25% of the proceeds going to the United States. He said Blackwell and later chips were not included.

On January 13, 2026, BIS announced a rule moving license review for the H200, AMD’s MI325X and similar chips to China to case-by-case. Applicants must show three things:

• the exports will not reduce global semiconductor production capacity currently available to U.S. customers;
• the Chinese buyer has export compliance procedures, including customer screening;
• the product passed independent third-party testing in the United States for performance and security.

This is a licensing policy, not a general license. Each sale still needs approval.

Companies headquartered in China

On May 31, 2026, BIS issued guidance confirming that a license requirement dating from November 2023 still applies to exports of advanced computing items to entities headquartered in China (or another Country Group D:5 country) or Macau, or whose ultimate parent is, even when the entity is located elsewhere, according to Holland & Knight. BIS said bona fide data center operators need not stop using equipment already in place until further notice.

Have the licensed chips actually shipped?

H200 chips have shipped to China only in small numbers, according to reporting through August 2026. In July 2026, CNBC reported that Under Secretary Kessler told Congress that very few H200 shipments had taken place. In late August 2026, the South China Morning Post reported that Nvidia confirmed its first H200 shipments to China and still forecast no China data center compute revenue.

How do the export rules compare?

Of the five rules and policies below, only the AI Diffusion Rule goes unenforced; the last row is GAO’s decision about it.

Rule or action	Date	What it does	Status (Oct 2026)
Advanced computing controls	Oct 7, 2022; expanded Oct 17, 2023	License required for advanced AI chips to China and other listed destinations	In force
AI Diffusion Rule	Published Jan 15, 2025	Three country tiers; chip caps; license for closed weights ≥1026 (4E091)	Announced rescinded May 13, 2025; not enforced; no formal rescission found
H20 license requirement	April 2025	License needed for H20 to China	License requirement applies
H200 / MI325X review policy	Announced Jan 13, 2026	Case-by-case review with supply, compliance and testing conditions	In force; few shipments reported
Headquarters guidance	May 31, 2026	Confirms 2023 license requirement for China- or Macau-headquartered entities anywhere	Guidance; requirement in force
GAO decision B-337935	May 12, 2026	Non-enforcement announcement is a rule under the Congressional Review Act	No resolution of disapproval found

Do export controls cover open-weight models?

No. The 2025 rule’s model-weight control, ECCN 4E091, covered only closed weights, and the rule expressly left open-weight models out.

That exclusion matters to the open-model community, because a rule that controlled open weights would function as a license on publication. For how open models are licensed privately, see open-weight model licenses.

Where does the shared 10²⁶ number come from?

The same number appears in a sequence of rules. Executive Order 14110 used 10²⁶ operations for its reporting duty in 2023. The Diffusion Rule reused it for model weights in 2025. California and New York reused it in 2025 and 2026. The reasoning and the criticism are in compute thresholds.

Is export licensing a license on AI?

For chips, yes. For models, only on paper. Export licensing is the only frontier AI permit the United States runs today, and it licenses chips rather than models as long as the model-weight control stays unenforced.

Several items here, including license approvals and shipment volumes, change month to month and are stated as of October 2026. This guide is general information, not legal advice.