How US AI Chip Export Controls Work (as of October 2026)

1. What's controlled: ECCN 3A090, 4A090 and two numbers

US export controls on AI chips sit in the Export Administration Regulations (EAR), run by the Commerce Department's Bureau of Industry and Security (BIS). Each controlled item has an Export Control Classification Number (ECCN). AI accelerators fall under ECCN 3A090, which names GPUs, TPUs, neural processors and ASICs among the integrated circuits it covers, and the servers and boards built around them fall under 4A090.¹

The October 2023 rule replaced the original 2022 parameters with two numbers.¹ Total processing performance (TPP) is 2 × MacTOPS × the bit length of the operation, summed over all processing units on the chip, at the highest rate the manufacturer claims. The rule's own example: 300 teraFLOPS at 16 bits reaches TPP 4,800, as do 600 tera integer operations at 8 bits. A chip that supports both sparse and dense matrices is rated on its dense figure. Performance density is TPP divided by the "applicable die area" in mm², counting logic dies made on non-planar (FinFET-class) transistor processes.¹

A chip with TPP of 4,800 or more, or TPP of 1,600 or more with density of 5.92 or more, is 3A090.a; a tier below that is 3A090.b (the table above gives the bands). Chips that aren't designed or marketed for data centers and stay below TPP 4,800 fall outside 3A090.¹

2. Who needs a license: country groups and headquarters

Whether a license is needed depends on the destination's Country Group in supplement no. 1 to part 740. The 2023 rule requires a license for 3A090 and 4A090 items going to Country Groups D:1, D:4 and D:5, which BIS describes as countries of national security concern, missile technology concern and US arms embargoes respectively, unless the destination is also in Country Group A:5 or A:6.¹

The license review policy then sets the odds. Applications for Macau and D:5 destinations are reviewed under a presumption of denial, as are those for any entity headquartered in, or whose ultimate parent is headquartered in, Macau or a D:5 country, wherever the entity sits. Other destinations get a presumption of approval.¹ BIS's May 31, 2026 guidance confirmed that this headquarters test, first introduced on November 17, 2023, is still enforced.⁸ The 2023 rule also created License Exception Notified Advanced Computing (NAC) for 3A090.b chips and for non-data-center 3A090.a chips; shipments under it to Macau or D:5 require advance notice to BIS.¹

3. HBM and the foreign direct product rules (December 2024)

The December 2024 rule added high-bandwidth memory (HBM) as ECCN 3A090.c: memory with a "memory bandwidth density" above 2 GB per second per mm², measured as the package's bandwidth divided by its area. BIS said every HBM stack then in production exceeded that threshold, and that it used density rather than raw bandwidth so the control would still apply to a larger number of smaller stacks.² Chips that co-package HBM with logic whose main job is processing aren't 3A090.c, though they may fall under 3A090.a or .b depending on their TPP and density.²

HBM needs a license to Macau and D:5 destinations. A new License Exception HBM covers memory below 3.3 GB/s per mm² moving between packaging sites owned by US or allied-headquartered companies that track every unit.² The same rule added two foreign direct product (FDP) rules, which make certain foreign-made chipmaking equipment subject to the EAR when it is bound for advanced fabs in D:5 or Macau, or for Entity List parties marked with Footnote 5, and is the product of US technology or software or contains a component that is. Foreign-made 3A090.c memory also became subject to the EAR where it meets the scope of the existing advanced computing FDP rule.²

4. The AI diffusion rule, and what "rescinded" means here

The Framework for Artificial Intelligence Diffusion, published January 15, 2025, would have made 3A090.a and 4A090.a licensable worldwide. It added License Exception AIA for 18 listed destinations plus the United States, a low-volume exception (LPP) of up to 26,900,000 TPP per consignee per year, a cumulative country allocation of 790,000,000 TPP for 2025 through 2027, and ECCN 4E091 for the weights of closed models trained with 10²⁶ or more operations. Compliance was due May 15, 2025.³

On May 13, 2025, BIS announced the rule's rescission, said its under secretary had instructed enforcement officials not to enforce it, and said BIS "plans to publish a regulation formalizing the rescission and will issue a replacement rule in the future." The same release announced guidance on the risks of using specific Huawei Ascend chips.⁴ As of October 13, 2026, no rescission rule appears in the Federal Register.¹¹ The July 2026 UAE rule describes the result: the worldwide license requirement still sits in § 742.6(a)(6)(iii)(A), but BIS enforces it only for D:1, D:4 and D:5 destinations (except those also in A:5 or A:6) and for entities headquartered in D:5 or Macau.⁹

5. The affiliates ("50%") rule and its one-year stay

Until September 2025, Entity List restrictions applied only to the entities named on the list. The affiliates rule, effective September 29, 2025, extended them to any entity at least 50% owned, directly or indirectly, individually or in the aggregate, by one or more listed entities, and did the same for listed military end users and certain sanctioned parties. BIS said the 50% standard follows long-standing Treasury Department practice.⁵ It applies to every item subject to the EAR, not only AI chips.

On November 12, 2025, BIS stayed it. The stay took effect November 10, 2025 and runs to November 9, 2026, "absent a future extension"; the same rule writes the affiliates provisions back into the EAR effective November 10, 2026.⁶ No extension had been published in the Federal Register as of October 13, 2026.¹¹

6. H200-class chips to China: case-by-case review (January 2026)

The January 15, 2026 rule changed the review policy, not the license requirement. Exports from the United States to end users in China or Macau of chips with TPP below 21,000 and "total DRAM bandwidth" below 6,500 GB/s, which the rule says covers the Nvidia H200 and the AMD MI325X, move from a presumption of denial to case-by-case review, provided the applicant certifies nine conditions.⁷

Among them: enough US supply that no US order is delayed, and no diversion of foundry capacity from US customers; aggregate TPP exported to China and Macau of no more than 50% of the aggregate TPP of the same product shipped to US customers for US end use; know-your-customer and physical security measures; a list of remote cloud users in eight named destinations, China and Macau among them; and a review of every shipment, before export, by a US-headquartered third-party testing lab with no stake in the deal, testing in the United States (a representative sample of each batch is allowed). Reexports, and exports to entities headquartered in D:5 or Macau but located elsewhere, stay under a presumption of denial.⁷

Neither this rule nor Proclamation 11002 of January 14, 2026, which set a 25% tariff on imports of certain advanced computing chips not used for the US technology buildout, describes a revenue-sharing arrangement on exports to China, so this explainer doesn't cover one.⁷¹⁰

7. The UAE moves to Country Group A:5 (July 2026)

The rule effective July 10, 2026 removed the United Arab Emirates from Country Groups D:3 and D:4 and added it to A:5, which opens License Exception Strategic Trade Authorization (STA) to the UAE government and approved entities. Licenses are still required for 3A090 and 4A090 items going to the UAE, except where the ultimate consignee and all end users are UAE government agencies or entities listed in a new supplement no. 8 to part 740.⁹

That supplement lists G42 (with G42 Cloud Technology) and Core42, whose approval expires April 6, 2027 absent further notice, and eight US-headquartered companies and their subsidiaries: Amazon, Apple, Google, Meta, Microsoft, OpenAI Group PBC, Oracle and X.AI. BIS also said it intends to review favorably license applications involving MGX for chips and servers exported to the UAE.⁹

8. What to watch next

On the current texts, the next scheduled change is November 10, 2026, when the affiliates rule's provisions return unless BIS publishes an extension first.⁶ The replacement for the diffusion rule that BIS announced in May 2025 had not appeared in the Federal Register as of October 13, 2026.⁴¹¹ This explainer describes the rules as published; it isn't legal advice, and classification or licensing questions go to BIS through the procedures in part 748 of the EAR.⁸