US export controls on advanced chips are part of the Export Administration Regulations (EAR), administered by the Bureau of Industry and Security (BIS) in the Commerce Department. Items are classified under Export Control Classification Numbers; advanced computing chips fall under ECCN 3A090, and computers and servers containing them under 4A090. When a rule applies, exporting, re-exporting or transferring such items to a covered destination or end user requires a license from BIS. [2][9]

This article summarises public rules and commentary as of September 2026 and is not legal advice. The regulations are detailed and change often; companies should rely on the current EAR text, BIS guidance and qualified export control counsel.

BIS issued its first advanced computing and semiconductor manufacturing rule on 7 October 2022 as an interim final rule, which let it take effect before public comments closed. GAO notes that this approach was chosen to avoid stockpiling of controlled items. The rule restricted exports of advanced chips, supercomputer-related items and chipmaking equipment to China. [1][5]

The original chip thresholds combined a processing performance level of 4,800 with a chip-to-chip interconnect speed of 600 gigabytes per second. The October 2023 update removed the interconnect parameter, kept a performance threshold, and added a new performance density measure; BIS said the updates were meant to shut off pathways used to evade its restrictions. [3][4][1]

Since the October 2023 update, the main metric has been total processing performance (TPP), defined as 2 times a chip's multiply-accumulate throughput in trillions of operations per second, times the bit length of the operation. A second metric, performance density, relates TPP to the chip's die area, so that efficient designs with a smaller footprint can also be captured. [2]

Under the 2023 rule, a chip is in ECCN 3A090.a if its TPP is 4,800 or more, or if its TPP is 1,600 or more with a performance density of 5.92 or more. ECCN 3A090.b captures a lower band, such as a TPP of 2,400 up to 4,800 with a performance density of at least 1.6 but below 5.92. Chips that are not designed or marketed for use in data centers and have a TPP below 4,800 are carved out of parts of these controls. [2][3]

BIS released three rules on 17 October 2023 that took effect in November 2023. Besides the new TPP and density thresholds, they extended license requirements to more than 40 additional countries viewed as posing diversion risk, including Country Groups D:1, D:4 and D:5, applied a presumption of denial for Macau and D:5 destinations such as China and a presumption of approval for most other newly covered destinations, and created a notification-based license exception, NAC, for certain lower-risk chips. [1][2]

BIS made further changes in April and September 2024. On 2 December 2024 it added controls on high-bandwidth memory (HBM) above a memory bandwidth density of 2 gigabytes per second per square millimetre, extended rules on foreign-made chipmaking equipment, and added 140 entities to the Entity List. GAO's review of the earlier rules found that companies reported compliance challenges such as unclear classification numbers and definitions. [6][5]

In January 2025 BIS published the Framework for Artificial Intelligence Diffusion. It would have let 18 allied destinations, including Japan, the United Kingdom, Germany and Taiwan, receive advanced chips under a license exception, placed caps on the total computing power that could be installed in most other countries, set up validated end-user programs for data centers, and controlled the weights of certain closed AI models. Most compliance obligations were due to start on 15 May 2025. [7]

On 13 May 2025, before those obligations applied, Commerce announced it was rescinding the rule, arguing that it would have burdened companies and relegated dozens of countries to second-tier status. BIS issued guidance at the same time warning about the use of certain Huawei Ascend chips, the use of US chips to train or run Chinese AI models, and diversion tactics. In July 2026, according to Morgan Lewis, Under Secretary Jeffrey Kessler said the administration no longer intends to replace the rule and is focusing on new rulemaking. [8][14]

BIS guidance dated 31 May 2026 clarified that licenses are still required for 3A090 and 4A090 items destined for entities headquartered in China, other D:5 countries or Macau, or whose ultimate parent is headquartered there, wherever those entities operate. The guidance lets bona fide data center operators continue ongoing use and servicing of such items until further notice. [9]

A BIS rule effective 15 January 2026 changed the review policy for exports to China and Macau from a presumption of denial to case-by-case review for chips with a TPP below 21,000 and total DRAM bandwidth below 6,500 gigabytes per second, a band that includes Nvidia's H200 and AMD's MI325X. Exporters must certify, among other things, that the chip is commercially available in the US, that US customers will not face delays, that China and Macau shipments stay at or below 50% of US shipments, that buyers apply know-your-customer checks including for remote access, and that an independent US lab has tested the chips. Re-exports and more capable chips remain under a presumption of denial. [10][11]

A presidential proclamation under Section 232, effective the same day, imposed a 25% tariff on advanced chips in similar performance bands that are not destined for the US supply chain, with exemptions for uses such as US data centers and research. The administration presented the combined policy as allowing lower-risk trade while preserving national security. [12]

Views on the change differ. Critics at the Center for a New American Security argue that end-use certifications are hard to verify, that the volumes allowed could substantially add to China's AI computing capacity, and that the supply certifications are difficult to meet while leading-edge capacity is tight. Supporters of the approach, including the administration, argue that case-by-case licensing with conditions is preferable to blanket prohibition. [13][12]

Separately, a BIS rule effective 10 July 2026 moved the United Arab Emirates into a more favourable country group while keeping license requirements for advanced chips except for listed approved entities, including UAE government bodies, two UAE-based AI companies and subsidiaries of US hyperscalers. [14]

In China, the 2026 opening has so far translated into limited volumes. Tom's Hardware reported in August 2026 that ByteDance and Tencent had each received about 10,000 H200 accelerators, around 2.5% of the more than 400,000 units that ByteDance, Alibaba and Tencent were collectively approved to buy, and that China's own planning agency now licenses such imports case by case. The same report cites a TrendForce projection that domestic chips will take nearly 90% of China's high-end AI chip market in 2026. [15]

Elsewhere, license requirements for D:1, D:4 and D:5 countries and entity-based rules mean that large GPU deployments in parts of the Middle East and Asia depend on licenses or approved-entity arrangements, as the UAE rule shows. With the diffusion rule rescinded, there are no country-level caps on installed computing power of the kind it proposed in 2025. [2][14][8]

From a market point of view, export controls act as one of several filters on where data center GPUs can be installed and rented. They help explain why the newest accelerators appear first and most widely in the United States and allied markets, and why some regions list older or lower-specification models. Rules have changed several times since 2022, so any regional picture should be treated as a snapshot.

Kovara does not provide compliance advice, but it can show you what is actually listed where. Check GPU availability by region to see which accelerators providers offer in each location, compare prices across regions, or ask Kova which regions list a particular GPU today.