FabricationTrade restrictions

US semiconductor export controls: AI chip compliance

An analytical guide to US semiconductor export controls. Learn how the 2026 BIS update impacts AI chip compliance, tariffs and European infrastructure.

, AI Researcher11 min read
In this article8
  1. The 2026 BIS update: TPP thresholds and the end of blanket denial
  2. The parallel 25% tariff on imported computing chips
  3. Supply quotas and the 50% domestic shipping cap
  4. Extraterritorial reach: the FDPR and third-country subsidiaries
  5. The independent US testing requirement
  6. Closing the cloud loophole: KYC and remote access controls
  7. Frequently asked questions
  8. Sources
The container ship CMA CGM Kerguelen at Burchardkai container terminal in Hamburg
Burchardkai, Hamburg. Export controls written in Washington are enforced at European loading docks. Photo: Matti Blume, CC BY-SA 4.0, via Wikimedia Commons

The 2026 BIS update: TPP thresholds and the end of blanket denial

On 15 January 2026, the US Department of Commerce's Bureau of Industry and Security (BIS) published a final rule in the Federal Register (91 FR 1684) revising the licensing policy for exports of advanced computing semiconductors to China and Macau. This regulatory action amends 15 CFR Parts 742, 744, and 748, marking a structural departure from the blanket presumption of denial established in October 2022 and expanded in October 2023. Under the updated framework administered by BIS[1], qualifying chips transition to a discretionary, case-by-case licensing review policy provided exporters fulfill strict technical, supply, and compliance verifications.

The case-by-case review applies specifically to advanced computing commodities classified under Export Control Classification Number (ECCN) 3A090 that meet two distinct technical ceilings: a Total Processing Performance (TPP) of less than 21,000 and a total dynamic random-access memory (DRAM) bandwidth of less than 6,500 GB/s. These thresholds bring commercial accelerator lines, such as the Nvidia H200 and AMD Instinct MI325X, within the scope of potential export licensing, whereas previous rules blocked direct shipments of these performance tiers to Chinese buyers. While this creates an avenue for approved sales, the scope is narrow: hardware exceeding either threshold remains subject to an outright presumption of denial.

Regulatory regimeLicensing standardKey performance thresholdsCovered hardware examples
October 2022 Interim Final RulePresumption of denialAggregate bidirectional transfer rate of 600 GB/s or more (original ECCN 3A090)Nvidia A100, H100
October 2023 Updated RulePresumption of denial / Notified Advanced Computing (NAC)ECCN 3A090.a covers ICs with a TPP of 4800 or more, or a TPP of 1600 or more with a performance density of 5.92 or moreNvidia H800, L40S, RTX 4090
January 2026 Final Rule (91 FR 1684)Case-by-case review (under conditions)TPP below 21,000 and total DRAM bandwidth below 6,500 GB/sNvidia H200, AMD Instinct MI325X

The transition away from blanket denial does not constitute an open general license. Exporters must submit comprehensive documentation proving that the destination facility meets physical security requirements, that the hardware was commercially available in the United States prior to the rule's publication, and that the transaction complies with all end-use restrictions under EAR Part 744. Furthermore, the rule explicitly restricts this case-by-case standard to direct exports originating within the United States, maintaining strict curbs across foreign-produced alternatives.

The parallel 25% tariff on imported computing chips

Concurrently with the BIS export rule, the White House announced on 14 January 2026 an immediate 25% tariff on a narrow group of semiconductors meeting the same performance thresholds, following completion of an investigation under Section 232 of the Trade Expansion Act of 1962[3]. The Department of Commerce determined that imported semiconductors, semiconductor manufacturing equipment and derivative products threatened to impair US national security, and the duty applies to imports of the same class of advanced computing hardware the export rule covers[4]. Imports that support the US technology supply chain and domestic manufacturing capacity are outside its scope.

The two measures move in the same direction without being formally linked. The export rule requires third-party testing in the United States by a US-headquartered lab, so a chip fabricated or packaged abroad has to reach US soil before it can be certified for export; the tariff separately raises the cost of importing that same class of hardware. Neither instrument references the other.

  • US data centers: Facilities dedicated to AI training, inference, simulation, or synthetic data generation requiring greater than 100 megawatts (MW) of new electrical grid load[5].
  • Domestic research and development: Systematic research programs conducted by academic, civil, or corporate laboratories within the United States.
  • Emerging growth enterprises: Startups meeting the statutory definition under 15 U.S.C. Section 77b(a)(19).
  • Civil and consumer applications: Dedicated non-data center deployments, including gaming, personal workstations, automotive systems, and factory robotics.
  • Public sector infrastructure: Direct hardware acquisitions by US federal, state, and local governmental bodies.

For enterprises evaluating global hardware procurement, these tariff rules alter the economics of compute allocation. While US domestic hyperscalers and qualified 100 MW facilities retain duty-free access, non-exempt international buyers and intermediary distributors face increased landed costs when routing advanced hardware through North American supply corridors.

Supply quotas and the 50% domestic shipping cap

To stop export licensing from draining compute needed by the domestic technology sector, BIS added a new paragraph (dd) to Supplement No. 2 of EAR Part 748 setting out applicatio[3]n and certification requirements, including explicit supply conditions. An applicant must supply evidence that the aggregate Total Processing Performance of the covered circuits exported to China or Macau will not exceed 50% of the aggregate TPP shipped to customers in the United States for US end use, measured from the date the product started shipping to commercial US customers. The cap is a ratio of compute, not of units[3].

Separately, paragraph (dd)(1)(i) requires the applicant to report the total number of units already shipped to commercial US customers, along with the product's TPP, total DRAM bandwidth, interconnect bandwidth, copackaged DRAM capacity and peak power consumption. Because the ceiling is measured in aggregate TPP, a slower domestic rollout tightens the exporter's foreign allocation, and a higher-performance part consumes the allowance faster.

  1. Baseline domestic reporting: Exporters must report how many units of the item have been sold to US customers for use in the United States as of the application date[3].
  2. No domestic delivery delays: Applicants must certify that US supply of the product is sufficient so that filling the China or Macau order will not delay or deprive any US end-user orders for advanced-node chips[3].
  3. Foundry non-diversion pledges: The applicant must certify that global foundry capacity that would otherwise produce similar-node or more advanced chips for US end users will not be diverted to the licensed export[1].
  4. Baseline specification disclosure: The application must document the item's performance, including total processing performance, total DRAM bandwidth, interconnect bandwidth and copackaged memory[3].

These conditions shift the compliance burden upstream to chip designers and supply chain planners. International buyers tracking NVIDIA AI chip pricing must account for these production controls, as artificial supply caps directly influence secondary-market availability and hardware lead times across non-US jurisdictions.

Extraterritorial reach: the FDPR and third-country subsidiaries

The reach of US export controls extends far beyond direct shipments from US ports. Under the Advanced Computing Foreign Direct Product (FDP) rule in EAR Section 734.9(h), items produced outside the United States can be subject to US export jurisdiction when they are the direct product of specified US-origin technology or software, including electronic design automation tools, or are made by a plant (or major component of a plant) that is itself the direct product of such US-origin technology[6]. Because advanced semiconductor manufacturing globally relies on US toolsets and electronic design IP, foreign-produced AI accelerators remain legally bound by the Export Administration Regulations.

A critical feature of the January 2026 framework is the asymmetrical licensing standard between domestic exports and foreign re-exports. While direct exports from the US customs territory benefit from case-by-case review, the licensing policy for re-exports and in-country transfers of these chips to Macau or any Country Group D:5 destination, including China, remains a presumption of denial[4]. Consequently, European and Asian intermediaries cannot independently re-export covered AI chips to Chinese entities without securing direct US authorization.

Jurisdictional triggerApplicable EAR ruleLicensing standardCompliance impact for European firms
Direct US export to China or Macau15 CFR Part 742 license review policyCase-by-case reviewRequires US testing, the 50% domestic supply ratio, and KYC certification
Foreign re-export / in-country transfer15 CFR Part 742 license review policyPresumption of denialProhibits secondary transfers from EU hubs without explicit BIS authorization
D:5-headquartered entity or its subsidiary (any location)15 CFR 744.23 / Supplement No. 2 to Part 748Presumption of denialBlocks routing compute to European subsidiaries of Chinese parent entities
Advanced computing foreign direct product rule15 CFR 734.9(h)License required by destination and end userBinds foreign foundries using US EDA software and fab equipment

Enforcement extends aggressively to corporate structures and intermediary routing. The rule maintains a presumption of denial for exports to entities that are headquartered in, or whose ultimate parent company is headquartered in, Macau or a destination specified in Country Group D:5, including end users located outside those destinations, so an order routed through a third-country subsidiary of a Chinese parent is still denied[1]. European companies managing global distribution networks risk substantial administrative penalties and denial orders if foreign subsidiaries transfer controlled compute assets without verified authorizations under US chip export controls.

The independent US testing requirement

A central technical hurdle under the January 2026 rule is the mandatory pre-shipment physical testing verification set forth in Supplement No. 2 to Part 748 (paragraph dd)(3). To qualify for case-by-case licensing, exporters must submit every export batch to an independent, qualified third-party testing lab headquartered in the United States, with the testing itself performed in the United States. In-house testing by the semiconductor manufacturer or software-based self-attestation is legally insufficient.

The laboratory must evaluate a representative batch sample selected independently by the testing facility rather than chosen by the exporter. This evaluation confirms that the technical capabilities and functions of the commodity match what the license application declares, covering total processing performance, total DRAM bandwidth, interconnect bandwidth and copackaged DRAM capacity.

  • Total Processing Performance (TPP): Measured across integer and floating-point data types to ensure total compute throughput remains strictly below the 21,000 ceiling.
  • Total DRAM bandwidth: Verification that high-bandwidth memory (HBM) interfaces operate below the 6,500 GB/s limit.
  • Interconnect bandwidth: One of the parameters the qualified lab must be able to confirm, covering chip-to-chip communications used for clustering and high-density pod formations[1].
  • Co-packaged DRAM capacity: Physical audit of memory density directly integrated on the package substrate.
  • Power consumption profile: Empirical verification of peak electrical draw at maximum sustained TPP to ensure thermal limits correspond to approved specifications.

To guarantee institutional impartiality, BIS imposes strict structural separation rules. The testing facility must be headquartered in the United States, must not be owned or controlled by entities headquartered in Country Group D:5 or Macau, and must possess zero financial interest, equity stake, or common ownership with the exporter, manufacturer, or end-user. Before export, the applicant certifies to BIS that every shipment has been reviewed by such a lab, and supplies the lab's name and US address as part of that certification.

Closing the cloud loophole: KYC and remote access controls

Beyond physical hardware consignments, the updated regulatory framework closes loopholes surrounding remote compute access. Cloud infrastructure providers and receiving data center facilities hosting advanced computing items must implement exhaustive Know Your Customer (KYC) frameworks to screen out prohibited end-users. These obligations apply to foreign facilities operating export-licensed hardware to deliver Infrastructure-as-a-Service (IaaS) solutions.

Under the rule's unique application requirements, license applicants must enumerate the physical security protocols and automated identity verification measures deployed by the receiving data center. The receiving operator must provide verified contractual guarantees that restricted actors cannot bypass physical controls or access virtualized compute slices for training high-capability models or military simulations.

  • Remote user disclosures: Applicants must provide a list of remote end users located in, or whose ultimate parent company is headquartered in, Belarus, China, Cuba, Iran, Macau, North Korea, Russia or Venezuela[4].
  • Military end-user prohibitions: The applicant must certify that the items are not intended for prohibited end users or end uses under EAR Part 744, including military and military intelligence end uses[4].
  • Managed IaaS obligations: The receiving facility must manage and limit Infrastructure as a Service, including AI model training and inference, to prevent unauthorised access to the covered hardware[1].
  • Algorithmic access restrictions: Technical barriers preventing prohibited third parties from accessing, querying, or fine-tuning models hosted on the licensed cluster.

These stringent compute screening rules carry significant consequences for European enterprise architecture. Organizations operating hybrid infrastructure must navigate how US extraterritorial jurisdiction intersects with local compliance, particularly when evaluating data sovereignty for AI inference against multi-jurisdictional cloud compliance requirements. For infrastructure planners, verifying hardware supply chains and remote API governance is now an essential element of baseline regulatory risk management.

Frequently asked questions

What is the Total Processing Performance (TPP) threshold?

Under the January 2026 BIS update, chips with a TPP below 21,000 and a total DRAM bandwidth below 6,500 GB/s are eligible for case-by-case export review, rather than a presumption of denial.

Does the 25% US tariff apply to European AI chip imports?

The 25% tariff targets imported semiconductors meeting the same performance thresholds as the export rule. Imports for qualifying domestic uses, including US data centers, US research and development, US startups, and public sector applications, are exempt.

How does the Foreign Direct Product Rule (FDPR) affect European foundries?

The advanced computing FDP rule extends US export jurisdiction to foreign-produced chips manufactured using US equipment, technology or design software, so European foundries and their customers need BIS authorisation for covered shipments even when no US border is crossed.

What are the independent testing requirements for US AI chip exports?

Before each shipment, the chips must be tested and certified by a qualified independent laboratory that is US-headquartered, has no ownership ties to embargoed countries, is not financially affiliated with the exporter or end user, and performs the testing on US soil.

Are infrastructure-as-a-service (IaaS) providers subject to BIS export controls?

Yes. Where the chips will be used in an IaaS context, the exporter must identify intended remote end users in countries of concern and the consignee must run Know Your Customer screening to block prohibited parties from remote access.

Sources

  1. Federal Register, Volume 91 Issue 10 (Thursday, January 15, 2026) govinfo.gov
  2. Commerce Imposes Significant New Controls on Advanced Semiconductors akingump.com
  3. BIS Revises Export Review Policy for Advanced AI Chips Destined for China and Macau morganlewis.com
  4. BIS Revises License Review Policy for Advanced Computing Commodities (AI Semiconductors) to China and Macau… sanctionsnews.bakermckenzie.com
  5. A Mixed Bag of Chips: Significant New Import and Export Changes for Advanced Semiconductors wsgr.com
  6. 15 CFR § 734.9 - Foreign-Direct Product (FDP) Rules. law.cornell.edu

Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score

Related reading

All Fabrication →