Yes, AI chips are export-controlled. For US-origin hardware the controls sit in the Export Administration Regulations (EAR), which the Bureau of Industry and Security (BIS) administers. Most high-performance GPUs and accelerators fall under ECCN 3A090, and the boards and servers built around them fall under 4A090. A chip’s total processing performance (TPP) and performance density decide whether it is caught. Whether a particular shipment then needs a licence depends on the destination, the parties to the deal, who owns them and the end use. That last step is a legal determination: engineering supplies the parameters and trade-compliance counsel signs it off.

Which AI chips fall under US export control: TPP and the 3A090 bands

TPP is 2 × MacTOPS × the bit length of the operation, summed over every processing unit on the chip at the highest rate the manufacturer claims. 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.

ECCN 3A090.a catches an integrated circuit with a TPP of 4,800 or more, or a TPP of 1,600 or more with a performance density of 5.92 or more. ECCN 3A090.b catches a TPP from 2,400 up to 4,800 with a density from 1.6 up to 5.92. High-bandwidth memory has its own entry, 3A090.c, which BIS added in December 2024. Chips that co-package HBM with processing logic are assessed under .a or .b instead.

Worked example: reading a datasheet against the thresholds

1. Start from the rule’s own example. 300 teraFLOPS at 16 bits gives a TPP of 4,800, and so do 600 tera integer operations at 8 bits. The working: a FLOP count treats the multiply and the add as separate operations, so 300 teraFLOPS is 150 MacTOPS, and 2 × 150 × 16 = 4,800.
2. Use the dense, manufacturer-claimed peak. TPP uses the highest rate the manufacturer claims, and a chip with both sparse and dense ratings is rated on its dense figure.
3. Check density when TPP is between 1,600 and 4,800. At exactly TPP 1,600, the 5.92 density line is crossed when the applicable die area is about 270 mm² or less (1,600 ÷ 5.92 ≈ 270). For 3A090.b at TPP 2,400, a density of 1.6 is reached at 1,500 mm² or less (2,400 ÷ 1.6).
4. Classify the board or server separately. Boards and servers built around these chips fall under their own entry, 4A090.

Why a benchmark score is not a classification input

TPP is set in the rule at the manufacturer’s highest claimed rate. A benchmark measures what an executor actually delivers, and the two answer different questions. LynxBenchAI scores each test over one timed window after a discarded warm-up, on stock PyTorch. In the 26Q3 results the NVIDIA H200 scores 1045.0 GT, 1914.6 in inference and 480.8 in compute. The January 2026 rule names that same device as an example of a chip inside its case-by-case tier. None of those scores changes the device’s legal position, and a LynxBenchAI result cannot stand in for the third-party lab certification the rule requires.

The January 2026 case-by-case tier for China and Macau

In January 2026 BIS moved direct exports from the United States to China or Macau from a presumption of denial to case-by-case review. The move covers chips with a TPP under 21,000 and a total DRAM bandwidth under 6,500 GB/s, such as the NVIDIA H200 or AMD MI325X. To qualify, the applicant must certify all nine conditions in Supplement No. 2 to Part 748, paragraph (dd):

• the chip is commercially available in the United States;
• US supply is sufficient, so no US order is delayed;
• the exports do not divert foundry capacity from US customers;
• aggregate TPP shipped to China and Macau is no more than 50% of the aggregate TPP of the same product shipped to US end users;
• there is no prohibited end use or end user;
• the ultimate consignee keeps documented KYC and physical security measures;
• the applicant lists remote end users in Belarus, China, Cuba, Iran, Macau, North Korea, Russia and Venezuela, or whose ultimate parent is headquartered in one of them;
• each shipment is reviewed before export by a US-headquartered third-party lab with no stake in the deal, testing in the United States (a representative sample of each batch is allowed);
• the applicant reports TPP, total DRAM bandwidth, interconnect bandwidth, copackaged DRAM capacity and peak power at max TPP.

Reexports and in-country transfers of these chips to Macau or Country Group D:5 stay under a presumption of denial. Morgan Lewis (January 2026) advises export managers to build lead time and budget for the per-shipment lab step and to hold every shipment until the lab certification has been submitted to BIS.

Can this accelerator SKU legally go where we plan to ship it?

The answer is a legal determination for each SKU, each destination and each set of parties. It depends on the item’s technical classification and on the destination, end user and end use combined. A marketplace listing, a vendor datasheet or an internal engineering review can feed that determination, but none of them is the determination. Shortlisting SKUs that fit your power budget and marketplace is a separate sourcing exercise. Export-control clearance comes after it, and the shortlist does not replace it.

Eligibility of each SKU in our jurisdiction

Destination screening starts with the Commerce Country Chart and the country groups. Since the 2023 rule, 3A090 and 4A090 items need a licence to Country Groups D:1, D:4 and D:5 unless the destination is also in A:5 or A:6. Applications for Macau and D:5 destinations are reviewed under a presumption of denial. Country status also moves. A rule effective 10 July 2026 moved the United Arab Emirates into A:5, yet licences are still required for 3A090 and 4A090 items going there, except where the consignee and all end users are UAE government bodies or entities on a new approved list. In May 2025 BIS also imposed additional licence requirements on AI chip exports to Malaysia and Thailand, citing known diversion.

A domestic or locally sourced accelerator does not automatically remove EAR exposure. The EAR reaches all US-origin items wherever they are, and foreign-made commodities that incorporate controlled US-origin content, in some cases in any quantity. The foreign direct product rules extend jurisdiction to some foreign-produced items made with specified US technology or software. The performance side of a domestic-versus-imported trade-off is a measurement question. If the domestic part is not in the LynxBenchAI hardware directory, the LynxBenchAI Personal Edition measures that hardware on the public PyTorch stack, without running your own application.

How do we screen a new destination and its parties before the first shipment?

Screen every party to the transaction, not only the end user. BIS’s restricted-party lists apply to all items subject to the EAR, including EAR99, and they apply even where the item’s classification and destination would otherwise need no licence. A listed applicant, purchaser or consignee triggers them just as a listed end user does. The lists that matter most here work differently:

• Entity List: imposes licence requirements and limits most licence exceptions for listed persons and addresses.
• Unverified List: no licence exceptions may be used, and a statement is needed from the party even for shipments that need no licence.
• Military End-User List: not exhaustive, so you must do your own diligence on whether an unlisted party meets the definition.

The US government’s Consolidated Screening List gathers parties under US export restrictions, including the Entity and Unverified Lists. A likely match calls for further due diligence before the deal proceeds. BIS’s May 2025 counter-diversion guidance adds two further steps. Check whether the customer’s ultimate parent is headquartered in D:5 or Macau, and obtain an end-user certification that names every party and the intended end use. Parties also may not “self-blind” to avoid a licence requirement. Where these checks sit among the other gates of a hardware shortlist is covered in compliance and procurement gates for shortlisting AI accelerators.

What the BIS 50% affiliates rule changes about who may receive AI chips

The affiliates rule took effect on 29 September 2025. It extended Entity List restrictions to any entity that one or more listed entities own 50% or more of, directly or indirectly, individually or in aggregate. It did the same for listed military end users and certain sanctioned parties, and it applies to every item subject to the EAR. BIS stayed the rule from 10 November 2025 to 9 November 2026, with the provisions written back into the EAR from 10 November 2026. That date is not settled. On 23 September 2026 the US announced a two-month extension of the related US-China agreement to 10 January 2027. As of 1 October 2026, Miller Canfield noted that BIS had not published a notice moving the affiliates rule’s effective date. When ownership cannot be determined, BIS guidance indicates that a licence may be needed before proceeding.

Keep this separate from the headquarters rule, a distinct licence requirement. Since 17 November 2023, a licence has been required to export advanced computing items to entities headquartered in D:5 or Macau, or with an ultimate parent headquartered there, even when the entity itself sits elsewhere. In practice, under the January 2026 rule, an export to a Singapore customer owned by a Chinese parent is still reviewed under a presumption of denial.

Who signs off that we are compliant before we award the framework agreement?

TechnoLynx works on the engineering side, which supplies the parameters for the compliance file, but trade-compliance counsel signs off, and the person with that authority should be named before the award, not found afterwards. A sign-off that covers several member states or markets should rest on a documented opinion per SKU and per destination, because licence requirements depend on the destination. The stakes are not abstract. The maximum administrative civil penalty is the greater of $374,474 per violation or twice the transaction value, and a wilful violation can carry a criminal fine of up to $1,000,000 per violation and up to 20 years in prison. Classifications also go stale. BIS warns that ECCNs change over time, so review each classification against the current Commerce Control List.

EU dual-use, EAR or ITAR: which regime governs the shipment

Regulation (EU) 2021/821 is the EU’s dual-use export regime. Whether a shipment to an EU distributor needs an authorisation under it is a determination against the EU control list. EU counsel should make it, because the sources behind this page cover the US side. On the US side, items exclusively controlled by the Department of State under ITAR are not subject to the EAR. The Commerce Control List order of review puts the “600 series” and 9x515 analysis ahead of the other ECCNs.

Does the EAR still apply when a US-origin GPU module is built into a product shipped from outside the US?

Often, yes. The EAR follows US-origin items wherever they are located, and foreign-made commodities that incorporate controlled US-origin commodities can be subject to it in any quantity or above de minimis levels. For some combinations there is no de minimis level at all, such as certain foreign-made computers containing US-origin 3A001 semiconductors bound for specified destinations. So for a product shipped from a third-country factory, do not assume the EAR no longer applies. For AI chips headed to China, reexports from third countries remain under a presumption of denial.

Which licence exceptions can cover a commercial export of AI chips, and when they cannot

License Exception STA exists. When the UAE joined A:5, STA became available to the UAE government and approved entities, yet 3A090 and 4A090 items still needed licences outside the listed parties. License Exception ENC is the main exception for Category 5 Part 2 encryption items and covers most encryption products to most destinations once reporting and classification requirements are met. The January 2025 AI Diffusion Rule introduced License Exceptions AIA and LPP. Commerce announced that rule’s rescission on 13 May 2025 and said a replacement rule would follow, so those exceptions should not be relied on without counsel confirming their status. Confirm the conditions of License Exception NAC with counsel against the current EAR text.

Several situations close off exceptions regardless of the item. No exception may be used for parties on the Unverified List. Most exceptions are limited for Entity List parties. For entities headquartered in D:5 or Macau, BIS says to seek licences unless a §740.2(a)(9)(ii) exception is available.

What ECCN applies to a Jetson-class module once it is integrated into our drone?

The supplier’s classification covers the supplier’s module and nothing more. NVIDIA states that Jetson modules and developer kits are ECCN 5A992.c. That entry is a unilateral US control within Category 5 Part 2, the information-security category. It is a documented baseline for the component, but it does not classify the drone, the flight-control board or the system as integrated. Each of those is classified on its own nature and technical parameters.

The order of review matters for a drone. Under the Commerce Control List order of review, the analysis starts with the 9x515 and “600 series” entries before the other ECCNs. USML Category VIII lists, among other items, certain autonomously cooperating UAVs and mission computers specially designed for aircraft in that category. “Specially designed” has slightly different definitions under the EAR and ITAR. BIS notes that a part that is not specially designed under ITAR can still be specially designed under the EAR, for example under a 600-series .x paragraph.

If the RF link implements cryptography, it falls under Category 5 Part 2, where most items sit in 5A002, 5A004 or 5A992 and ENC is the main exception. A later fleet revision that swaps the RF link or the compute module reopens the classification.

Is a formal CCATS ruling from BIS worth getting, or is self-classification defensible?

There are three routes to an ECCN. You can ask the manufacturer, producer or developer. You can self-classify, which needs a technical understanding of the item and familiarity with the ten-category structure of the Commerce Control List. Or you can request an official classification from BIS through the SNAP-R portal. A request must recommend a classification and explain it against the ECCN’s technical parameters, or name the ambiguities that prevented a recommendation. One request covers up to six items, with exceptions for related items. BIS assigns each classification a CCATS number. Neither the classification nor the number may be cited as evidence that the US government has determined the items are subject to the EAR. BIS will also advise whether a licence is required for a given end use, end user or destination, but that advice does not bind BIS to grant one.

Situation	Self-classification	Formal CCATS request
Component with a current supplier ECCN, checked against today’s CCL	Usually defensible, with the check documented	Rarely worth the wait
Integrated system with no clear precedent (RF link, flight control, possible 600-series pull)	Weak footing	Worth it for certainty
Low risk tolerance, large contract or a counterparty asking for a ruling	Possible, but hard to defend	Preferred
Need to know whether a specific deal needs a licence	Not answered by classification alone	Ask for an advisory opinion with the request, knowing it is non-binding

A repeatable process for new distributor markets runs the same checks before every first shipment: country chart, party screening, ultimate-parent check and end-user certification. Counsel is brought in only when a destination, party or item is ambiguous or high-risk. Re-run the classification whenever the GPU module, the RF link supplier or any other controlled component changes, and review it on a calendar basis as well, because the lists and rules change underneath a checklist that nobody maintains.

How exposed are we to allocation cuts and price rises over a five-year product life?

Allocation and price risk belong to a specific vendor and product line, and no spec sheet tells you about them. For each candidate platform, set the vendor’s published lifecycle and longevity statements against what it actually did in past shortages. Then ask whether the vendor will put a multi-year supply commitment or a second source in a contract. A roadmap slide is not a guarantee, and an allocation assessment goes stale quickly, so date it and re-run it before each volume commitment. If no vendor will commit, a dual-sourcing plan should exist before mass procurement, even though a second source adds validation cost.

What contractual guarantees protect us if a GPU vendor misses delivery on a multi-year contract?

Mechanisms include committed volumes and dates, liquidated damages for late delivery, substitution rights and SLA credits. Compare vendors on how enforceable those terms are, not on list price, and prefer a concrete remedy over “best effort” language. Given supply constraints, no single vendor’s unconditional multi-year guarantee is risk-free, and penalties reduce the damage of a slip without preventing it. On export-sensitive contracts, also ask how the contract allocates the risk of a licence being denied or a rule changing mid-term. A contract lawyer needs to judge enforceability, and a technical evaluation cannot do it.

Lock-in risk: SDK openness, second sources and end-of-life record

Lock-in is decided by three things that have nothing to do with runtime performance. The first is how open the SDK and software stack are. The second is whether a pin-compatible or functionally equivalent second source exists. The third is how the vendor has handled long-term support and end-of-life in the past. Rate each candidate on all three. For a high-risk part that still gets selected, add contractual mitigations such as source escrow or support SLAs. Benchmark results do not measure any of this, and a published score says nothing about how long a module will be supported. A second source is rarely a drop-in swap, so expect re-validation. Whether a non-NVIDIA software stack is mature enough for your workloads is a separate question about toolchains, not about contracts.

Qualifying a fallback accelerator in parallel with the primary choice

Qualify a backup board in parallel with the primary. Pick it for architectural similarity to the primary, to keep re-engineering small, and limit the validation to the core model ported and benchmarked. Full qualification is only worth it if the primary’s supply risk materialises. TechnoLynx does that porting and validation work through the AI Porting & Deployment Pack. The first step is defining the smallest validation scope that fits the primary’s timeline. Then the engagement selects a backup board compatible with your pipeline, ports the inspection model to its SDK, benchmarks accuracy and latency against the primary and documents the result for a fast switch-over.

LynxBenchAI’s published results help with the hardware side of that choice. They show how each device runs the public PyTorch stack your own models run on, so they are broadly indicative for custom PyTorch models. They do not run or measure your application. The 26Q3 results cover 48 devices, mostly desktop, workstation and data-centre GPUs plus some CPUs and integrated graphics, and they do not include Hailo or Qualcomm QCS parts. Where a candidate is covered, read its page in the directory. Where it is not, the Personal Edition is the way to measure the hardware.

This route does not suit every team. If the backup board needs a substantial re-architecture of the pipeline, it may not be ready in time when you actually need it. Parallel work also takes engineers away from the primary path. Teams with spare in-house porting capacity, or a primary vendor willing to contract a second source, may not need an outside engagement. Performance parity between the two boards still has to be validated, not assumed, and the backup needs its own export classification before anyone relies on it.

Vendor roadmap stability for a long-lived framework contract is a due-diligence question, not a benchmarking one. Check contractual support terms, the vendor’s history of keeping past roadmap promises, independent tracking, and the hardware’s export-control and sovereignty eligibility for the programme. Treat that eligibility as something that can change mid-contract. The affiliates rule took effect on 29 September 2025, was stayed from 10 November 2025, and its resumption date was still unconfirmed as of 1 October 2026.

The facts engineering owes the compliance file

Applicants report TPP, total DRAM bandwidth, interconnect bandwidth, copackaged DRAM capacity and peak power at max TPP. They must also explain any change to those specifications since the model launched. That list is a useful template even for shipments far from China. Engineering produces the parameters, the rated dense peak, the die area and a record of every firmware or component change. Counsel turns them into an ECCN, a destination analysis and a licence decision.

FAQ

Is an ECCN the same as an HS code?

No. An HS code classifies goods for customs tariffs, and an ECCN classifies items for export control on the Commerce Control List. An ECCN is a five-character code whose first digit gives the category and whose letter gives the product group.

Is EAR99 an ECCN?

EAR99 is the designation for an item that is subject to the EAR but not described by any ECCN. EAR99 items can still need a licence when a restricted party or a prohibited end use is involved.

Do export controls apply to software?

Yes. ECCNs cover commodities, software and technology, so software can carry its own classification.