Summary
- The advanced-computing licence ticket is a formula, not a test: total processing performance is twice a “theoretical peak” throughput multiplied by the bit length of the operation and aggregated over every processing unit, and performance density divides that by the part’s own die area in square millimetres.
- The density term punishes tighter integration. Halve the applicable area at constant throughput and the ratio doubles, so the same operation count in less silicon can cross into the stricter paragraph — a layout decision becomes the licensing variable.
Start with the entry rather than the announcement. Under ECCN 3A090, paragraph .a covers integrated circuits with one or more digital processing units having total processing performance of 4800 or more, or of 1600 or more together with performance density of 5.92 or more. Paragraph .b takes the layer beneath: performance of 2400 or more but under 4800 with density from 1.6 up to but not including 5.92, or performance of 1600 or more with density from 3.2 up to but not including 5.92 (the amended entry as printed in February 2025). One paragraph keys on a single quantity. The other keys on a ratio.
Then the definitions, which is where this article actually lives. Technical Note 2 defines total processing performance as twice MacTOPS multiplied by the “bit length of the operation”, “aggregated over all processing units on the integrated circuit”. MacTOPS is “the theoretical peak number of Tera” operations per second “for multiply-accumulate computation (D = A × B + C)”. Two adjectives carry the weight: theoretical and peak. Not sustained, not measured under load, not benchmarked by anybody. A datasheet number of precisely the family that marketing departments have always printed.
Performance density then divides that figure by “applicable die area”, which “is measured in millimeters squared and includes all die area of logic dies manufactured with a process node that uses a non-planar transistor architecture” (Technical Note 2). The entry stops short of the hard cases, and the bureau’s own FAQ decides them: SRAM sitting on the logic die is not to be excluded, high-bandwidth memory is excluded as memory rather than logic, and for chiplets only each logic chiplet built on a non-planar node counts. The same FAQ labels the density definition Technical Note 4, while the February 2025 entry numbers it inside Technical Note 2. The numbering moved. The arithmetic did not.
Now the portion that deserves a ruler laid on the page. My arithmetic, not the text’s: at a performance figure of 1600, the density limb of .a admits a die no larger than 1600 ÷ 5.92, which is 270.3 mm². The .b edges fall at 2400 ÷ 1.6 = 1500 mm², at 1600 ÷ 3.2 = 500 mm², and at 2400 ÷ 5.92 = 405.4 mm². Because density is performance divided by area, halving the applicable area at constant performance doubles the ratio. A part with unchanged throughput can cross from .b into the density limb of .a on layout alone.
Direction matters here, and it runs against intuition. Fewer square millimetres is normally the achievement — the same work in less silicon, better yield, better thermals, better cost. Under this parameter it is the liability. At 4800 or more the first limb of .a captures the part with no density condition at all, which means the ratio decides something only between 1600 and 4799. Inside that band, integration is what the gate measures.
Then there is the shape of the printed bands, which do not tile the space. A part at performance 3000 and density 1.0 fails .a on both limbs, sitting below 4800 and below 5.92, and fails .b on both, sitting below 1.6 and below 3.2. A design can fall outside both paragraphs on the printed numbers alone. That is a reading of the ranges as written, and it is the kind of gap that later guidance fills or a revision closes.
The gate itself is spread across two sections. The licence-requirement table states “RS applies to 3A090.a” with “To or within any destination worldwide, see Sec. 742.6(a)(6)(iii)(A) of the EAR”, and that paragraph opens “(A) Worldwide license requirement. A license is required to export, reexport, or transfer (in-country) items specified in” a list that includes 3A090.a (15 CFR 742.6). For .b the destination list narrows to the UAE and Country Groups D:1, D:4 or D:5, excluding places also in A:5 or A:6. Separately, 15 CFR 744.23 adds an end-use control: “In addition to the license requirements for items specified on the CCL, you may not export, reexport, or transfer (in-country) without a license any item subject to the EAR described in paragraphs (a)(1) through (4) of this section when you have” knowledge of the covered destination, end use or end user. Its (a)(3) is headed “Advanced computing items”, and (c) allows no licence exceptions past the prohibition except the ones listed.
Which section currently carries the .a requirement is less settled than the paragraph numbers suggest. The bureau’s enforcement guidance of 31 May 2026 says the requirement “was first introduced on November 17, 2023” and “was implemented via an end-user control in Sec. 744.23(a)(3) of the EAR”, that the January 2025 AI Diffusion Rule moved the .a requirement into 742.6 as a worldwide requirement, that the bureau would not enforce that rule’s new compliance requirements, and that a requirement nonetheless continues under 742.6(a)(6)(iii)(A) for .a items going outside the United States to entities headquartered in, or ultimately parented in, D:5 or Macau. Both sections carry cross-references to 2026 amendments I did not read: 91 FR 46260 for part 742, and for part 744 a history through 91 FR 1687 with a later rule said to move the .a requirement back into 744.23(a)(3)(i)(A). Which text binds today depends on an effective date I did not establish. The October 2023 interim final rule that introduced this parameter, replacing a “bits × TOPS” metric with performance values, went unread as well; its content here is a description, not a quotation.
Nothing in the documents read for this piece states how the bureau verifies a performance figure or a die area. That is an argument from silence inside the provisions actually read, not a finding that no verification mechanism exists. What the record does show is where the method lives: the counting rules that decide whether on-die SRAM or high-bandwidth memory occupies the denominator come from an FAQ issued by the same bureau that writes the entry.
The named company documents the consequences without ever printing the parameter. Its quarterly report for the period ended 27 October 2024, filed on 20 November 2024, says the October 2023 requirements cover products “exceeding certain performance thresholds, including, but not limited to, the A100, A800, H100, H800, L4, L40, L40S and RTX 4090”, and that “To date, we have not received licenses to ship these restricted products to China.” A list of part numbers stands where the threshold would be. The reader learns which products fell inside. The reader does not learn the number that put them there.
By the following fiscal year the parameter had changed species. The annual report for the fiscal year ended 25 January 2026 records that “in April 2025, the USG informed us that it requires a license for export to China (including Hong Kong and Macau) and D:5 countries, or to companies headquartered or with an ultimate parent therein, of our H20 integrated circuits and any other circuits achieving the H20’s memory bandwidth, interconnect bandwidth, or combination thereof.” Not a computed throughput, not a ratio over area: a shipped product used as the yardstick for everything that matches its bandwidth. A $4.5 billion first-quarter charge for H20 excess inventory and purchase obligations follows in the same filing, then licences from August 2025 covering certain H20 shipments, an expectation attributed to USG officials of 15% or more of revenue from licensed sales with no codifying regulation published, and a February 2026 licence for small amounts of H200 to specific China-based customers, conditioned on US inspection before shipment and carrying a 25% tariff on importation into the United States.
Where those sentences sit is part of the finding. The 2024 passage is under Item 2, Management’s Discussion and Analysis, in a subsection headed “Global Trade”; the fiscal 2026 passage sits in “Government Regulations” inside Item 1, Business. Neither was read from a risk-factor heading, and the April 2026 quarterly report states that, other than the factors it lists, “there have been no material changes from the risk factors previously described under Item 1A” of the annual report. The licensing constraint is disclosed as an operating condition of the business, not as a hazard in the risk section.
So the checkable point is the definition, and anyone holding a design file can check it: where the threshold is denominated, who supplies the numerator, who supplies the denominator, and how the same silicon moves across. Nine documents sit behind this piece. Not one of them is a certificate of measurement. Ask why a particular accelerator needed a licence and the honest answer is a division — with the dividend and the divisor both supplied by the company selling it (NVIDIA’s directory entry).
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






