Skip to content
Komposite News
Markets · Technology · Capital
Thursday, September 10, 2026Global EditionGot a tip?
Export Controls

The H200 licence rule, line by line

The Treasury Secretary said on Friday that China should get no Nvidia chips above the H100. The rule Commerce published in January sets its line by performance, and the H200 passes it with room to spare. Our reading of the rule and Nvidia's spec sheets.

Dana Whitfield · House byline
November 14, 20264 min read
A black probe station with a binocular microscope mounted over a small test board, against a white backgroundRgarrigan, CC BY-SA 4.0, via Wikimedia Commons
A probe station with a microscope above a test board, used to check the electrical behaviour of chips and circuit boards, in a file photograph from 2016.Photograph via Wikimedia Commons

The US rule that allows Nvidia's H200 to be licensed for China, published by the Commerce Department on 15 January 2026, does not name a ceiling by product. It sets two numbers: total processing performance below 21,000 and memory bandwidth below 6,500 gigabytes a second. By our arithmetic from Nvidia's own specifications, the H200 sits at 15,832 and 4,800, well inside both.

That detail matters after Friday. Scott Bessent, the Treasury Secretary, said on CNBC: "I am frustrated when President Trump made it very clear that China should not receive any NVIDIA chips above H100s," according to the network's transcript. On the measure the rule uses for computing power, the H200 is not above the H100. The two chips have the same processing figures. What the H200 adds is memory: 141 gigabytes against 80, and 4.8 terabytes a second of bandwidth against 3.35, according to Nvidia's product pages.

What does the rule actually change?

It moves licence applications for chips under the two thresholds from a presumption of denial to case-by-case review when the buyer is in China or Macau. The rule names "the NVIDIA H200 or AMD MI325X" as examples. It is a review standard, not a general licence: every export still needs a licence from the Bureau of Industry and Security, and the bureau and other reviewing agencies decide each one.

The conditions are specific. The exporter must certify how many units have shipped in the United States, that exports will not delay American orders or divert foundry capacity from chips for American customers, and that "the aggregate shipments of the product to China and Macau will be no more than 50% of the total product shipped to customers for end use in the United States of that product". The buyer must run know-your-customer checks "to screen and prevent unauthorized remote access". Every shipment must be reviewed before export by "a qualified third-party testing lab" headquartered in the United States, which may test a sample from each batch. The applicant must also list remote users in Belarus, China, Cuba, Iran, Macau, North Korea, Russia and Venezuela.

The relaxation is narrower than its headline. It covers exports from the United States to buyers located in China or Macau. Re-exports from third countries keep the presumption of denial, and so do exports to companies headquartered, or with a parent headquartered, in Macau or a country in the group the regulations call D:5, which includes China, when the buyer is located somewhere else, such as a subsidiary in South-East Asia. The bureau classed the rule as "significant" under the executive order on regulatory review, and published it without prior notice or comment, which the export-control statute allows.

Komposite's arithmetic from Nvidia's published specifications for the H100, H200 and HGX B200. Nvidia quotes FP8 performance with sparsity; dense figures are half. Total processing performance (TPP) is dense operations per second times bit length, the measure in Export Control Classification Number 3A090. Bandwidths converted at 1 TB/s = 1,000 GB/s. Our figures are not an official classification.
ChipDense FP8 (teraFLOPS)TPP, our arithmeticMemory bandwidth (GB/s)Inside the January rule's limits?
H100 SXM1,97915,8323,350Yes
H100 NVL1,670.513,3643,900Yes
H200 SXM1,97915,8324,800Yes
H200 NVL1,670.513,3644,800Yes
B200 (per GPU in an HGX B200)4,50036,000Not listedNo: TPP above 21,000
Rule's ceilingBelow 21,000Below 6,500

How do the chips compare with the thresholds?

The H200 uses about three-quarters of each allowance. Its total processing performance is 75 per cent of the 21,000 ceiling and its bandwidth 74 per cent of 6,500, by our division. The H100, at 3,350 gigabytes a second, uses about half the bandwidth allowance. Nvidia's Blackwell B200 is excluded on processing alone: at 4.5 petaFLOPS of dense FP8 per chip, half the sparse figure Nvidia publishes for an eight-chip HGX board, its total processing performance is about 36,000.

Total processing performance is the export-control measure of raw arithmetic: operations per second multiplied by the bit length of each operation. Nvidia's headline FP8 numbers assume sparsity, a technique that skips zeros, and the company states that dense performance is half. On dense figures, the H100 and H200 in their SXM form both come to 1,979 teraFLOPS of FP8, which gives 15,832.

Who holds the pen on licences?

Commerce. The rule was issued by the Bureau of Industry and Security, and Bessent made the same point about the separate question of blacklisting on Friday, saying that "Commerce BIS has the pin", as the CNBC transcript renders it. Commerce did not respond to Fortune's request for comment on his remarks. The rule itself expected modest use: the bureau estimated "100 more license applications per year" because "industry is more likely to submit licenses when there is a case-by-case review".

Use has been modest in Nvidia's accounts too. Its filing for the quarter to July recorded shipments under the H200 programme at less than 1 per cent of data-centre revenue, as Komposite reported in our reading of that 10-Q, and Beijing has restricted purchases by Chinese buyers. Enforcement has meanwhile become Washington's main tool, as we noted when export controls shifted from rulemaking to settlements.

The rule has been in force for 303 days, by our count. Nvidia reports quarterly results on Tuesday 17 November, the first chance to see whether H200 shipments to China have moved from the margin of its accounts.

MORE ON THIS STORY