🔬 Semiconductors
758 H100s Is the First Count Past 12 Million TPP
The November 19 rule requires a license once access passes 12 million aggregate TPP. Dense math puts that at 758 H100s, or 334 B200s.
Seven hundred fifty-eight.
A Bureau of Industry and Security rule effective November 19, 2026, requires a license for covered advanced-computing items when you know they are going to a data center, anywhere, that is providing or will provide access to more than 12,000,000 aggregate total processing performance, at one site or through linked sites, to or for a company headquartered in Macau or Country Group D:5, or whose ultimate parent is. Divide 12,000,000 by 15,832 and the quotient is 757.96, so 757 Nvidia H100 SXM chips of that access stay under the line and 758 is the first whole count over it.
Dense peak, not the sparse headline
Aggregate TPP is the sum of each chip's total processing performance. Technical Note 2 defines TPP as 2 times MacTOPS times the bit length of the multiply inputs, takes the highest product across bit lengths, and scores a chip that does both sparse and dense math on the dense value. BIS dated its press release November 16 and said Treasury and the National Security Council directed the text. Morrison Foerster records public inspection that day, effect on November 19, and comments due January 19, 2027, the date the Federal Register prints.
Sparse does not count.
Nvidia's H100 page lists the SXM at 3,958 teraFLOPS of FP8 and footnotes the row "with sparsity." Nvidia's Hopper note says that feature doubles tensor-core throughput, so half of 3,958 is 1,979 dense teraFLOPS, and an FP8 multiply's inputs are 8 bits, which makes 1,979 times 8 equal 15,832 TPP. That is 15,832 divided by 21,000, or 75.39 percent, of the case-by-case ceiling for a China or Macau end user: TPP under 21,000 and DRAM bandwidth under 6,500 GB/s. Applications outside that window, including ones for a D:5 or Macau headquarters, carry a presumption of denial, while section 744.24 is case by case.
Different door.
H200 SXM pages print that same 3,958 FP8 figure, the same sparsity footnote, 141 GB, 4.8 TB/s, and up to 700 watts. Memory is not part of TPP, so an H200 shares the 15,832, and 4.8 TB/s is 4,800 GB/s, or 4,915.2 if a terabyte is 1,024 gigabytes, both under 6,500, so an H200 still fits both windows of that China review.
| Chip, as published | Dense rate used | Bits | TPP | Versus 21,000 | First count over 12,000,000 |
|---|---|---|---|---|---|
| H100 SXM and H200 SXM, FP8 | 1,979 TFLOPS | 8 | 15,832 | 75.39% | 758 (over by 656) |
| HGX B200, per GPU, FP4 or FP8 | 9,000 or 4,500 TFLOPS | 4 or 8 | 36,000 | 171.43% | 334 (over by 24,000) |
| HGX B300, per GPU, FP4 | 13,500 TFLOPS | 4 | 54,000 | 257.14% | 223 (over by 42,000) |
| 3A090.a floor, the rule's example | 600 TOPS at 8 bits | 8 | 4,800 | 22.86% | 2,501 (2,500 lands on the line) |
Nvidia's HGX table lists an eight-GPU B200 at 72 dense petaFLOPS of FP4 and 72 sparse petaFLOPS of FP8, with dense equal to half the sparse spec. Per GPU that is 9,000 teraFLOPS at 4 bits, or 4,500 at 8 bits, and both products are 36,000 TPP. B300's FP4 cell is 108 dense petaFLOPS across eight GPUs, 13,500 teraFLOPS each, times 4 bits is 54,000, which beats that column's FP8 product, so the highest-TPP instruction selects 54,000. Its INT8 cell, 3 POPS with the sparse footnote, is smaller once halved.
757 times 15,832 is 11,984,824, short by 15,176. 758 times 15,832 is 12,000,656, over by 656. Chips at the rule's 4,800 example hit 12,000,000 exactly in a stack of 2,500, which is not "more than," so 2,501 is the first count over that floor. Not a campus.
At the top of the H100 SXM label, up to 700 watts, 758 chips are 530,600 watts, or 530.6 kilowatts of GPU nameplate. DGX B200 lists 64 TB/s of memory bandwidth across eight GPUs, which is 8 TB/s each: 8,000 GB/s in thousands, or 8,192 if a terabyte is 1,024, and either one clears 6,500, so a B200 misses the China bandwidth window as well as the 21,000 TPP window.
A place license, then an access license
Covered items need a license to or within Indonesia, Malaysia, Singapore, Brazil, and Thailand, unless the ultimate consignee, the end user, or its ultimate parent is headquartered in the United States and the exporter files the section 743.10 report by the first business day of the next month. Baker McKenzie, in its November note, read that same list, the same 12,000,000 test, and that reporting exclusion. Section 744.24 does not repeat the U.S.-headquarters exit. A data center that will not hand over its know-your-customer file is, under Note 3, a red flag that presumptively supplies knowledge, and linked sites include halls joined only by a network.
Only one carve-out sits inside that section: a commercial recommender used solely to rank, filter, or suggest products, services, or content. That definition excludes any system designed or used to generate novel text, image, audio, video, software, or other content, even if the system also recommends.
On May 31, 2026, BIS told exporters the headquarters license still applied outside D:5 and Macau, and that bona fide operators did not have to stop use, storage, disposal, or servicing "until further notice." Holland & Knight noted that "bona fide" was never defined and no date was given. A June 17 FAQ said the guidance covered 3A090.a and the related .z items, and a June note at International Trade Law repeated the until-further-notice line. Paragraph (b)(3) now says a data center already providing a significant volume of this compute must obtain a license "with 90 days" of the effective date, or face future export restrictions, if it wants to keep providing that access.
November 19 plus 90 days is February 17, 2027: eleven days finish November, December takes 31, January takes 31, and 11 plus 31 plus 31 is 73, so the remaining 17 days land on February 17, a date the rule itself does not print.
BIS prices the net paperwork change, which also subtracts two repealed license exceptions, at about 970 added submissions, 440 hours, and $16,720, so 16,720 divided by 970 is $17.24 a form and 440 divided by 970 is 0.454 of an hour, or 27.2 minutes. November's rule says this control extends the October 7, 2022 interim rule and the October 2023 rule at 88 FR 73458, and it narrows how a designer located or parented in Country Group D:1, D:4, or D:5, or in the five new countries, can beat the presumption that an applicable advanced logic chip is 3A090.a unless that designer is approved or authorized.
What You Can Do
Renting GPUs, record the customer's ultimate-parent headquarters before the next box ships, sum dense TPP of the access those D:5 or Macau parents receive, and fold digitally linked halls into that sum, because a customer who withholds the file is the red flag in Note 3. Shipping into the five countries for a U.S.-headquartered cloud, file the section 743.10 report by the first business day of the next month if you want the destination license to fall away, and do not treat that filing as a waiver of the 12,000,000 test. A system that generates novel text, image, audio, video, or software is outside the recommender definition even when it also ranks results. Goods already aboard a carrier on November 19, under an actual order, keep the old eligibility only through December 21, 2026, comments close January 19, 2027, and our 90-day addition lands on February 17, 2027.
Limitations
This does not prove that any named hall is over the line, or what share of a cluster a model uses. Nvidia has not published a TPP, so 15,832 is half of a sparsity-footnoted 3,958 rather than a dense number printed beside it, and the B200 and B300 figures divide an eight-GPU total by eight. B300's 3 POPS INT8 cell loses only because its dense half is smaller than the FP4 product. Both the 1,000 and the 1,024 bandwidth conversions leave the under-or-over calls unchanged. Paragraph (b)(3) says "significant volume" and "with 90 days" without repeating 12,000,000, so February 17 is our addition, BIS's $16,720 is a net paperwork estimate, and we did not determine which designers sit on supplement number 6 to part 740.
The Strongest Case Against This Analysis
Grant the objection its full strength. A datasheet peak is not access anyone can use, H100 clusters do not train at 1,979 dense teraFLOPS, and a building with 758 installed H100s is not what section 744.24 describes. What is prohibited is knowledge that the data center will provide more than 12,000,000 aggregate TPP to or for those companies, so renting 757 H100-equivalents to a parent headquartered in China stays under this arithmetic however full the rest of the hall is. Case-by-case review, unlike a presumption of denial on a direct D:5 shipment, means a license can still issue, and treating 758 as the size of a forbidden data center overstates a knowledge test.
Count the access.
That objection is right about the legal object and wrong about the unit. Technical Note 2 takes the highest theoretical rate the manufacturer claims and then drops the sparse figure, and delivered tokens are not in the formula, so an operator who wants to stay under still counts chips of access at 15,832 each, across linked sites. 757 times 15,832 leaves 15,176 of room, and the next chip spends it. Section 744.24's case-by-case standard is real, and this piece does not turn the rental path into a denial. It names the count at which the question starts for an H100 SXM scored the way the note scores it.
The Bottom Line
If you rent GPUs to a company whose parent sits in Macau or Country Group D:5, count dense TPP before the next shipment. On the rates Nvidia prints, 12,000,000 aggregate TPP is 758 H100s or H200s, 334 B200s, or 223 B300s, and 757 of the H100s are still 15,176 under. A U.S. headquarters can clear the five-country purchase license with a monthly report. It cannot, on the text of section 744.24, clear the access test the same way. Review of that license is case by case, comments close January 19, 2027, and a hall already providing this access is on a 90-day clock we add up to February 17.