On November 16, 2026, the Bureau of Industry and Security (BIS) at the US Department of Commerce published a new rule requiring export licenses for advanced semiconductors destined for data centers that provide computing power to companies headquartered in China, Macau and elsewhere (hereafter, "Chinese-headquartered companies"). Under existing provisions, the criteria cited as examples included whether a Chinese-headquartered company was the purchaser, or a major customer using more than 50% of a data center's computing power. The new rule adds an absolute threshold of 12 million TPP (Total Processing Performance, a measure of computing performance used in export controls) and ties the results of customer checks (KYC) at foreign data centers to the exporter's knowledge determination. As a result, even a Chinese-headquartered company renting only part of a large data center may fall under the export license requirement, depending on the scale of its usage. What will determine the rule's effectiveness is how TPP is calculated, and the mechanism under which an exporter is presumed to "know" the end use when customer-check information is unavailable.

AD

Data centers anywhere in the world providing Chinese-headquartered companies with more than 12 million TPP are covered

The newly created §744.24(a)(1) requires a license for the export, reexport and in-country transfer of advanced computing chips (Export Control Classification Number 3A090, excluding 3A090.c), as well as computers containing them (4A090) and similar items. It applies when an exporter knows that a single data center, or "connected data centers," provides or will provide access to computing power exceeding 12 million TPP in aggregate to or for Chinese-headquartered companies. The text specifies the scope as "any destination worldwide." No license exceptions apply, and applications are reviewed case by case.

A Chinese-headquartered company here means a company headquartered in US export control Country Group D:5 or Macau, or one whose ultimate parent is headquartered in those locations. The rule was published as an interim final rule without prior notice and comment; it takes effect on the date of publication in the Federal Register, and comments will be accepted for 60 days after publication. Federal Register publication is scheduled for November 19. If published as planned, the comment deadline would be January 18, 2027. The official dates will be fixed in the Federal Register version.

The measure also newly establishes license requirements for advanced computing items bound for five countries—Indonesia, Malaysia, Singapore, Thailand and Brazil—and formally rescinds the January 2025 AI Diffusion Rule. In addition, "commercial recommender systems," meaning systems whose use is limited to recommending products or content, fall outside the §744.24 license requirement if the provider notifies BIS of the counterparty's name, the aggregate TPP provided, and confirmation that the use is limited to recommendation. However, the rule's preamble explains that, because of the relationship with existing provisions, such providers will also need a license after 90 days. On January 12, the US House passed a bill to bring remote access under export controls, but this rule introduces a license requirement for transfers of advanced chips without waiting for a change in law. Determinations based on headquarters location were clarified in BIS's May 31 guidance (Japanese).

Japan belongs to Country Group A, so the newly created license requirements for the five countries do not apply to it. However, because §744.24 applies regardless of where a data center is located, facilities in Japan could be covered. For example, if an exporter knows that a Japanese data center provides Chinese-headquartered companies with more than 12 million TPP in aggregate computing power, a license is required to export or transfer advanced chips to that facility. The same applies where customer-check information cannot be obtained and the exporter is presumed to "know" the end use. Domestic operators that provide computing power to Chinese-headquartered companies will be asked by exporters to supply information on their customers' ownership and usage.

758 H100s exceed 12 million TPP, as an absolute threshold joins the usage-share test

TPP is calculated by doubling MacTOPS (the number of multiply-accumulate operations per second, in units of one trillion) and multiplying by the bit length of the operation (3A090 Technical Note 2). Because one multiply-accumulate counts as two operations, this is equivalent to multiplying performance expressed in TFLOPS by the bit length. The calculation uses the manufacturer's published theoretical maximum performance, and where multiple bit lengths are supported, the highest TPP is used. Where both sparse and dense matrix operations are supported, the dense matrix performance is used.

According to NVIDIA's published specifications, the H100 SXM delivers 1,979 TFLOPS of FP8 dense matrix performance. Multiplying by 8 bits gives a TPP of 15,832 per unit. For the B300, the HGX B300, which carries eight units, is rated at 108 PFLOPS of FP4 dense matrix performance, or 13.5 PFLOPS per unit. Multiplying by 4 bits gives a TPP of 54,000, which is higher than the FP8-based figure of 36,000, so 54,000 is used for the B300. The B200's TPP, calculated the same way, is 36,000.

These figures show how many GPUs it takes to exceed the 12 million TPP threshold. Based on NVIDIA's published performance, 758 H100s or 223 B300s exceed the threshold. With 757 H100s the total is 11,984,824 TPP, just short of 12 million. Converted to servers with eight GPUs each, the threshold is crossed at the 95th H100 server or the 28th B300 server. For products with different configurations, such as GB200 and GB300, the TPP per unit differs. Similar conversions have also been presented in analyses overseas.

Two definitions bear on how far multiple chips and data centers are aggregated. The newly created §772.1 defines "aggregate TPP" as the sum of the TPP of the covered chips. §744.24(d)(2) defines "connected data centers" as two or more independent data centers that are linked by a network and jointly share computing resources, balance loads and replicate data, including digital connections. As a result, even if facilities are split across multiple sites, they are subject to aggregation if operated as a single whole. The text does not make clear, however, whether this is limited to facilities run by the same operator, or whether the scope of aggregation varies with the distance between facilities or the form of contract.

The existing §744.23(a)(3) has also required licenses for exports "for" Chinese-headquartered companies. In the preamble to this rule, BIS answers questions received about that wording. The examples it gave of where the existing provision applies are where a Chinese-headquartered company is a direct or indirect purchaser, and where it is a major remote user consuming more than 50% of a data center's total computing power. The new rule adds the absolute threshold of 12 million TPP to these examples. The 50% figure is only an example of application and not the sole criterion under the existing provision.

Item Examples given in the preamble for the existing provision (§744.23(a)(3)) New §744.24
Covered parties Chinese-headquartered company is a direct or indirect purchaser Provided to or for Chinese-headquartered companies
Usage share Major remote user consuming more than 50% of the data center's total computing power No share-based criterion
Absolute computing power No example given More than 12 million TPP in aggregate at a single or connected data centers
Customer checks and transition No example given If information cannot be obtained, the end use is presumed "known" (Note 3). Existing provision requires a license within 90 days of the effective date; applications reviewed case by case

Note: The column for the existing provision reflects examples given in the preamble to this rule; the column for the new provision is based on the text of the rule.

A test based on usage share alone means that, for the same amount of computing power used, the larger the overall data center, the smaller the share. An absolute threshold, by contrast, makes the amount of computing power provided to Chinese-headquartered companies the basis for judgment. A customer renting only a portion of a large data center could therefore be covered if its usage reaches the equivalent of 758 H100s.

AD

If customer-check information cannot be obtained, the exporter is presumed to "know" the end use

The new license requirement applies when the exporter "knows" of the covered use. The rule builds the customer-check information from foreign data centers into this determination. According to Note 3 to §744.24(a)(1), an exporter that claims it did not know of the covered use must have obtained, and be able to present, sufficient information on the results of customer checks carried out by the foreign data center. The items to be confirmed include whether the customer is directly or indirectly owned or controlled by Chinese individuals or companies. If the data center does not or cannot provide the necessary information, this is treated as a red flag, and unless it is resolved before the transaction, the exporter is presumed to "know" of the covered use.

In the preamble, BIS explains that it also expects foreign data centers to conduct sufficient checks on their customers. One aim is to prevent circumvention through corporate restructurings and spin-offs. The checks are not limited to direct contract counterparties. According to the preamble, the condition "for Chinese-headquartered companies" is also met where the exporter knows that a customer that is not a Chinese-headquartered company is training or serving AI models for a Chinese-headquartered company.

The rule also addresses changes in use after chips have shipped. Under §744.24(e), exporters and others who ship covered items without a license on or after the Federal Register publication date must notify recipients of three points: that a BIS license is required if the items are used to provide more than 12 million TPP of computing power; that responsibility for obtaining the license lies with the party carrying out the transfer, such as the data center operator; and that the recipient must follow the EAR (Export Administration Regulations) customer-check guidance (Part 732, Supplement No. 3). A change from a use that does not require a license to one that does is treated as an "in-country transfer," extending the rules to holders who were not involved in the original export transaction (Note 1).

A transition period applies to data centers already providing substantial computing power to Chinese-headquartered companies before Federal Register publication. Under §744.24(b)(3), those that wish to continue providing it must obtain a license within 90 days of the rule's effective date. If publication occurs as scheduled on November 19, the deadline would be February 17, 2027. The grace period is limited to data centers that were providing such power before publication and does not apply to new provision or changes in use after publication.

Ambiguity remains over how TPP is aggregated, and whether Note 3 survives in the Federal Register version

The biggest uncertainty in determining the rule's scope is the unit by which the 12 million TPP is calculated. §744.24(a)(1) describes the covered companies in the plural, but does not say whether TPP is calculated per customer or aggregated across all provision to Chinese-headquartered companies. The definition in §772.1 likewise does not specify which customers are included in the total. If calculated per customer, contracts split into 757 H100-equivalents or fewer would each fall below the threshold on their own. If aggregated across all Chinese-headquartered customers, the total provided to multiple customers could exceed the threshold even if each one's usage is small.

For license applications, too, case-by-case review has been adopted rather than a presumption of denial. In May, before the rule was published, Noah Tan of the Carnegie Endowment analyzed a proposed restriction on third-country cloud services, citing as challenges evasion methods that distribute training across multiple clusters, opaque corporate ownership, and BIS's limited enforcement capacity.

Some provisions, however, would not change under either interpretation of the TPP aggregation unit. §744.24(a)(1) applies regardless of a data center's location, and Note 3 provides that if a foreign data center does not supply customer-check information, the exporter is presumed to know of the covered use. In other words, the mechanism under which an exporter asks the destination data center to confirm customer information, and the data center investigates its customers' ownership, works independently of how TPP is aggregated. Its scope covers data centers that will receive EAR-controlled advanced chips going forward, as well as those that change the use of existing equipment.

This outlook could change in two main ways. One is if the presumption in Note 3 is removed or relaxed in the Federal Register version scheduled for November 19, or in FAQs or guidance BIS issues afterward. If exporters could claim they did not know of the covered use without obtaining customer-check information from foreign data centers, the reach of checks extending to users who rent only part of a large facility would narrow.

The other is if, after the comment period ends or ahead of the deadline for existing operators to obtain licenses, revisions exclude cloud services shared by multiple customers from §744.24 or create a new license exception. If publication occurs as scheduled, the comment deadline would be January 18, 2027, and the license deadline for existing operators February 17 of the same year. Any such changes can be confirmed in the Federal Register.

What matters first is how the Federal Register version, scheduled for November 19, describes the scope of TPP aggregation and the presumption in Note 3. These provisions will determine how far the new export control requirements can be applied to Chinese-headquartered companies that use part of a large data center.