ai insights

Commerce Rule Sets a 12 Million TPP License Trigger for Overseas AI Data Centers

BIS will require a license for chips bound for data centers giving China-headquartered firms over 12 million aggregate TPP. Who must act and when.

The Commerce Department’s Bureau of Industry and Security (BIS) released an interim final rule on November 16, “Addressing Remote Access to Advanced Semiconductors,” that makes the customer list of an overseas data center a licensing question for whoever ships chips to it. Under a new § 744.24 of the Export Administration Regulations (EAR), a license is required to export, reexport or transfer advanced computing items when the shipper knows, or is aware of a high probability, that they are destined for a data center, anywhere in the world, that provides or will provide access to more than 12,000,000 “aggregate TPP” to companies headquartered in, or ultimately parented in, Macau or a Country Group D:5 destination. China is in D:5. The rule also adds a license requirement for the same items going to Indonesia, Malaysia, Singapore, Thailand and Brazil, and it formally rescinds the AI Diffusion Rule. Rule text, FR Doc. 2026-23529 · BIS press release

Nothing is in force yet. The public-inspection copy schedules Federal Register publication for November 19, and the rule takes effect on publication. This article works from that copy, BIS’s press release and earlier guidance and, where noted, a Nixon Peabody client alert. S5 Labs is not a law firm and this is not legal advice. If you ship, host or rent this hardware outside the United States, read the published rule and talk to export counsel.

What takes effect and when

This is an interim final rule, so there was no proposed-rule stage. Comments are open for 60 days after publication at regulations.gov, under ID BIS-2026-0628. BIS’s press release says the rule was written “by and at the direction of the Department of Treasury and the National Security Council,” a role the rule text does not mention. Rule text · BIS press release

MilestoneWhat the rule saysDate if published November 19
Effective dateDate of Federal Register publicationNovember 19, 2026
Savings clause for shipments already aboard a carrierComplete within 30 days of publicationDecember 19, 2026
Comment deadline60 days after publicationJanuary 18, 2027
License for existing arrangementsWithin 90 days of the effective dateFebruary 17, 2027

The right-hand column is our arithmetic. The rule still carries bracketed placeholders, and January 18, 2027 is a federal holiday, so use the dates printed in the published version. The savings clause covers only shipments that lost a license exception or no-license-required status because of this rule and were en route on the publication date.

The 12,000,000 TPP trigger

Section 744.24(a)(1) bars exporting, reexporting or transferring the listed items “without a license if you have ‘knowledge’” that the item is destined to a data center “that is providing or will provide access to more than 12,000,000 ‘aggregate TPP’ either at one data center or through ‘linked data centers’” to or for covered companies. The listed items are those classified under ECCN 3A090 (except 3A090.c, high-bandwidth memory), 4A090 and the related “.z” entries. Rule text, § 744.24

This is an end-use control on the shipper. The duty sits with the exporter, reexporter or transferor who has knowledge; the rule does not ban remote access and does not regulate cloud customers directly. The control also follows the hardware. Note 1 applies the prohibition whether or not the person holding the item was a party to the original export, and the preamble treats a change from a permitted to a prohibited end use as an in-country transfer.

Splitting capacity across linked sites does not help. Linked data centers are those whose networking lets facilities “share resources, balance workloads, and replicate data as a unified whole,” and the definition “includes digitally-linked infrastructure, not just physically-linked infrastructure.” Intermediaries count too: the preamble says the control applies where a data center serves a non-D:5 customer that is training models for, or providing models to, a D:5 company.

No license exception overcomes the prohibition unless BIS says so, and applications are reviewed case by case.

The text leaves three things open. It says “companies” in the plural without saying whether the 12,000,000 is summed across all covered customers or measured per customer. We think a sum is the natural reading. It also gives no measurement period and does not say whether the count is installed capacity or capacity actually allocated to covered customers.

What 12,000,000 TPP is in chips

The rule contains no GPU count and names no chip or vendor. Aggregate TPP is the sum of the total processing performance of every covered chip in a given group. The technical notes to ECCN 3A090 define TPP as 2 × MacTOPS × the bit length of the operation, using the peak dense-matrix rate the manufacturer claims. One of the 3A090.a thresholds is a TPP of 4,800, so the trigger equals 2,500 chips sitting exactly at that threshold. Other prongs of 3A090 reach chips with a TPP as low as 1,600, so it is not a floor. Rule text, § 772.1 and ECCN 3A090

Data-center GPUs score well above 4,800. The analyst newsletter 456X Research, in a partly paywalled post, puts 12,000,000 TPP at roughly 760 Nvidia H100 GPUs, using about 15,800 TPP per H100. That is an analyst estimate, and faster chips would reach the trigger with proportionally fewer units. 456X Research

Who counts as a covered customer

The test turns on headquarters and ultimate parent, wherever the customer operates: a company “headquartered in, or whose ultimate parent company is headquartered in, Macau or a destination specified in Country Group D:5.” The word China does not appear in § 744.24. An eCFR snapshot dated November 10, which predates the rule, lists China in D:5 alongside Russia, Iran, North Korea, Belarus, Cuba, Venezuela and others. Macau is named separately because it is not in D:5. Hong Kong has no row of its own in that table, and the rule never mentions it. eCFR, Supplement No. 1 to Part 740

Staying under 12,000,000 TPP does not put a deployment in the clear. BIS calls the new control a “logical extension of existing controls,” and the older one in § 744.23 remains. BIS guidance from May 31, 2026 says a license is already required for advanced computing items for entities headquartered in D:5 or Macau “even if the entities themselves are located outside” those places. The new preamble says that control reaches cases where a covered entity was the purchaser, directly or indirectly, or was “the exclusive or primary (greater than 50 percent of total compute) remote end user” of the capacity. BIS guidance, May 31, 2026 · Rule text, preamble III.B

Five countries get a destination license requirement

Separately from the end-use trigger, the rule amends § 742.6 to require a license to send the covered items “to or within” Indonesia, Malaysia, Singapore, Brazil and Thailand. There is no TPP floor beyond the classification itself. The rule’s summary says these countries present a “higher-than-normal risk of diversion,” and Commerce can add countries. Applications are reviewed case by case, with weight on the destination’s own export controls and its coordination with BIS on enforcement. Rule text, § 742.6

The requirement does not apply when the ultimate consignee or end user is headquartered in, or ultimately parented in, the United States, “provided the exporter, reexporter, or transferor follows the reporting requirements in § 743.10.” Those reports are monthly and due by the first business day of the month after the transaction. Each one names every party and its ultimate parent, the destination and end users, and each product’s ECCN, SKU, TPP, quantity and date of export. They go to the same BIS web address as the recommender carve-out submissions described below, a link that, according to Nixon Peabody’s alert, was not yet functional when the alert was published. Rule text, § 743.10 · Nixon Peabody

Singapore’s trade and digital ministries said in a joint statement, reported by Al Jazeera, that they are studying the measure and engaging U.S. authorities on its scope and implementation. We found no on-record reaction from the other four governments. Al Jazeera

Customer checks, existing arrangements and the recommender carve-out

Know your customer. To show it lacked knowledge, a shipper is “expected to obtain and be able to provide sufficient information from the foreign data center” about that data center’s customer due diligence, including whether any customer is directly or indirectly owned or controlled by covered persons. A data center that declines or is unable to provide it creates “a red flag that presumptively provides you with ‘knowledge’” unless the shipper resolves the flag before proceeding. Rule text, § 744.24(a)(1), Note 3

Changes after delivery. For shipments made on or after publication without a license, the shipper must inform the recipient that BIS authorization is required if the hardware will be used above the threshold for covered companies, and that the transferor, “e.g., the data center operator,” must obtain it. The rule says “inform”; Nixon Peabody’s alert reads that as written notice. Rule text, § 744.24(e) · Nixon Peabody

Existing arrangements. Any data center that was providing a “significant volume of compute power” to covered companies before publication and wants to continue “must obtain a license with 90 days of the effective date of this IFR or face future export restrictions.” (The “with” is the rule’s typo.) “Significant volume” is not defined and is not equated with 12,000,000 TPP. Rule text, § 744.24(b)(3)

Recommender systems. The prohibition does not apply when access is “only being provided” for a commercial recommender system and the party submits the covered company’s name, the TPP provided and its due diligence to BIS. The definition is narrow: a tool “exclusively designed and exclusively used to rank, filter, or suggest commercial products, services, or content.” Any system that generates text, images, audio, video or software is excluded “regardless of whether the system also performs recommendation functions.” The preamble and the regulatory text do not line up on duration. Discussing the older § 744.23 control, the preamble says operators serving recommender workloads “may continue providing these services for 90 days, but must obtain a license” afterward, while § 744.24(b)(2) has no sunset. Get an answer from BIS or counsel before relying on the carve-out. Rule text, § 744.24(b)(2), (d)(1) and preamble III.B

The AI Diffusion Rule formally ends on publication

The rule “formally rescinds the regulations implementing the Biden-era AI Diffusion Rule,” which it says “has never been enforced” and whose rescission BIS announced in May 2025. That removes the country tiers, the Data Center Validated End User program and License Exceptions AIA and LPP. License Exception ACM stays. It also removes ECCN 4E091, the control on model weights. BIS writes that “a more effective approach … is through the use of specified destination-based or end-use license requirements.” Rule text, section IV · AI Diffusion Rule, 90 FR 4544

Export controls still reached models this year. In June, controls requiring nationality verification took Claude Fable 5 offline for about 18 days.

Foundries and packaging houses get their own change: a new § 744.23(a)(3)(i)(C) requires a license when a front-end fabricator or packaging and test company ships advanced-node chips specified in 3A090.a that were designed by a company that is neither an approved nor an authorized IC designer. Our TSMC coverage has background on foundry capacity. Rule text, preamble III.B

Nvidia’s governance week

The rule arrived days after a leadership change at Nvidia, whose reported financing backstop for an OpenAI data-center campus we covered in August. The board accepted Jensen Huang’s resignation as President, CEO and director on November 11 and announced it on November 12, with CFO Colette Kress as interim CEO. Neither the release nor the 8-K gives a reason. On November 8 the board had said it was starting an independent investigation in light of recent reporting that Huang had been the target of a foreign intelligence effort, and that it had no reason to believe national security or Nvidia secrets were compromised. Press accounts of the circumstances rest on unnamed sources, and NPR says it has not independently verified them. Nvidia release on SEC EDGAR · Form 8-K · Board statement, November 8 · NPR

On November 17 Nvidia announced a standing board committee to oversee national security policy, export control compliance and global trade risk management, with the company’s National Security and Global Trade Compliance teams reporting directly to it and to Kress. The release mentions neither the BIS rule nor Huang. Nvidia newsroom

No primary source connects the rule to Huang’s exit. CBS notes the timing, and TechCrunch states, without naming a source, that the rule was a direct response to a chip smuggling scandal at Nvidia. BIS’s release and the rule say nothing about it. CBS News · TechCrunch

CNBC, citing two unnamed people, reports that Nvidia told buyers that for now it would not sell chips to customers who rent more than 50% of their GPUs to Chinese companies. That report concerns a bill the House passed in January, not this rule, and CNBC says Nvidia and BIS did not respond to a request for comment. CNBC

What to do now

Operators outside the United States. Total the TPP of covered hardware you make available to customers headquartered or parented in Macau or a D:5 country, across every site that shares workloads, and trace ownership to the ultimate parent. If you already serve those customers, plan around the 90-day license window; “significant volume” is undefined, so ask counsel before assuming you are below it.

Exporters, resellers and server builders. Add the data center’s customer due diligence to your order paperwork now, because under Note 3 a refusal is a red flag that presumptively gives you knowledge. For sales into the five listed countries, decide whether the U.S.-headquartered exclusion applies and who files the monthly report.

Teams renting GPU capacity in Southeast Asia or Brazil. Expect ownership questionnaires from your provider, and ask whether its hardware supply depends on a license that has not been granted. If your capacity runs on custom accelerators such as Google’s TPUs, ask the vendor for the ECCN and TPP figures. The test is performance, and the rule names no vendor.

Anyone affected. Use the comment period to press BIS on the open questions: how the 12,000,000 is aggregated, what “significant volume” means, and whether the recommender carve-out expires. Until BIS answers, plan around the stricter reading of each.

Sources

The rule and BIS documents

Nvidia

Analysis and coverage

Continue reading.

Insight7 min read

Nvidia's $250 Billion Backstop Says More About OpenAI's Credit Than Its Compute

Nvidia may guarantee $250B so OpenAI can lease a 10GW Ohio campus. Its own credit default swaps hit a record the same day. What that signals.

Insight13 min read

NVIDIA RTX Spark and Nemotron 3.5 Bolt: A 74B Model in 24GB, Before Anyone Can Test It

NVIDIA's Bolt packs a 74B-A9B MoE into 24GB with 2-bit QAT; RTX Spark offers up to 128GB. What fits, what's unpublished, what to test first.

Insight4 min read

NVIDIA Agrees to Buy Hugging Face. Open-Model Builders Should Read the Commitments.

NVIDIA's $12.93B Hugging Face agreement promises hardware choice. What the filing says, what remains unproven, and how builders can preserve portability.