Export controls have always followed the box. The rule the Commerce Department released on November 16 follows the login.
For four years, Washington’s AI chip restrictions worked on shipments: which accelerators could go to which countries and which buyers. Chinese companies have been barred from buying top Nvidia parts directly since 2022, as NBC News recounts. They could still rent the same parts from a data center in Johor or Singapore and train over the network. The new interim final rule, “Addressing Remote Access to Advanced Semiconductors” (RIN 0694-AK50), goes after that rental path. Read the operative text and the mechanism is not a chip ban at all. It is a customer screening mandate that lands on every overseas data center that wants to keep receiving U.S. silicon.
That matters well beyond China. If your company rents GPU capacity in Southeast Asia or Brazil, your provider is about to start asking who owns you.
What the rule switches on, and when
The Federal Register public inspection copy shows the rule is scheduled for publication on November 19, 2026, and it takes effect on publication. Comments are due 60 days later, which puts the deadline around January 18, 2027 (our arithmetic from the publication date). The rule text BIS posted does four things:
- A new worldwide data center end use control, § 744.24. You may not export, reexport or transfer covered advanced chips and servers without a license if you know the item is going to a data center, anywhere, that “is providing or will provide access to more than 12,000,000 ‘aggregate TPP'” to companies headquartered in Macau or a Country Group D:5 destination, or whose ultimate parent is. TechCrunch puts that threshold at roughly the compute of about 750 Nvidia H100 or H200 chips.
- A license requirement for five countries. Shipping advanced chips to Indonesia, Malaysia, Singapore, Thailand or Brazil now needs a license, reviewed case by case. The rule calls them “countries which present a higher-than-normal risk of diversion.”
- Tighter presumptions for foundries and chip packagers making advanced logic chips for designers in licensed destinations.
- Formal rescission of the Biden era AI Diffusion Rule, which the rule notes “has never been enforced.”
The BIS press release says the rule was “written by and at the direction of the Department of Treasury and the National Security Council.” That is unusual enough that Chris McGuire, a former Biden AI adviser now at the Council on Foreign Relations, wrote, “I have never seen a BIS rule that was not written by BIS,” as The Next Web reported. He still called it “a major step forward,” per NBC.
The timing is not subtle either. The rule lands days after Jensen Huang resigned as Nvidia CEO amid reporting that linked him to a former executive of Megaspeed, a Singapore company. NBC, citing the New York Times and Financial Times, reports Megaspeed planned to rent computing power to Alibaba and ByteDance; The Next Web notes Megaspeed has denied wrongdoing. Markets read the rule as material: Nvidia fell as much as 2.9% and the Philadelphia Semiconductor Index as much as 5.5% on the day, according to Bloomberg figures cited by The Next Web.
Three details that make the threshold bigger than it looks
The 12 million TPP number is the headline. Three pieces of drafting decide how often it actually bites.
It aggregates across “linked data centers.” The count runs “either at one data center or through ‘linked data centers’,” and the definition is broad: data centers joined by networking infrastructure that lets them “share resources, balance workloads, and replicate data as a unified whole,” adding that this “includes digitally-linked infrastructure, not just physically-linked infrastructure.” A regional provider with three modest halls in Singapore, Johor and Batam, tied together as one availability region, should expect BIS to read them as one pool. Spreading a customer’s capacity across sites is not a way under the line.
“For” reaches the customer’s customer. The preamble says the phrase covers cases where there is knowledge that a data center “is providing remote access to a non-D:5 customer that is in turn training SI models for, or participating in the training of SI models for, or providing SI models to a D:5 company.” A Singapore-registered startup training models on contract for a Shenzhen parent’s affiliate is inside the rule, even though its own incorporation papers are clean.
The old test was majority use; the new test is volume. The preamble restates how BIS already read the existing control: it covered a data center if a D:5 headquartered entity “was the exclusive or primary (greater than 50 percent of total compute) remote end user of the data center computing capacity.” Under § 744.24, a Chinese-parented tenant using 5% of a very large campus can trip the license requirement if that 5% exceeds 12 million TPP. That is the loophole being closed, and it is why large neutral colocation sites, not just dedicated China-facing ones, are now in scope.
One word on vocabulary: the rule uses “Super Intelligence” or “SI” wherever older rules said “artificial intelligence,” citing Executive Order 14434. Same chips, same models, new label.
The sentence that moves the work onto data centers
The most consequential language sits in a note, not the headline control. Note 3 to § 744.24(a)(1) says that to show you lack “knowledge,” exporters “will be expected to obtain and be able to provide sufficient information from the foreign data center regarding that data center’s ‘know your customer’ due diligence checks.” Those checks must cover whether customers are “subject to direct or indirect ownership or control” by persons in D:5 or Macau. Then:
“A foreign data center declining or being unable to provide information sufficient to verify compliance with this end use control shall be considered a red flag that presumptively provides you with ‘knowledge’ unless you are able to resolve that red flag before proceeding with the transaction.”
Read that from Nvidia’s or Dell’s or Supermicro’s side. A data center that cannot produce its customer ownership diligence is, by default, a data center you knowingly ship to in violation. No compliance team will accept that exposure, so the practical rule is simple: no KYC file, no next shipment. The burden of proof has moved from the government to the operator’s customer database.
Two more provisions keep that obligation alive after delivery. Note 1 says the prohibition applies “regardless of whether the person in possession of the EAR item was a party to the transaction” that originally shipped it. And new paragraph (e) treats a change from a permitted use to a prohibited one as a “transfer (in-country)” that needs authorization. A cluster leased to a domestic bank in 2025 and re-let to a D:5 parented tenant in 2027 is a new licensable event, even though the GPUs never moved.
Note 2 closes the integration route: the prohibition applies even when the U.S. chip is “incorporated or commingled into or with a foreign-made item,” and de minimis relief is unavailable where there is knowledge of the prohibited end use.
Who carries what
| Party | What the rule asks of it | Clock |
|---|---|---|
| Chip, server and system exporters | Obtain and be able to produce the destination data center’s KYC diligence; license where knowledge exists; monthly reports for U.S. headquartered exclusions in the five countries | Effective on publication, scheduled Nov 19 |
| Overseas data centers already serving D:5 or Macau headquartered customers above the threshold | Obtain a license to continue | Within 90 days of the effective date, which our arithmetic puts at about February 17, 2027 |
| Data centers in Indonesia, Malaysia, Singapore, Thailand, Brazil | Every new advanced chip shipment needs a license unless the consignee or end user is U.S. headquartered and the exporter reports monthly | Effective on publication |
| Operators serving recommender workloads only | Submit the customer name, total TPP provided and the diligence performed to BIS to use the carve-out | Effective on publication; existing services get 90 days |
| Shipments already in transit | Can proceed if completed within 30 days of publication, per Nixon Peabody’s alert | 30 days |
The five-country exclusion is narrower than it first sounds. It applies to consignees or end users “headquartered in, or whose ultimate parent company is headquartered in, the United States,” and the new § 743.10 report must list every party and its parent companies with addresses, the ECCN at paragraph level, product name and SKU, the TPP of each product, quantity and date, filed “no later than the first business day of the month following the transaction.” The exclusion names one country. A Japanese, Korean, European or Singaporean headquartered operator building in Johor gets case-by-case licensing like everyone else.
What changes for an enterprise renting GPUs in the region
Most readers here are not exporters and have no Chinese parent. The rule still reaches them, through their providers.
Expect an ownership questionnaire with your next order
Note 3 requires data centers to run diligence on direct and indirect ownership and control. The cheapest way for an operator to produce an auditable file is to ask every customer, not just suspicious ones. Enterprise buyers with capacity in the five countries should expect beneficial ownership forms, parent company declarations and renewal clauses that let the provider terminate if your ownership changes. Companies with any D:5 linked shareholders, joint venture partners or offshore holding structures should work out their answer before the provider asks, because “declining or being unable to provide information” is exactly the red flag the rule names.
Capacity from U.S. providers and local neoclouds will diverge
Under the five-country rule, AWS, Microsoft, Google and Oracle facilities in Malaysia or Singapore should be able to keep receiving chips under the U.S. headquartered exclusion, with their suppliers filing monthly. A Malaysian, Singaporean or Indonesian neocloud, or a non-U.S. telecom building sovereign GPU capacity in those markets, waits on a case-by-case license. Joshua Kurlantzick of the Council on Foreign Relations told Fortune, as quoted by The Next Web, “For Malaysia, the controls could be disastrous.” For buyers, the near-term effect is predictable: new local capacity gets slower and more expensive to stand up, and the regional price gap between hyperscaler and local GPU rental could narrow or flip. If you were planning to move a training or inference workload to a regional neocloud for data residency or price, ask now whether its 2027 expansion depends on licenses it does not yet hold. The same question about counterparty fragility came up in our look at what Nvidia’s leadership change means for neocloud buyers; this rule adds a regulatory version of it.
“Recommender system” will not cover your chatbot
The carve-out is real but tight. A “commercial recommender system” means a tool “exclusively designed and exclusively used to rank, filter, or suggest commercial products, services, or content to users based on user behavior, transaction history, or item attributes.” It excludes any system “designed or used to generate novel text, image, audio, video, software, or other content in response to user inputs, or any system that incorporates such model, regardless of whether the system also performs recommendation functions.” A shopping assistant that writes product summaries is out. A retrieval system with a generative answer layer is out. Only classic ranking survives, and using it requires a filing with BIS naming the D:5 company, the total TPP provided and the diligence performed.
Your provider’s tenancy decisions are now partly yours
Because a change of use counts as a transfer, providers will start writing permitted use clauses into GPU contracts and policing them. Expect acceptable use terms that forbid resale or sub-tenancy to unscreened parties, and audit rights over who is actually using reserved capacity. If you resell compute internally across a group with foreign affiliates, or run a platform where third parties fine-tune models on your reserved cluster, map that before your provider asks. The rule’s “for” language means a downstream customer training for a D:5 company is your provider’s problem, and therefore yours.
What is still unsettled
Several things will move between now and February.
- The comment period could change the text. This is an interim final rule, effective immediately but open for comment until roughly January 18. The threshold, the linked data center definition and the reporting fields are all fair game.
- Enforcement posture is unknown. The rule adds an “is informed” process under § 744.24(a)(2) that lets BIS impose license requirements on specific parties by letter, with oral notice confirmed in writing within two working days. Nixon Peabody flags these letters as one thing to watch. So is the list of additional countries the Secretary of Commerce can designate.
- Gaps remain. McGuire, while praising the rule, said loopholes remain, including some covering activity inside the U.S., per The Next Web. The rule’s own diligence standard also leans on foreign operators’ records that BIS cannot directly audit.
- Politics may shape licensing speed. The Next Web reports Trump meets Xi in Shenzhen this week, and NBC notes that Treasury Secretary Scott Bessent had publicly criticized Commerce for its reticence the week before. Case-by-case review means approval timelines will track policy temperature, not a published service level.
The short version for anyone buying compute outside the U.S.: the chip export question has become a customer identity question, and your provider now has a legal reason to ask it. Have your ownership answer ready, read your next GPU contract for permitted use and termination clauses, and if a regional provider’s growth plan rests on shipments to Malaysia, Singapore, Indonesia, Thailand or Brazil, price in the license queue. For a related look at why where you run models is now as much a governance call as a cost call, see our breakdown of the Cloudflare cache leak and open model hosting, and for the arithmetic of owning capacity outright, Revolut’s self-hosted coding model case.
