BIS Requires Licenses for AI Chips Shipped to Five Countries

A BIS interim final rule requires licenses for advanced AI chips going to Singapore, Malaysia, Indonesia, Thailand and Brazil, and for any data center giving China- or Macau-parented companies more than 12 million TPP of compute. US-parented customers are excluded only if the shipper files monthly reports.

By Rajesh Beri·November 16, 2026·10 min read
Share:
Rows of GPU server racks in a humid Southeast Asian colocation hall, one rack door open with a clipboard of blank customer ownership forms hanging from it.

Illustration generated using AI

If you rent or plan to rent GPU capacity in Singapore, Malaysia, Indonesia, Thailand or Brazil, every new advanced chip shipped to your provider will need a US export license once the rule takes effect. The exception is chips serving a US-headquartered customer, and only if the shipper files monthly reports naming that customer's ultimate parent. That is the core of the interim final rule the Commerce Department posted on November 16. A second provision reaches further: any data center, anywhere, that gives companies headquartered in China, Macau or another Country Group D:5 destination access to more than 12 million "aggregate TPP" of compute needs a license too. Expect ownership questionnaires, reporting clauses and slower capacity additions in those five markets, starting with your next contract.

The posted text still has placeholders where the Federal Register date belongs. The rule takes effect on publication, and comments are due 60 days after that. So the clocks have not started yet, and that gap is your window to get ahead of what your providers will ask for.

What the Remote Access Rule Actually Requires

The rule controls remote access by attaching conditions to the chip shipment. BIS cannot directly license a cloud login: under current law BIS does not treat cloud computing services as exports, which is why the House passed the Remote Access Security Act 369-22 in January 2026 (it has not cleared the Senate). This rule gets there another way. It prohibits exporting, reexporting or transferring advanced chips when the sender has "knowledge" they are headed to a data center serving the restricted companies, and it treats a later switch to that use as a new in-country transfer that needs authorization.

The BIS rule text does four things:

  1. A data-center end-use control (new § 744.24). A license is needed when chips go to a data center "to or within any destination worldwide" that provides or will provide more than 12,000,000 aggregate TPP, at one site or across "linked data centers," to companies headquartered in Macau or D:5, or whose ultimate parent is. "Linked" explicitly includes digitally linked infrastructure, so splitting a cluster across two buildings does not reset the count.
  2. A destination license for five countries (§ 742.6). Exports, reexports and in-country transfers of advanced computing items to or within Indonesia, Malaysia, Singapore, Brazil and Thailand now need a license, reviewed case by case on how well each country enforces export controls.
  3. Tighter presumptions for foundries and packagers. Chip designers in licensed destinations get fewer ways around the presumption that their advanced logic chips are controlled.
  4. Formal rescission of the AI Diffusion Rule. The country tiers, License Exception AIA, the Data Center Validated End User program and the model-weights control (ECCN 4E091) are all removed.

The press release says the aim is that controls "cannot be circumvented through remote computing arrangements involving data centers in third countries."

How Big Is 12 Million TPP?

Twelve million TPP is a mid-sized training cluster, roughly 760 H100 or H200 GPUs. TPP (total processing performance) is the export-control compute metric: twice the chip's peak multiply-accumulate rate times the bit length of the operation, using dense figures. GPU Smith's worked derivation from Nvidia's datasheets puts an H100 SXM or H200 at 15,832 TPP, a B200 at about 36,000 and a B300 at about 54,000. On those derived figures, 12 million TPP works out to roughly 758 H100s, 333 B200s or 222 B300s.

That is small enough to catch a single tenant's dedicated training pod at a regional neocloud. It is also aggregate across the restricted companies a data center serves, so a provider with several smaller China-parented tenants can cross the line without any one of them being large.

Who Is Exempt, and What the Exemption Costs

US-headquartered customers are exempt from the five-country license, but the exemption comes with paperwork that names you. Under new § 743.10, the exporter, reexporter or transferor must file a report for each qualifying shipment listing every party, including "the ultimate parent(s) of the ultimate consignee and end user," with names, addresses and phone numbers, plus each product's ECCN, SKU, TPP and quantity. The rule calls this a "Monthly reporting requirement"; reports are due by the first business day of the month after the transaction.

If you are a US company buying a dedicated cluster in Singapore through a neocloud, your provider's chip supplier needs your corporate parent details to use the exclusion. If you are headquartered in Europe, Japan or India, there is no exclusion at all: the chips going into your capacity in those five countries move under a license.

Commercial recommender systems get a carve-out from the data-center control. The rule defines them narrowly as a tool "exclusively designed and exclusively used to rank, filter, or suggest" products or content, and anything that generates text, images, code or other content does not qualify, "regardless of whether the system also performs recommendation functions." The text is inconsistent on timing: the preamble says recommender operators may continue for 90 days and then need a license, while § 744.24(b)(2) says the prohibition does not apply if the operator submits the D:5 company name, total TPP and its diligence. Raise it in a comment letter, and do not plan around the more generous reading.


The Red Flag That Moves Diligence Onto Your Provider

The clause that will change your contracts is Note 3 to § 744.24(a)(1). To show it lacked "knowledge," a supplier "will be expected to obtain and be able to provide sufficient information from the foreign data center" about its know-your-customer checks, including whether any customer is directly or indirectly owned or controlled from D:5 or Macau. A data center "declining or being unable to provide information sufficient to verify compliance" is a red flag that "presumptively provides you with 'knowledge'" unless resolved before the transaction.

Chip suppliers will push that burden to the data centers, and data centers will push it to you. Expect three new asks in neocloud and colocation paperwork:

  • Ownership attestations covering your ultimate parent and any D:5 or Macau shareholders, direct or indirect.
  • Use attestations, because the control also reaches a non-D:5 customer that is "training SI models for, or participating in the training of SI models for, or providing SI models to a D:5 company." A joint venture or model-supply partner with a China-parented company can bring your tenancy into scope.
  • Change-of-use notice, because § 744.24(e) makes a shift from permitted to prohibited use an in-country transfer requiring BIS authorization. Your provider will want a contractual duty to tell it before your workload or customer mix changes.

The rule also says plainly that corporate restructurings and spinoffs are not a way out: diligence "will ensure that companies cannot circumvent new controls through corporate restructurings or spinoffs."

Why This Gap Mattered, and Who Doubts the Fix

The rule closes a channel BIS spent most of 2026 trying to shut. After the Diffusion Rule was dropped in May 2025, Reuters reported via The Star that a chip industry source estimated hundreds of thousands of advanced chips may have reached overseas subsidiaries of Chinese firms. BIS answered with May 31, 2026 guidance confirming that the November 2023 ultimate-parent license requirement still applied, while letting bona fide data centers keep running chips already installed "until further notice." Asia Times reported that Megaspeed's Malaysian subsidiary had bought nearly US$2 billion in Nvidia chips and that Bridge Data Centers removed Megaspeed from its Malaysian hub in April 2026.

The strongest case that this is manageable comes from BIS's own burden estimate. The rule projects about 100 additional license applications a year (an estimated $1,900 in added cost) and a net increase of roughly 970 submissions in its license-exception reporting collection (an estimated $16,720). Burden estimates like these are BIS's own projections, not measurements. If they are right, most enterprise capacity in the five countries keeps flowing under the US-parent exclusion with a monthly filing.

The skeptics point at verification. A Liaoning-based columnist quoted by Asia Times wrote that "Proving that an overseas chip buyer is ultimately Chinese-owned is far harder than it sounds." That cuts both ways for you: the rule answers it by putting a presumption of knowledge on anyone who cannot verify, which is why your provider will ask you for more than it used to. Trade coverage also noted that the rule came with a sharp sell-off in semiconductor and AI infrastructure stocks, and unite.ai's summary flags the same placeholder dates noted above.

The last time an export rule hit AI vendors, enterprises learned the vendor-side lesson in a day: our coverage of the Fable 5 export-control shutdown showed how fast access can vanish. This rule moves slower and lands on infrastructure, where lead times are already long. As our piece on Microsoft's installed-but-unpowered GPUs argued, delivery dates are the term that matters, and a license queue is one more thing in front of delivery.

What to Do Before the Clock Starts

The 90-day window in the rule belongs to data centers already serving D:5 companies, and it runs from the effective date. Your own window is shorter in practice, because providers will rewrite paperwork before they file anything.

This Week:

  1. List every GPU, colocation and managed inference contract with capacity physically in Singapore, Malaysia, Indonesia, Thailand or Brazil, including capacity your hyperscaler or a provider such as Nebius, Lambda or Together AI places there for you. Ask each provider in writing where your cluster sits today.
  2. Get your trade-compliance lead and your corporate secretary to write down your ultimate parent, its headquarters and any D:5 or Macau shareholders. That one page is what every provider will ask for.

This Month:

  1. Map every partner that trains on your clusters or receives models you trained there. Flag any JV, reseller or model customer headquartered in or parented from China or Macau, since the control reaches compute used to train or supply models for them.
  2. Ask each provider in those five countries three questions: Will new capacity for us ship under the US-parent exclusion or a license? Who files the monthly § 743.10 report? What happens to our delivery date if a license is needed?
  3. If you are not US-headquartered, assume every new chip into those markets for your workloads needs a license, and price a fallback region now. Our GPU cloud comparison is a starting list.

Before Renewal:

  1. Add a regulatory-delay clause: if an export license stalls committed capacity, you get relief from minimum commits or a substitute region at the same price. The same logic as our Equinix commitment-term piece applies: cap the commit you make against a date the provider cannot control.
  2. Agree the change-of-use notice language yourself, before the provider's template arrives, so a new customer or workload does not become a breach by surprise.
  3. If you run regulated workloads in Singapore, fold this into the third-party AI inventory that MAS's AI risk guidelines already require.

The Bottom Line

This rule extends export control from the chip shipment to whoever ends up using the compute, and it does so by making every party in the chain prove who its customers are. For a US enterprise with clean ownership, the cost is a monthly report and a longer questionnaire. For a non-US multinational, or anyone with a China-parented partner on shared infrastructure, it is a license queue in five of the fastest-growing capacity markets in Asia and Latin America. Comments run 60 days from Federal Register publication, and the recommender-system timing conflict is the kind of ambiguity a comment can fix. Send your provider list and your ownership page to trade compliance this week.

Continue Reading

Share:

Frequently Asked Questions

What does the BIS remote access rule for AI chips require?

It requires a license to send advanced AI chips to any data center worldwide that gives companies headquartered or ultimately parented in China, Macau or other Country Group D:5 destinations more than 12 million aggregate TPP of compute. It also requires a license for advanced chips going to or within Indonesia, Malaysia, Singapore, Brazil and Thailand, and it formally rescinds the AI Diffusion Rule.

How many GPUs equal 12 million TPP?

Using TPP values derived from Nvidia datasheets, an H100 or H200 is about 15,832 TPP, so 12 million TPP is roughly 758 H100s. A B200 at about 36,000 TPP puts the line near 333 GPUs, and a B300 at about 54,000 TPP near 222. The threshold is aggregate across one site or linked data centers.

Are US companies exempt from the five-country license requirement?

Shipments to an end user headquartered in the US, or whose ultimate parent is, are excluded, but only if the exporter files monthly reports under new section 743.10. Each report lists every party and the ultimate parent, plus each product's ECCN, SKU, TPP and quantity, and is due by the first business day of the following month.

When does the BIS remote access rule take effect?

The rule takes effect on its Federal Register publication date, which was still a placeholder in the text BIS posted on November 16, 2026. Comments are due 60 days after publication, and data centers already serving D:5 companies have 90 days from the effective date to obtain a license.

Newsletter

More on AI Infrastructure, every Sunday

The week's enterprise AI news, AI Infrastructure included, in one Sunday email.

Weekly, free. Unsubscribe any time. Privacy

Latest Articles

View All →