Ai Engineering 7 min read

US closes the cloud loophole: new BIS rule licenses AI compute for China-linked customers and five diversion hubs

BIS's November 16 interim final rule requires a license for advanced AI chips going to any data center that will give China- or Macau-headquartered firms more than 12 million aggregate TPP, adds license requirements for Brazil, Indonesia, Malaysia, Singapore and Thailand, and formally buries the AI Diffusion Rule.

The US Commerce Department’s Bureau of Industry and Security (BIS) released an interim final rule on November 16 that, for the first time, puts a license requirement on renting American AI compute to Chinese companies through data centers outside China. The rule, titled “Addressing Remote Access to Advanced Semiconductors” (RIN 0694-AK50), also adds a new license requirement for advanced chips shipped to Brazil, Indonesia, Malaysia, Singapore and Thailand, and formally removes what was left of the Biden-era AI Diffusion Rule.

The BIS press release says the rule is meant to ensure that existing restrictions “cannot be circumvented through remote computing arrangements involving data centers in third countries.” It also says something unusual for an export rule: it was “written by and at the direction of the Department of Treasury and the National Security Council,” not by BIS alone. It lands four days after Jensen Huang’s resignation from Nvidia, a story driven by the Megaspeed affair and questions about whether Southeast Asian data centers were feeding Chinese demand for restricted chips (see our report on the board probe and Megaspeed).

The remote-access control: 12 million TPP

Until now, the US controlled where chips physically went. A Chinese company that could not buy Nvidia accelerators could, in many cases, rent them from a data center in a third country. The new section 744.24 of the Export Administration Regulations targets that gap directly.

According to the rule text, a license is now required to export, reexport or transfer advanced chips and servers (ECCNs 3A090 and 4A090 and their .z variants) when the exporter has “knowledge” that they will go to a data center, anywhere in the world, that will provide more than 12,000,000 “aggregate TPP” of computing power to companies headquartered in Macau or a Country Group D:5 destination (which includes China), or whose ultimate parent is headquartered there.

Three details make the control broader than a simple threshold:

  • Linked data centers count together. The 12 million TPP is measured “either at one data center or through linked data centers,” and BIS explicitly includes digitally linked infrastructure, not just facilities on the same campus. Splitting capacity across sites does not get an operator under the line.
  • It follows the chips after export. If a data center that bought chips legitimately later starts serving significant compute to D:5-headquartered companies, the owner or operator must obtain a license at that point.
  • Existing customers get 90 days. Any data center already supplying such customers must obtain a license within 90 days of the rule’s effective date “or face future export restrictions.”

For scale, Total Processing Performance is roughly dense 16-bit throughput multiplied by bit width. By that measure an H100 SXM sits near 15,800 TPP and a B200 near 36,000, so 12 million TPP is on the order of 760 H100s or about 330 B200s: a meaningful training cluster, but far below hyperscale. That is our arithmetic, not BIS’s, but it shows the threshold is aimed at frontier-model training rather than casual cloud use.

The one carve-out is for “commercial recommender systems,” such as those run by online retailers and social media companies, which BIS says “do not pose the same national security threat.” Using it is not automatic: the provider must file the customer’s name, the total TPP supplied and the due diligence performed through a BIS submission portal, and that information goes to the interagency End User Review Committee.

Due diligence moves to the exporter

The compliance burden is the part chip vendors and cloud operators will feel first. Note 3 to the new section says that to show they lack “knowledge,” exporters will be expected to obtain enough information from a foreign data center about its know-your-customer checks, including whether customers are directly or indirectly owned or controlled by persons in D:5 countries or Macau. If a data center declines or is unable to provide that information, the rule treats it as a red flag “that presumptively provides you with knowledge” unless resolved before shipping.

In practice, that pushes Nvidia, AMD, server makers and resellers to audit the customer books of the data centers they sell to. The preamble also restates BIS’s view that the older end-use control in section 744.23 was never limited to direct sales: it already covered cases where a D:5 entity was the purchaser or “the exclusive or primary (greater than 50 percent of total compute) remote end user” of a facility. BIS invites operators that have been doing this to file voluntary self-disclosures.

Five countries with “a higher-than-normal risk of diversion”

The second pillar is destination-based. Under a revised section 742.6, a license is required to export, reexport or transfer advanced AI chips to or within Indonesia, Malaysia, Singapore, Brazil and Thailand, which the rule describes as presenting “a higher-than-normal risk of diversion.” The Commerce Secretary, in consultation with the Secretary of State or the National Security Advisor, can add more countries. Applications are reviewed case by case.

There is a significant exclusion for American companies. Shipments to an ultimate consignee or end user headquartered in the United States, or whose ultimate parent is, do not need a license if the exporter files monthly reports under a new section 743.10. Each report must name every party and their ultimate parents, the destination, the ECCN, product name and SKU, TPP, quantity and export date. The upshot is that a US hyperscaler building in Johor or Singapore keeps shipping, while locally owned operators in those markets now need individual approval.

The rule also warns that Commerce “will be actively monitoring for countries that tolerate evasion,” and lists possible consequences: import restrictions, new license requirements, a downgrade in a country’s export-control status, and Entity List designations for the people involved. Megaspeed, the Singapore-based operator with data centers in Malaysia at the center of the Nvidia story, has denied diverting chips to China, and no court or regulator has made a public finding against it.

The AI Diffusion Rule is formally gone

BIS announced in May 2025 that it would not enforce the January 2025 “Framework for Artificial Intelligence Diffusion.” This rule finishes the job. It removes ECCN 4E091, the export control on closed model weights, along with the model-weights foreign direct product rule and the related red flag. It deletes the three-tier country system, License Exception AIA, License Exception LPP and the Data Center Validated End User authorization. BIS says model weights are better handled through destination-based and end-use license requirements, which means there is no longer a standalone US control on exporting frontier model weights.

The rule also tightens a semiconductor manufacturing pathway: front-end fabs and packaging houses (OSATs) making advanced-node logic chips for designers that are neither approved nor authorized under the existing 3A090 framework now need a license for any destination, which closes a route for China-linked chip designers using foundries in third countries.

One more sign of the times sits in a footnote: under Executive Order 14434, “Inaugurating the Era of Super Intelligence,” the rule uses “Super Intelligence” or “SI” everywhere earlier regulations said “artificial intelligence.”

What to watch

  • Federal Register publication. The posted text leaves the dates blank: the rule takes effect on publication and comments are due 60 days later, under a regulations.gov docket not yet assigned. The 90-day window for existing data centers starts at that effective date.
  • Licensing volume. BIS’s own paperwork estimate is strikingly small: about 100 additional license applications a year and roughly 970 extra report submissions. Whether that holds will depend on how many Southeast Asian operators turn out to have D:5 customers.
  • Nvidia’s earnings on November 17. Interim CEO Colette Kress will face questions about how much Southeast Asian and Brazilian revenue now needs licenses, and Nvidia has not publicly commented on the rule.
  • Host-country responses. Singapore, Malaysia and the other three countries are now named in US regulation as diversion risks. How they respond to that label, and whether Commerce follows through on the threatened consequences, will decide whether this becomes a durable compute border or another rule that the market routes around.
Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

How to Run TPU Workloads on Google Cloud with Ray 2.55

Learn how to provision Google Cloud TPUs, handle slice topologies, and deploy machine learning models using Ray 2.55 and the KubeRay Operator.

Ai Engineering

Jensen Huang resigns as Nvidia CEO: Colette Kress takes over, and the AI industry loses its backstop of last resort

Nvidia said on November 12 that Huang had resigned as CEO and director, effective immediately, with CFO Colette Kress as interim CEO. The exit puts more than $100 billion in lease guarantees, a $12.9 billion Hugging Face deal and US chip policy in new hands days before earnings.

Ai Engineering

Nvidia's board opens an independent probe after reports Jensen Huang was a Chinese intelligence target

On November 8, Nvidia's independent directors launched an investigation and Huang admitted an affair. The harder questions concern Megaspeed, chip diversion and what the Trump administration knew, with earnings on November 17 and a House briefing the next day.

Ai Engineering

Blackwell fingerprints in DeepSeek, Kimi and MiniMax weights: how checkpoint forensics reads the GPU behind a model

A November 12 report by former OpenAI researcher Will DePue says six open-weight models from four Chinese labs were initialized on export-controlled Nvidia Blackwell GPUs, using leftover random-number patterns in the published weights. Here is how the method works, what it proves and what it does not.

Ai Engineering

SpaceX Seeks $40 Billion in Debt to Buy Nvidia Chips

The Financial Times reported October 7 that SpaceX is seeking roughly $10 billion in bank loans plus $30 billion in investment-grade debt, led by Apollo and Pimco, to fund an Nvidia chip purchase, one of the largest corporate debt raises ever attempted.