RealityRewi.red

Washington closed the cloud loophole on AI chips. It also stopped calling it AI

A new Commerce Department rule means a data center anywhere on Earth needs a US license if it rents serious compute to Chinese companies. The 132-page text also swaps "artificial intelligence" for "Super Intelligence," by executive order.

Reality Admin//Nov 16, 2026//4 MIN READ
Washington closed the cloud loophole on AI chips. It also stopped calling it AI

Since 2022, the United States has banned the export of its best AI chips to China. For most of that time, a Chinese company that wanted to train a large model on those chips had an obvious workaround. It didn't need to own them. It could rent them, over the internet, from a data center in Malaysia or Singapore or anywhere else with good power and loose questions.

On Monday the Commerce Department moved to shut that door. An interim final rule titled "Addressing Remote Access to Advanced Semiconductors" puts a licence requirement on exports of advanced chips to any data center, in any country, that supplies a big enough slice of compute to companies headquartered in China, Macau or the rest of the Country Group D:5 list. TechCrunch notes that D:5 covers China, Russia, Iran, North Korea and 18 other countries.

The rule calls it "a logical extension of existing controls." That is a fair description. It is also a large one.

Twelve million TPP, linked or not

The trigger is a number. A data center needs a licence if it will provide access to more than 12,000,000 "aggregate TPP," a measure of total processing performance, to D:5 or Macau companies. TechCrunch estimates that works out to the combined power of about 750 Nvidia H100 or H200 chips. That is not a frontier training cluster, but it is enough to matter.

The obvious dodge would be to split a customer across several sites, each under the line. The rule closes that by counting "linked data centers," and it is explicit that this includes "digitally-linked infrastructure, not just physically-linked infrastructure." It also reaches middlemen. If a data center knows it is renting to a non-Chinese customer that is in turn training models for a D:5 company, that counts. Chip exporters, data centers and even foreign foundries are told to do "robust due diligence" on their customers, including any direct or indirect ownership by people in the restricted countries.

Operators already doing this business get 90 days after the rule takes effect to obtain a licence or stop. One carve-out survives: compute used purely for "commercial recommender systems," the kind of models that pick your next product or your next video, is exempt, because in Commerce's words those "do not pose the same national security threat."

The stated reasons are not subtle. The rule says access to that much compute could help foreign actors "accelerate the development of weapons of mass destruction (WMD), execute sophisticated zero-day cyberattacks on critical infrastructure, and disrupt global military balances of power."

Five countries on notice

The second half of the rule is about places, not customers. Exports of advanced chips to Indonesia, Malaysia, Singapore, Thailand and Brazil now need a licence too. Commerce says these countries "present a higher-than-normal risk of diversion." Shipments to US-headquartered companies in those countries are exempt, as long as they file reports.

There is a stick for governments as well. Commerce says it "will be actively monitoring for countries that tolerate evasion" and lists possible responses, including import restrictions, new licence requirements, a downgrade in the country's export control status, and Entity List designations for people involved.

The timing is pointed. TechCrunch reports that Megaspeed, a company the US government reportedly wants on a restricted list, bought thousands of Nvidia chips to provide compute to Chinese businesses, and that Reuters reported last week that Chinese AI firms were racing to lock up computing capacity overseas. Southeast Asian cloud providers that built a business on that demand now have three months to get Washington's permission.

The rule also formally buries the Biden administration's AI Diffusion Rule from January 2025, the tiered country-by-country cap that was supposed to govern global chip sales. The Trump administration said in May 2025 it would scrap that framework. The new text adds, a little drily, that the Diffusion Rule "has never been enforced."

The footnote

Read the document and something odd jumps out on the first page. It never talks about AI chips. It talks about "advanced Super Intelligence (SI) semiconductors" used "to train advanced SI models." Even the recommender-system exemption is defined as a kind of "Super Intelligence" system.

Footnote 1 explains. Under Executive Order 14434, titled "Inaugurating the Era of Super Intelligence," the terms "Super Intelligence" and "SI" are now used in place of "Artificial Intelligence" or "AI." So the federal government's main tool for keeping advanced computing out of China now describes a model that recommends sneakers with the same label that AI labs use for a system smarter than every human alive.

That matters a little more than branding. Export rules get read by lawyers in dozens of countries, and the definitions are what they fight over. A data center operator in Johor trying to work out whether its customer's workload is a "commercial recommender system" or an "advanced SI model" will be parsing a vocabulary set by executive order.

The rule takes effect the day it is published in the Federal Register, with comments due 60 days later. Interim final rules apply first and take questions afterward. The question most operators will have is how Commerce plans to see who is really on the other end of an encrypted connection. The 132 pages don't answer it.

Sources: US Department of Commerce: Addressing Remote Access to Advanced Semiconductors (interim final rule), TechCrunch.

// RELATED_SIGNALS

More from the feed