
Illustration
Developments3 min read
US rule requires licences for data centres renting AI compute to China
The US Commerce Department issued an interim final rule on 16 November requiring licences for data centres worldwide that could supply AI compute to companies in China. It applies above 12 million TPP, roughly 750 Nvidia H100 or H200 chips, TechCrunch reported, and existing operators have 90 days after it takes effect to apply. Whether a given GPU provider serving European customers is covered remains open.
Key facts
- The US Bureau of Industry and Security published the rule on 16 November 2026; it requires licences for AI chips and servers going to data centres worldwide that could supply compute to companies in China or other D:5 countries.
- The threshold is more than 12 million TPP, roughly 750 H100 or H200 chips, TechCrunch reported; existing operators have 90 days after the rule takes effect to apply.
- Licences are also required for advanced chips going to Malaysia, Singapore, Indonesia, Thailand and Brazil, TechCrunch reported; chips for US-based companies are exempt.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
According to TechCrunch, the rule requires licences to export, re-export or transfer AI chips and servers to data centres anywhere in the world if the operator may supply compute to companies in China or other restricted countries. These are countries in the D:5 group, including China, Russia, Iran and North Korea. Indirect access through intermediaries and brokers is covered too, and operators and exporters must check who owns their customers.
The threshold is more than 12 million TPP (total processing performance, a US measure of computing power), TechCrunch said. Regardless of where operators are based, licences also become mandatory for advanced chips going to Malaysia, Singapore, Indonesia, Thailand and Brazil, which Commerce considers at higher risk of diversion, the outlet reported; chips destined for US-based companies are exempt. Existing operators have 90 days after the rule takes effect to apply. The rule also rescinds the Biden administration's AI Diffusion Rule from January 2025.
The Commerce Department's press release gives none of these figures. It says existing restrictions must not be circumvented through data centres in third countries. The Federal Register text was still pending on 17 November, according to Dr. Web, one of few German-language reports, which also mentions a 60-day comment period.
AI chips: what changes for data centres in Southeast Asia and Europe
About 75% of China's overseas AI compute is forecast to sit in Southeast Asia by the end of 2026, CNBC reported, citing October estimates from the Institute for Progress. Oracle cancelled a multibillion-dollar chip-leasing deal with Tencent covering about 100,000 chips, CNBC said, citing the Financial Times. Jessica Brandt of the Council on Foreign Relations told CNBC the rules will likely make these data centres "slower and costlier". Her colleague Chris McGuire warned that Chinese companies could pivot to data centres in Virginia.
The Commerce Department's press release does not mention Europe. CNBC reported on 5 November that the Chinese AI lab Moonshot AI had sought Nvidia chips through European data centres. Nvidia has warned buyers, CNBC reported, that a bill that has not yet passed, the Remote Access Security Act, could create legal risk. For now the company reportedly will not sell to customers that rent out more than half their GPUs to Chinese firms. For background on Nvidia, see Huang's resignation and the earnings without Huang.
Analysis
The rule targets operators, but their customers may end up paying. Companies in Europe that rent GPU capacity for model training, image generation or their own language models rely on platforms that are likely to pass new screening duties on to them, through ownership questionnaires, tighter contract clauses or higher prices. Whoever builds on someone else's infrastructure inherits its terms and risks, and the export rule is a concrete case. A data centre in Germany, Austria or Switzerland offers no shelter on TechCrunch's reading, since the duty covers data centres worldwide. What counts is a provider's customer base, not its location.
Teams that rent AI compute have two simple steps. First, ask the provider in writing whether its systems exceed the 12 million TPP threshold and how it screens customers. Second, check the contract for who bears price rises or termination if a licence is required. This does not replace legal advice, but it is a useful basis for the next call with the provider.
Got AI news? Submit news →
FAQ
What does the new US AI chip export rule do?
According to Commerce, it stops restrictions being bypassed through data centres in third countries. Licences are required where data centres could supply compute to companies in China or other D:5 countries.
Does the rule apply to data centres in Europe?
TechCrunch says the licence requirement covers data centres worldwide above 12 million TPP. The Commerce release does not mention Europe, so whether individual providers are affected is open.
How long do existing operators have?
Existing operators must apply for a licence within 90 days after the rule takes effect, TechCrunch reported.
Source
More sources
- TechCrunch – Schwelle 12 Mio. TPP, D:5, Länderliste, 90 Tage, abgerufen 19.11.2026
- CNBC – Südostasien, Oracle/Tencent laut Financial Times, Zitate, abgerufen 19.11.2026
- CNBC – Remote Access Security Act, 50-Prozent-Grenze, abgerufen 19.11.2026
- Dr. Web – Deutschsprachige Zweitquelle: Federal-Register-Fassung ausstehend, 60 Tage Kommentarfrist



