Washington tightens restrictions on AI chips exported to China

The United States is preparing new rules to block remote access to advanced AI chips via data centers outside the country. The measure mainly targets China and expands controls to several destinations considered at risk.

The U.S. Department of Commerce is preparing a new crackdown on advanced AI chips, with a clear goal: to prevent companies linked to China from accessing remote computing power via data centers located outside the United States. According to the interim rule published on November 16, 2026, licenses will be required for the export, re-export, or transfer of AI chips and servers to data centers worldwide when they can provide compute to entities in China or other restricted countries. This issue matters for a market where the boundary between hardware, cloud, and usage control is increasingly blurred. And the safeguards are not just for show.

Le nouveau filtre vise l’accès distant, pas seulement la livraison physique

Washington no longer limits itself to monitoring chip shipments to embargoed countries. The Bureau of Industry and Security’s text extends control to infrastructures located abroad as soon as they can serve, directly or indirectly, companies based in a Category D:5 country, including China, Russia, Iran, and North Korea. The logic is simple: block circumvention via third-party data centers, where U.S. restrictions had lost some effectiveness.

The threshold is precise: more than 12 million TPP, an American indicator of total computing power. This roughly represents the maximum capacity of about 750 Nvidia H100 or H200 chips, depending on the source. The rule also applies to indirect access, via intermediaries or brokers, and requires operators to verify the ownership structure and activity of their clients. Control therefore does not stop at the data center’s door; it follows the path of usage.

Des pays tiers dans le viseur, car le risque de diversion inquiète Washington

The rule also adds licensing requirements for exports of advanced AI chips to Malaysia, Singapore, Indonesia, Thailand, and Brazil. The Commerce classifies them among destinations with a higher-than-normal risk of diversion. However, deliveries intended for U.S.-based companies remain excluded from this part of the system, subject to additional reporting obligations.

Data centers already in operation and providing compute to affected companies must obtain a license within 90 days of the rule’s entry into force. BIS also specifies that controls apply to related infrastructures, whether physically or digitally connected and sharing workloads, resources, or data as a single unit. In this context, defining the “same” data center becomes almost an Olympic discipline for lawyers.

Le texte remplace l’AI Diffusion Rule de l’ère Biden

The project formally cancels the AI Diffusion Rule, adopted under Joe Biden in January 2025 but never implemented. That rule imposed country-specific caps on exports of advanced AI chips and also regulated certain powerful AI models. The Trump administration had already rejected it in May 2025, arguing it unnecessarily complicated life for American companies and their allies.

In return, the new framework focuses on end users, cumulative capacity, and foreign compute arrangements. The Commerce says it wants to prevent the use of these infrastructures to train or run advanced models that could support military or intelligence activities. The justification is primarily security-related, but its effects are very concrete for cloud providers, data center operators, and chip vendors: controls no longer only concern what is shipped, but what the recipient can later do with what they received.

Key points

  • Rule published on November 16, 2026, by the U.S. Commerce Department.
  • Trigger threshold: more than 12 million TPP.
  • Licenses required for five countries considered high risk.
  • Existing data centers have 90 days to comply.
  • The rule replaces the AI Diffusion Rule from January 2025.
  • The text targets entities linked to China, according to BIS.

By the numbers

  • 12 million TPP — computing power threshold, according to BIS, 2026.
  • 750 chips — approximate equivalent in Nvidia H100 or H200, according to the source.
  • 5 countries — Malaysia, Singapore, Indonesia, Thailand, and Brazil.
  • 90 days — deadline given to already affected data centers.
  • 2025 — year of the now rescinded AI Diffusion Rule.

To read