Skip to content

China drafts export controls on top AI models and chips

A draft rule covers 23 AI technologies, from model weights to chip designs. Open-source releases get automatic approval, and API use abroad stays unrestricted.

· · 3 min read

Share
The Ministry of Commerce of the People's Republic of China building in Beijing on a sunny day
Image: EditQ / Wikimedia Commons, CC BY-SA 4.0
On this page

China wants a say over where its most advanced AI goes. The Ministry of Commerce and the Ministry of Science and Technology have released a draft update to the national catalogue of technologies that are prohibited or restricted from export, adding 23 AI technologies that span large models, AI agents, computing clusters, data processing and chip design, The Decoder reports. Bloomberg, via The Straits Times, describes it as part of a wider push covering about 40 technologies, including solar and biomedical ones.

What would be restricted

According to the draft as reported:

  • Frontier model know-how. Model weights, checkpoints, source code, architecture and training or inference methods are covered once a model passes 10^26 FLOPs of training compute, has at least 2 trillion parameters, or ranks in the global top 10 on public leaderboards for performance, usage or safety.
  • Models built on sensitive data. Systems trained on what Chinese law calls "important data" also fall under the rules. Lawfully public data doesn't count.
  • Chip design. Design technology for AI chips made on 14 nm or more advanced processes, or using 2.5D/3D packaging, plus some advanced memory designs.
  • Cyber tools. AI that can write or alter malware on its own to attack critical infrastructure, the military or government systems faces an outright export ban. AI for finding software vulnerabilities would be restricted but not banned.
SituationUnder the draft
Foreign users call a Chinese model via API, app or websiteNot controlled
Weights published as open sourceAutomatic licence once filed
Technology already public (open source, papers)Exempt
Covered model installed on servers abroadStandard export licence
Same model on the Chinese provider's own overseas cloudFast-track possible
Autonomous malware tools aimed at infrastructureExport banned
Summary of the draft as reported by The Decoder and Bloomberg; the final rules may change.

Why open-weight fans are watching

Chinese labs have become some of the biggest publishers of open-weight models, the kind people download and run on their own hardware (see our guide to buying a laptop for local AI). The draft protects that for now with automatic approval for open-source releases. Even so, Dean W. Ball, head of strategic futures at OpenAI, wrote that the move "does not mean 'China is shutting down open-weight models,'" but gives Beijing "the optionality to shut down future open-weight models at will." He called it a "mirror image" of US controls on model weights.

A Commerce Ministry spokesperson said the rules are meant to ensure "the safety and controllability" of the technology and to prevent its misuse against critical infrastructure, defence industries and core government systems. Beijing also said the measures are not aimed at any particular country.

What happens next

This is a draft, so thresholds and exemptions could change. In the meantime, Bloomberg reports that a temporary registration requirement for exports of the covered technologies took effect on 16 October, including deals already under way. For a sense of how big open models have become, see Meta's 816B Muse Spark Open and NVIDIA's Nemotron 3.5 Bolt.

Share

More from Computers & AI

View all →

Read next