China proposes export controls for 23 AI technologies

China is proposing tighter export controls for 23 artificial-intelligence technologies, including large models, AI agents, computing clusters and information-processing systems. The proposal is part of a wider draft revision covering 40 technologies across AI, biomedicine and solar power, according to an announcement from China’s Ministry of Commerce and Ministry of Science and Technology. ¹

What the draft would cover

The official notice says the authorities are asking for public feedback on additions to China’s prohibited and restricted technology catalogue. The consultation runs from October 16 to November 15, while exporters must register affected technology transfers before the revised catalogue formally takes effect. ¹

For large AI models, the draft identifies several triggers for controls. The Decoder reports that the proposal covers systems trained on China’s “important data”, as well as core model technologies such as weights, checkpoints, source code, architecture and training or inference methods. The thresholds described in the reporting include 10^26 FLOPs of training compute, at least two trillion parameters or a place in the global top ten of major evaluations. ²

The proposal also takes aim at autonomous systems capable of writing or altering malware for attacks on critical infrastructure, military targets or government agencies. Some vulnerability-finding tools would face export limits rather than an outright ban. The draft’s scope therefore reaches beyond consumer chatbots, covering the underlying components and infrastructure used to train, deploy and transfer advanced AI systems. ²

The rules are not framed as a blanket shutdown of open-weight models. China’s official notice says technology already made public through open-source releases or academic papers is outside the proposed management scope. It also describes an automatic-licensing route for eligible open-source releases and for certain overseas data-centre, chip-production and international research activities. ¹

That distinction matters for companies operating internationally. Using a restricted model through an API, app, website, SaaS service or lightweight local version would remain outside these export controls, according to The Decoder’s reading of the draft. Installing restricted AI technology on overseas local servers, however, would require a standard export licence, while some company-operated overseas cloud deployments could use the streamlined process. ²

The consultation is still at the proposal stage, so the final list, thresholds and implementation details could change. For European businesses, the move adds another layer to an already complicated AI supply chain: even if a model is publicly documented, its weights, training methods, source code or deployment environment may still have different regulatory treatment depending on how the technology is transferred.

Our opinion

This is best understood as a framework announcement, not a finished ban. The headline number is striking, but the practical effect will depend on the final catalogue, the meaning of its technical thresholds and how officials treat open releases that are already public. The automatic-licensing route suggests Beijing wants tighter control over strategic transfers without freezing the domestic AI sector. For companies in Britain and Europe, the immediate lesson is less about abandoning Chinese models and more about checking exactly which layer of an AI system is being moved, where it is hosted and whether the transfer is a licence-triggering export.