China's commerce and science ministries have proposed an update to the national list of restricted export technologies. The draft covers 23 categories of AI, including large models, agents, compute clusters, and data processing tools. AI systems capable of generating malware for attacks on critical infrastructure face a complete export ban, while tools for identifying software vulnerabilities will be subject to new export restrictions.
Export controls on large models apply if they are trained on sensitive data defined by the Data Security Law. Exemptions exist for public data and specific cases where sensitive data accounts for less than 0.1 percent of the training set. Controls also trigger for models exceeding 10^26 FLOPs of training compute, containing at least 2 trillion parameters, or ranking among the top 10 globally based on performance and usage metrics.
The regulations cover core technologies, including model weights, checkpoints, source code, and architecture. However, AI technology already released as open source or published in research papers remains exempt. Beijing is introducing a fast-track automatic licensing process for companies that publish restricted tech as open source, conduct research with foreign partners, or operate their own data centers abroad.
While API access, SaaS tools, and lightweight versions on local devices remain unaffected, installing restricted AI technology on local servers overseas requires a standard export license. Companies operating their own foreign cloud infrastructure can utilize the streamlined filing process. Until the final rules are enacted, businesses must register affected exports via the commerce ministry's portal.
