China's First Frontier AI Export Controls: Mofcom Announcement No. 48

China’s Ministry of Commerce and Ministry of Science and Technology proposed adding 40 technologies to China’s export-control catalogue, including 23 in AI, while immediately imposing a temporary registration requirement for certain transfers. The draft, open for comment until November 15, 2026, covers frontier-model weights, checkpoints, code, architecture, and training or inference techniques when a model meets any of three tests: at least 10^26 training FLOPs, two trillion parameters, or a global top-ten standing on specified measures. It also controls agent capabilities independently of the model, including advanced planning, multi-agent coordination, and tool use; high-performance agents designed for sensitive government, defense, or infrastructure uses, as well as certain offensive cyber technologies, could be prohibited outright. Other controls address AI chips, compute clusters, vulnerability discovery, and data-security tools.

The practical dividing line is whether technology itself moves abroad: API access, apps, SaaS, and lightweight on-device use are generally outside the large-model export rules, while local deployment of a restricted model overseas requires a license. The interim registration regime already applies to ownership transfers, overseas-use licenses, foreign changes of control, and other designated cases, including exports already underway. Publicly released technology is excluded, but newly developed restricted technology cannot simply be open-sourced without reporting and licensing. The draft’s changing “global top ten” test and unclear criteria for some agents and open-source cases complicate compliance. Companies may need to assess model versions, hosting arrangements, licensing deals, and intra-group technology sharing, as China’s rules create formal frontier-AI export controls alongside those of the United States.