China imposes export controls on 23 AI technologies
China has taken its most significant step since 2020 to protect its most sensitive artificial intelligence technologies. On October 16, the Ministry of Commerce (MOFCOM) and the Ministry of Science and Technology (MOST) issued Announcement No. 48, proposing to add 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export — 23 of them AI-specific.
The proposal, open for public comment until November 15, creates a formal outbound control regime that mirrors what already exists in the United States. Unlike the 2020 catalogue revision, which focused on personalized recommendation and AI interaction interface technology, the draft reaches frontier model weights, agent architectures, and AI-enabled offensive cyber capability, with explicit technical thresholds for each item.
What is being controlled
The 23 AI items break down into five categories: large models, agents, compute clusters, chip design, and information-processing technologies. Each has specific technical thresholds that determine whether the technology is prohibited or restricted.
For large models, the proposal prohibits the export of models trained on "core data" — defined by China's Data Security Law as data related to national security, the economy, public wellbeing, and collective interest. Models trained on "important data" (a lower-risk category, but still protected) would be restricted, requiring a license for export.
Model weights, source code, key checkpoints, architecture, and training and inference optimization know-how come under restriction if any one of three tests is met: at least 10²⁶ FLOPs of training compute, at least 2 trillion parameters, or a position in the global top-10 on public evaluations, global call volume, or comparable measures at the time of export.
For AI agents, the criteria are even more granular. Planning and task decomposition come under restriction when the system achieves chains of 20 or more subtasks, 80% decomposition accuracy, and 90% subtask success. Multi-agent collaboration is restricted at 60 or more concurrent agents, completion time of at most twice the single-agent minimum, and 85% agreement with reference answers. Tool calling requires 200 or more integrated third-party tools, 90% tool-selection accuracy, and 90% end-to-end success on tasks requiring three or more tools.
High-performance agents designed for critical infrastructure, defense, or core government operations are prohibited — with no licensing option. AI-enabled malicious-code generation, adaptive evasion, and variant generation with thresholds usable against high-sensitivity systems are also on the prohibition list.
What is already in force
While the catalogue revision is a draft, an interim pre-export registration requirement already applies. Multinational companies must report recipient details, destination, technology description, end use, and estimated value for technology ownership transfers, foreign-use authorizations, and changes in actual control of technology based in China.
MOFCOM has described the interim registration as covering exports that have started but are not yet complete — meaning companies that already have AI licensing agreements with Chinese partners may need to register immediately.
Generous carve-outs
The announcement includes significant exemptions that soften the impact on international collaboration. Products manufactured using restricted technologies are not subject to the restrictions. AI technologies already in the public domain are exempt. Automatic licensing is planned for open-source releases, Chinese companies operating overseas data centers (or deploying AI on their own cloud infrastructure abroad), overseas chip fabrication, joint R&D using foreign resources, and participation in international standard-setting.
This means most Chinese open-weight models — including Alibaba's 2.4-trillion-parameter Qwen3.8-Max — likely will not be affected, since they are already in the public domain.
Parallels with US controls
The 10²⁶ FLOP threshold in the Chinese draft mirrors the figure used in the US 2023 executive order and the 2025 US AI Diffusion rule (since rescinded). According to the Cloud Security Alliance, no Chinese model currently released exceeds that threshold — suggesting Chinese controls are more preventive than restrictive at present.
CSA notes that frontier AI supply chains now face formal outbound controls from both the world's two largest AI-producing jurisdictions. While US controls focus on chips and hardware, Chinese controls extend into model weights, agent architectures, and offensive cyber capabilities.
What to watch
China has had technology export controls since 2001, with revisions in 2006, 2014, 2017, and 2020. But this is the first time controls cover frontier model weights and agent architectures — the actual frontier of modern AI. If the final catalogue retains these items (which is likely, given the broad technical consensus behind the proposal), China will become the second country, after the US, to formally control the outbound flow of frontier AI technologies.
For AI companies on both sides of the Pacific, the message is clear: the era of free circulation of AI technology is ending, and chip geopolitics now includes model geopolitics.
Sources: AI Breaking Wire, China Briefing, JDSupra / Akin Gump
✓ Independent sources cross-checked and verified before publishing
