On October 16, China's Ministry of Commerce and Ministry of Science and Technology proposed adding 40 technologies to the national export-control catalogue, and 23 of them are artificial intelligence: large models, autonomous agents, computing clusters and chip designs, according to Reuters and Bloomberg reporting. The quiet revolution in the draft is scope. Until now, export controls on AI, from Washington and Beijing alike, mostly reached the hardware underneath the models. This proposal pulls the models themselves inside the perimeter, covering weights, checkpoints, source code and architecture for the systems that clear its thresholds.
Across the full catalogue, which also adds five biomedicine and twelve solar technologies, 14 entries would be prohibited for export and 26 restricted. Reports on the draft say the restricted tier covers core technology for models at or above 2 trillion parameters, trained with at least 10^26 floating-point operations of compute, or ranked in the global top ten, and the Associated Press notes that Moonshot AI's Kimi K3 and Alibaba's Qwen 3.8-Max already exceed the size threshold. Prohibited outright is core technology for models trained on sensitive national data, along with agents built for defence, critical infrastructure and core government work. Chip rules reach design know-how at 14 nanometres or below and chips built with advanced 2.5D or 3D stacking.
The carve-outs are as deliberate as the controls. Using a restricted model through an API, a website, an app or a software-as-a-service platform is not treated as an export, and technology already in the public domain through open-source releases or academic publication sits outside the catalogue, with open-source exports reportedly qualifying for a simplified or automatic licence. In plain terms, DeepSeek and Alibaba can keep selling AI services to the world while Beijing gains oversight of who gets the underlying technology and who may deploy restricted models on overseas servers. This is not a final rule: public consultation runs until November 15. But an interim pre-export registration requirement for covered transfers took effect immediately, including for deals already under way.
The timing is not subtle. Reuters reported in July that Chinese authorities had already discussed restricting overseas access to the country's most advanced models with Alibaba, ByteDance and Z.AI. Beijing has watched Washington use chip export controls as strategic leverage for years and has now written the mirror image for software. In the same week, Fortune reported that the White House has no plans to publicly release its own framework for vetting frontier models before release, after labs including OpenAI, Anthropic, Microsoft, Meta and Nvidia met in Washington to review the draft. Two capitals, two rulebooks, neither fully public, and both treating frontier models as strategic assets rather than ordinary software.
For people building with AI, the practical lesson is that supplier risk now runs in both directions. If your stack depends on Chinese open weights, the weights already published are outside this proposal's scope, but the next generation may arrive API-only, under licence, or not at all. That changes architecture decisions: pin the model versions you depend on, keep local copies where the licence permits, document provenance for due diligence, and keep a tested fallback provider for any workload you cannot afford to lose. Export control used to be a problem for chip buyers. As of this week, it is a line item in every serious model-selection review.
