For four years the story of AI export controls had one direction: Washington restricting what could flow into China. Advanced Nvidia chips, the software used to design them, the machines that make them, all blocked or licensed at the US border. On Friday, Beijing answered with a mirror. China's commerce and science ministries proposed adding 40 technologies to the national catalogue of controlled exports, and 23 of them are AI. The targets are not finished products but the ingredients: large models, intelligent agents, computing clusters, chip-design methods, and, most strikingly, the model weights and source code behind them.
The draft splits into two tiers. Fourteen technologies would be banned outright, including AI tools for generating and mutating malicious code, and models trained on what China's Data Security Law calls "core data", the information tied to national security and the economy. The other 26 would be restricted, meaning exportable only with a licence. Within the AI bloc, the restriction bites hardest on heavyweight systems: the draft reportedly covers model weights, checkpoints and architecture for anything above roughly two trillion parameters, or ranking among the world's top ten on performance, usage and safety. That is a description that fits China's own champions, Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max among them.
There is a deep irony here. China's open-weight models have been its great AI success of the past two years, precisely because they are freely downloadable. DeepSeek, Z.ai and Alibaba built global developer loyalty by giving away systems that Western labs charge for, and Beijing has treated that openness as soft power, a way to win markets and trust in countries wary of depending on the United States. Now the government wants a lever over the most capable of those releases. The draft keeps an open-source carve-out, so anything already published stays exempt, but it formalises a path to control future frontier models before they leave the country.
Two design choices soften the blow, and reveal the intent. Technologies already in the public domain are excluded, and, crucially, access through APIs, apps and cloud platforms is reportedly out of scope. A foreign developer calling a Chinese model over a hosted endpoint would sit outside the rules; only downloading the weights or licensing the code triggers them. That tells you Beijing's worry is the transfer of the technology itself, not foreign use of Chinese AI services. Companies like DeepSeek can keep courting overseas customers through the cloud even as weight downloads face new friction.
What unsettles trade lawyers is the timing. The catalogue is only a draft, open for public comment until 15 November. But a registration requirement took effect the same day it was published, and it reaches backwards: deals already in motion, mid-flight licensing arrangements, even ownership changes that put Chinese AI technology under foreign control, all must be registered now. Beijing frames the measure as keeping AI "safe and controllable". The subtext is harder to miss. The truce struck with Washington last month is fraying, and the world's two AI superpowers are now building the same kind of wall from opposite ends of the supply chain.