China's Ministry of Commerce has discussed with major domestic AI companies, including Alibaba, ByteDance, and Zhipu AI, the possibility of restricting the overseas transfer of key data used to train AI models, according to a Financial Times report relayed by chip analyst Jukan. The discussions reportedly also cover limits on foreign users downloading the actual weights of Chinese AI models, the trained parameters that let anyone run a copy of the model on their own hardware once obtained. Crucially, the report says China plans to keep letting overseas customers access these models and services normally, meaning the restriction would target the underlying training data and downloadable model files rather than cutting off API access or commercial use abroad. The move would mark a shift in how Beijing treats its AI industry, protecting the data and weights behind homegrown models like Alibaba's Qwen and Zhipu's GLM series the same way it already restricts sensitive technology exports, even as Chinese labs have built global usage partly by giving away open-weight models for free. If adopted, the policy would tighten China's grip on AI IP just as its open-weight models, from DeepSeek to Moonshot's Kimi, have been gaining significant traction with developers outside China.
Why it matters · If China locks down the training data and downloadable weights behind its AI models, it could slow the open-weight strategy that has made Chinese labs like Alibaba and Zhipu AI serious global competitors, even as it keeps the door open for paid access to their cloud services.
Worth asking · China wants to keep selling AI model access to the world while locking up the data and weights behind those models. Is this Beijing protecting a strategic asset the way it already does with chips, or does it undercut the open-weight strategy that made Chinese AI popular in the first place?