China weighs curbs on AI models, training data, and foreign chipmaking

Beijing is reportedly considering rules that would bar domestic firms from using TSMC and restrict exports of AI models and training data.

According to a Financial Times report cited by Tom's Hardware, China's Ministry of Commerce has consulted domestic AI and chip companies about a set of potential export restrictions. The measures under discussion would cover advanced AI models, training data, and overseas acquisitions of strategic technology firms. Regulators reportedly spoke with Alibaba, ByteDance, and Zhipu about limiting transfers of AI training data abroad and restricting foreign downloads of model weights.

The most consequential idea would prohibit domestic chip designers such as Alibaba, ByteDance, and Huawei from manufacturing advanced processors at foreign foundries like TSMC, which leads China's SMIC on process technology. Another proposal targets a loophole that let Meta acquire Manus for $2 billion, a deal Chinese authorities ordered undone. The measures could join China's export catalogue, which already lists rare-earth materials and battery production technology.

For shared AI progress, this points toward fragmentation. Compute, model weights, and training data are the raw inputs of a fast-moving field, and walling them off by nation slows the cross-border diffusion that has driven progress. Notably, overseas users could still access Chinese AI services remotely, and open-weight models from DeepSeek and Moonshot already circulate freely, unlike closed flagships from Anthropic and OpenAI.

The key caveat: none of this is enacted. The reporting rests on two people familiar with the talks, and proposals under consultation often change or stall. Watch whether any of it lands in the next revision of the restricted-technology catalogue.

Source: Tom's Hardware