China weighs AI export curbs on DeepSeek and top LLM models

If Beijing follows through on limiting overseas access to its strongest models, this would mark a sharp turn away from the last 18 months of aggressive open-weight and API exports. Qwen, Doubao, GLM‑5.x and DeepSeek’s R‑series have been the main price ceiling on what US labs could charge globally for frontier‑ish capabilities. Tightening that spigot would immediately reduce competitive pressure in many developing markets where Chinese models have gained share on price. ([ndtvprofit.com](https://www.ndtvprofit.com/technology/china-hardens-ai-wall-overseas-access-for-deepseek-future-models-to-be-curbed-11738661/amp/1))

Strategically, China appears to be converging with Washington on treating frontier models more like dual‑use strategic assets than generic cloud software. Discussions reportedly include making leaks or theft of proprietary AI tech a national‑security offence and scrutinising which foreign investors can back Chinese AI startups. That’s export control logic applied to models and weights, not just chips.

For the race to AGI, this is about who gets to train, fine‑tune and study models at the edge of capability. If future Qwen or GLM releases are effectively domestic‑only at the high end, research groups and startups outside China will lean even harder on US‑aligned ecosystems. It would also entrench a world where the most capable systems are clustered behind a small number of political borders, with cross‑border collaboration becoming the exception, not the norm.