China weighs export curbs on top AI models

China is considering restricting foreign access to its most advanced AI models, according to The Decoder, after talks between the Ministry of Commerce and major domestic companies. The outlet reports the discussions involved Alibaba, ByteDance, and the startup Z.ai, and covered limits that could apply to both released and unreleased models. The scope and timing remain unsettled, with officials debating whether rules would apply only to future systems.

If Beijing follows through, the change would land less like a technical policy tweak and more like a rerouting of global AI supply. The Decoder notes that Chinese models such as Alibaba’s Qwen, ByteDance’s Doubao, and Z.ai’s GLM-5.2 have gained international adoption on a simple proposition: improving capability at lower cost. It reports GLM-5.2 is approaching US frontier performance “at a fraction of the cost,” a pricing dynamic that has made Chinese APIs and open models attractive to startups and corporate teams trying to ship products without paying frontier-model rates.

The Decoder describes Chinese officials weighing use of national security law to treat theft or transfer of protected AI technology as a legal violation, alongside tighter controls on who can fund domestic AI startups. That points to a broader pattern: once a model is viewed as strategic infrastructure, access becomes something governments meter rather than companies sell. The US has already moved in that direction. The Decoder cites US restrictions that barred foreign nationals from accessing Anthropic’s most advanced models, leading the company to shut down availability worldwide before partially restoring access after adding safeguards; another model remains limited to a small set of vetted US organizations and partners.

Europe is stuck with the bill either way. The Decoder argues the EU relies heavily on foreign AI development and could lose access to cheaper alternatives if China restricts exports, while also facing US-led controls that narrow what European firms can buy or integrate. Brussels has announced ambitions to build domestic capacity, including the InvestAI initiative, but The Decoder notes planned EU data centers for AI development are behind schedule and the EU’s AI budget is small compared with what American tech companies spend.

For European companies, the immediate risk is not that models disappear overnight, but that access becomes conditional: on nationality checks, security reviews, licensing, and political alignment. The cheaper “good enough” model that once looked like a procurement choice starts to resemble a sanctioned import.

The Decoder’s reporting leaves the central fact hanging: the talks have happened, the rules are not written, and the customers most dependent on imported models are the least able to influence either capital’s decision.