China’s Ministry of Commerce has been consulting with leading domestic AI and semiconductor companies on potential export controls that could restrict foreign users from downloading the model weights behind the country’s most advanced AI systems, according to a Financial Times report, a move that would mirror a similar restriction the United States imposed on its own frontier models earlier this year.
Reuters separately reported that Chinese officials met with AI companies this month to discuss limiting foreign access to advanced models, including some not yet publicly released. Companies consulted reportedly include Alibaba, ByteDance and Zhipu. The proposals under discussion extend beyond software to hardware; with regulators also weighing whether overseas chipmakers such as Qualcomm and TSMC should be barred from manufacturing advanced chips designed by Chinese firms, and whether foreign acquisitions of Chinese companies working on agentic AI should face tighter scrutiny. Beijing has not finalized any of the proposals and continues to gather industry feedback.
The deliberations echo a decision the U.S. government made in June, when it ordered Anthropic to cut off foreign nationals from its two most capable models, Claude Fable 5 and Mythos 5, citing national security concerns tied to their cybersecurity capabilities.
Access was suspended for more than two weeks before being restored. Neither government has moved to ban AI exports outright; both are instead drawing a line around their most advanced systems and deciding, on a model-by-model basis, who outside their borders can access them and under what terms, a shift from the previous approach among frontier labs in both countries of competing largely by distributing models as widely as possible.
That dynamic falls directly across markets such as Australia’s, which has no frontier AI lab of its own and relies almost entirely on foreign model providers. Australian small and medium enterprise AI adoption reached 44% in February, according to the National AI Centre’s SME AI Pulse tracker, with Deloitte’s 2026 State of AI in the Enterprise report finding Australian organizations investing in AI at a pace comparable to global peers.
That adoption has been built largely on infrastructure Australia does not control; any enterprise, hospital network or government agency that has built workflows around a foreign frontier model is, in practice, one export-control decision away from losing access to it, as Anthropic’s customers briefly experienced in June.
The Australian Computer Society has estimated that competitive national AI infrastructure would require between AU$2 billion and AU$4 billion in investment, a fraction of the cost of training a genuine frontier model from scratch.
Analysis published in The Conversation has broken Australia’s AI sovereignty question into four components: where data physically sits, who controls the underlying compute, whether models are built domestically, and whether Australia sets its own AI rules rather than inheriting other countries’ export-control decisions by default.
Australia currently has only partial answers on each front. Providers including Macquarie Government, Vault Cloud and AUCyber offer sovereign data and compute services, but no Australian-built frontier model exists, and most model inference for Australian users continues to occur offshore.
