China's Ministry of Commerce sat down last month with Alibaba, ByteDance, and AI startup Z.ai to discuss restricting foreign access to their most advanced AI models. The talks covered both closed systems and open-weight models, meaning even freely downloadable versions could be affected. Officials also floated making the leak or theft of AI technology a national security offence.
No final rules exist yet. The scope is still being debated, and it may only apply to future models. But the direction is clear and it follows a series of steps Beijing has already taken: blocking Meta's acquisition of Chinese-founded startup Manus, tightening rules on overseas business dealings, and investigating startups that moved abroad for possible export violations.
The US moved first. In June, the Trump administration ordered Anthropic to cut off all foreign nationals from its two most powerful models, Fable 5 and Mythos 5. Because verifying nationality in real time is impossible, Anthropic had no choice but to shut the models down for every user in the world. They stayed offline for roughly three weeks before the Commerce Department lifted most restrictions. Mythos, the version built specifically for advanced cybersecurity tasks, remains available only to a vetted group of US organizations.
Chinese officials reportedly see Mythos as a direct threat, fearing it could be used to find and exploit software vulnerabilities in Chinese systems. A Chinese security executive has already called publicly for Beijing to build a domestic equivalent.
For businesses outside both countries, the practical lesson is the same regardless of which side restricts first. Any AI model, American or Chinese, can be restricted, repriced, or cut off based on decisions made in Washington or Beijing. The idea that Chinese open-source models offered a stable, sovereign alternative now looks much shakier.
Europe is the clearest example of the bind this creates. The EU depends on foreign providers for more than 80 percent of its digital products, services, and infrastructure. In AI specifically, since 2017, roughly 73 percent of foundational AI models have come from the US and 15 percent from China, leaving Europe almost entirely dependent on foreign-designed AI. In 2023, Europe attracted just 8 billion dollars in AI venture capital, compared to 68 billion in the United States.
The EU has a plan: 200 billion euros in AI investment under the InvestAI initiative, with 20 billion earmarked for four to five large AI data centers called gigafactories. The formal tender for those facilities has been pushed to summer 2026, with construction scheduled to begin in 2027. Any real output from those facilities will not arrive before 2028 at the earliest.
Meanwhile, the four largest US tech companies, Amazon, Alphabet, Microsoft, and Meta, could spend a combined 700 billion dollars on AI investment in 2026 alone. Europe's entire multi-year plan is roughly one-third of what four American companies plan to spend in a single year.
There is also a talent drain accelerating the problem. In 2025, 62 percent of EU AI researchers said they planned to move to the US or China. Beyond people leaving, the expertise of those who stay is itself being extracted: platforms now pay specialists to label data and evaluate AI outputs, feeding that knowledge directly into models trained and owned abroad.
For any business currently using Chinese AI tools as a cost-saving measure, the question worth asking now is not whether restrictions will come, but whether your operations could absorb a sudden cutoff. The US Anthropic episode showed how fast things can move: models went offline the same evening the government sent a letter. Building a workflow that depends entirely on a single foreign AI provider, from either country, is a supply chain risk that is now explicit rather than theoretical.