Chinese authorities have spent the past month in meetings with the country's top AI firms discussing whether to restrict overseas access to China's most advanced models, including ones that haven't shipped yet. Alibaba, ByteDance and startup Z.ai were all in the room, according to a Reuters report citing three people familiar with the talks, who weren't authorized to speak publicly.
Nobody involved would say exactly what the restrictions might look like. That's the maddening part of this story. Reuters couldn't determine a mechanism, and the companies aren't commenting on the record.
A tiered system, maybe
The closest thing to a blueprint comes from a May roundtable of Chinese legal scholars on open-source AI regulation, summarized in an official Supreme People's Court journal. The proposal on the table: basic open-source tools get a simple filing requirement, more advanced systems face security review, and the most sensitive frontier models get barred from public release entirely or restricted to domestic use only.
That's a legal academic's wishlist, not a policy. Whether it survives contact with Alibaba's Qwen team or ByteDance's commercial roadmap is a different question, and one Reuters' sources didn't answer.
Officials reportedly also floated making leaks of proprietary AI technology a national security offense, and new limits on who's allowed to fund domestic AI startups. Both would be consistent with a government that's been steadily tightening the leash on its AI sector all year.
The Manus precedent looms over this
Context matters here. In April, China's state planner, the National Development and Reform Commission, ordered Meta to unwind its roughly $2 billion acquisition of agentic AI startup Manus, a deal that had already largely closed, with employees moved and investors paid out. Meta reportedly began dismantling the arrangement in June, splitting Manus staff off from its internal systems. Beijing gave no detailed explanation beyond a one-line statement citing national security law.
Then in early June, authorities issued sweeping new rules tightening control over cross-border deals touching Chinese investors, technology, or data. China has also opened investigations into Manus and other AI startups that relocated abroad, checking whether they violated export control law.
Put those together and the pattern is obvious: Beijing is treating frontier AI the way it treats rare earths or advanced chips. A strategic asset you don't let walk out the door, even when your own companies built it to compete globally.
Why the open-weight crowd should be nervous
This is where it gets uncomfortable for a lot of people outside China who never thought about Chinese industrial policy until they started running a Qwen or GLM model in production. Since DeepSeek released its R1 reasoning model in January last year, open-weight Chinese models have become the default cheap option for startups and developers who can't afford frontier API pricing from OpenAI or Anthropic. If Beijing starts gating the top tier behind domestic-only rules, that option doesn't disappear overnight, but the newest and best versions might stop showing up on Hugging Face the day they're trained.
One industry newsletter noted that Chinese LLMs' global market share reportedly jumped sharply in the months after R1 launched, with the steepest gains in developing countries. If that pricing pressure lifts because the top models go dark internationally, US labs get some room back in exactly the markets where DeepSeek hit them hardest.
Worth remembering: this is one Reuters story built on three anonymous sources, and even they say any new rules might only apply to future models, not what's already out in the world. Nothing here is finalized. Watch for the actual draft rules rather than the tone of the meetings, which is the part that's hardest to verify from outside the room.
