China Mulls Curbing Foreign Access to Its AI Models

Ronni Holmvig Strøm · 2026-07-10

Reuters reported on July 7 that Beijing has held talks with Alibaba, ByteDance, and Z.ai about restricting foreign access to their most advanced models, including systems not yet released. Nothing is decided, and the ministries involved have said nothing on the record. But officials went far enough to sketch the options: a bar on public release, or a limit to domestic use only. The detail that matters is what such a move would cost China. Openness is the entire reason its models reached the world in the first place.

A Trailing Player Abandoning Its Best Lever

Chinese labs trail America's frontier by roughly seven months on the benchmarks that count. They closed as much of that gap as they did by giving the models away. Most ship open-weight, the weights published online for anyone to download and run, and the price of free is exactly why some U.S. businesses adopted them over proprietary American systems to cut costs. A trailing player does not surrender its biggest advantage on a whim. It does so when the concern has changed from commercial to national security.

Which is why escalation is the wrong word for it. The better one is convergence. Washington has already decided who may use Anthropic's Mythos and OpenAI's GPT-5.6, restricting both over the models' facility at finding software vulnerabilities. Late last month Z.ai released GLM 5.2, which it claimed matched Mythos at spotting bugs. Beijing is now looking at the same lever Washington already pulled, and for the same stated reason. Scott Singer of the Carnegie Endowment put it plainly: China "is going to have to balance the benefits of access to global markets with a desire to control a technology that is central for national security." That is the White House's conversation of the past year, arriving in Beijing on a delay.

The backdrop is a sharpening fight over distillation, the practice of using a stronger model's outputs to train a weaker one. Anthropic reported in February that DeepSeek, Moonshot, and MiniMax generated 16 million exchanges with Claude through some 24,000 fraudulent accounts, and in June it accused Alibaba of "brazenly" trying to distill Claude.

Alibaba has since banned Claude Code internally. Both governments now treat frontier capability as something to fence. The vocabulary of the "silicon curtain" gets the direction right and the shape wrong. Both camps are discovering the same fact about their own technology.

Open Weights Do Not Come Back

One fact bounds the whole exercise. Once a model's weights are published, they cannot be recalled. Any restriction China imposes would touch only future releases. Everything already downloaded stays downloaded, runs on hardware Beijing does not control, and keeps serving the U.S. businesses that built on it. The supply-chain shock the headline implies is real for the next generation of models and close to nil for the current one.

That asymmetry also carries a cost Beijing has to weigh. China's open models were a diplomatic instrument as much as a commercial one, pitched as a development tool for lower-income countries that could not afford the more expensive American offerings. Close the door and the pitch loses its floor. Silicon Valley has already noticed the opening. Mark Zuckerberg has promised "aggressive" pricing on Meta's first premium model, aimed squarely at the adoption that cheap Chinese weights once locked up. A restriction meant to protect a lead could hand away the audience that made the lead matter.

Two of the field's most capable builders have arrived, from opposite starting points, at the same recognition. Models that cross certain capability thresholds carry real security weight, and the people shipping them know it. Whatever Beijing decides, the weights that seeded a global ecosystem are already out in the world, and the fight over pricing the next generation has already begun.