Beijing weighs curbs on China's AI models used by 80% of open-source developers

China's potential restrictions on advanced AI model exports would reshape the global AI market just as US curbs on American frontier models push developers toward Chinese alternatives.

Beijing's discussions with top tech firms about restricting overseas access to China's most advanced AI models threaten to upend the global open-source ecosystem, where Chinese models now power 80% of developer projects using open-source tools, according to Andreessen Horowitz research.

"Chinese authorities have held meetings with top tech firms over the past month about potentially restricting overseas access to China's most advanced AI models, including those yet to be released," one of three people familiar with the discussions said, speaking on condition of anonymity.

The move comes weeks after the Trump administration imposed export controls on Anthropic's Claude Fable 5 and Mythos 5 models, forcing the company to disable both for more than two weeks. The US government also asked OpenAI to delay the public rollout of its GPT-5.6 series. Those actions, combined with rising costs of frontier models, have accelerated enterprise adoption of cheaper Chinese alternatives such as DeepSeek, Alibaba's Qwen and Z.ai's GLM-5.2, which climbed industry leaderboards after the Anthropic restrictions took effect.

If implemented, Beijing's restrictions would cut off global developers from the Chinese models that have become the backbone of open-source AI development. A study from MIT Management found open-source models accounted for a fifth of all AI token usage, and Chinese models dominate that segment. The bifurcation of the global AI ecosystem — with the US restricting its frontier models and China potentially doing the same — would leave developers with fewer options and higher costs, while accelerating the divergence of AI capabilities between the two blocs.

US Restrictions Already Reshaping the Market

The Trump administration's June 12 export control order against Anthropic gave the company 90 minutes to pull its Claude Fable 5 and Mythos 5 models after Amazon raised cybersecurity concerns about a potential jailbreak technique. Anthropic complied, and the models went dark until the Commerce Department lifted restrictions last week after the company implemented new safeguards.

The episode crystallized a central tension in US AI policy: the administration's light-touch regulatory approach collided with national security concerns, creating uncertainty for enterprises building on proprietary models. "When a company gets 90 minutes to pull a model deployed to hundreds of millions of people because of a competitor's complaint, that need becomes urgent fast," Felix Van de Maele, CEO of data intelligence platform Collibra, said.

OpenAI faced similar pressure when the US government asked it to stagger the release of GPT-5.6, limiting initial access to a small group of vetted partners. The restrictions showed how private AI firms could be forced to cut off thousands of customers with little notice.

Chinese Models Fill the Gap

Chinese open-source models have emerged as the primary beneficiaries of US regulatory uncertainty. Airbnb CEO Brian Chesky said the company was "relying a lot" on Alibaba's Qwen model, describing it as "very good" and "fast and cheap." Z.ai's GLM-5.2 gained Silicon Valley users during the Anthropic outage because it was immediately available and less expensive.

Alex Karp, CEO of Palantir Technologies, recently criticized the token-based pricing used by frontier labs like OpenAI and Anthropic, arguing enterprises want "control over their compute, their models, their data stack and their alpha." Chinese open-weight models, which can be downloaded for free and hosted in private clouds, offer exactly that.

The competitive dynamics are shifting rapidly. While US frontier models remain more capable on advanced benchmarks, the gap is narrowing. "We're starting to see the open models coming out of China catch up with the frontier," one AI industry executive said. "For a while there's been this narrative of America winning the AI race. While we've been winning on capability with the best models, we've been losing."

For investors, the stakes are substantial. US AI companies and their backers have committed hundreds of billions of dollars to developing frontier models, betting that corporations worldwide will use them as the intelligence engines for major business functions. If both the US and China restrict access to their best models, the addressable market for each shrinks, and the economics of massive training runs — some costing $4 billion or more — become harder to justify. Nvidia, whose H100 and Blackwell GPUs power much of the AI training infrastructure, faces a complex outlook: restrictions could slow model development, reducing GPU demand, while also pushing enterprises toward smaller, more efficient models that require fewer chips. Chinese AI chip makers such as Huawei, with its Ascend series, stand to gain if Beijing's restrictions push global developers toward Chinese hardware as well.

This article is for informational purposes only and does not constitute investment advice.