Chinese officials have held discussions with domestic AI labs about restricting overseas access to the country’s most advanced artificial-intelligence models, a shift from Beijing’s previous strategy of promoting open-source models globally, according to people familiar with the matter.
The potential restrictions would affect American companies that have increasingly turned to cheaper Chinese AI models as alternatives to U.S. frontier systems, and reflect a tightening of technology controls as the U.S.-China AI race intensifies.
The discussions mark a significant shift in China’s approach to artificial intelligence. Until recently, Beijing wanted to encourage the rapid spread of Chinese AI models globally, believing this represented a form of Chinese soft power. Many leading Chinese models are open-source, meaning anyone around the globe can download them without charge and generally use them with few restrictions. The new thinking, according to the people familiar: Not all technology should be open to everyone.
American companies have increasingly turned to Chinese AI models as lower-cost alternatives to products from OpenAI and Anthropic. Across Silicon Valley, models made by companies such as DeepSeek and Moonshot AI have become core to daily work at companies large and small, the Journal reported.
Andy Fang, co-founder of DoorDash, said in an X post Tuesday that his company routed the most complex tasks to Anthropic’s cutting-edge model Fable while delegating lower-level workloads to Kimi K2.6 from China’s Moonshot AI. DoorDash said the approach delivered better performance at a lower cost than using two Anthropic models.
Staff at legal-services platform Harvey typically turn to top models from OpenAI, Anthropic and Google for high-stakes tasks, said Niko Grupen, head of applied research at Harvey. For simpler work, lower-cost alternatives from Chinese companies such as DeepSeek and Zhipu have become the go-to models.
Startup Vercel said DeepSeek’s share of AI usage rose to 23% in June from 1% in April on the Vercel platform, while DeepSeek’s share of AI spending stayed in the low single digits.
Among the measures under discussion, according to the people familiar, is a more stringent regulatory review before labs can roll out their models. Companies seeking to introduce generative AI services currently must clear a vetting process by the government that focuses on safety and preventing abuse. Officials are considering requiring labs to defer public releases and restrict access by certain tiers of users, such as foreign entities, if a review determines their products contain sensitive technology, the people said.
China is also looking at tighter restrictions on exporting some AI technologies and on Chinese AI companies’ accepting foreign investment, some of the people said. Officials have consulted recently with Chinese companies and researchers to identify areas where China has developed a technology edge.
Beijing has been closely watching Washington’s moves to regulate Anthropic’s Mythos, which is capable of detecting cybersecurity flaws automatically. The White House banned foreign access to the model, prompting Anthropic to cut off access to all users. More recently, the White House moved to allow access for some users. To Beijing’s regulators, the back-and-forth in Washington has reinforced the idea that governments need to keep a tight grip on powerful AI technology to prevent misuse in areas such as cyberwarfare and bioweapons development, the Journal reported.
Industry officials on both sides of the Pacific generally agree that China’s top models trail the best the U.S. can offer, the Journal reported. But Chinese models have gained traction in the U.S. by offering good-enough capabilities at a fraction of the cost.
Any moves by China to tighten its grip on domestic models could risk alienating foreign users and slow global adoption, industry participants said.
The discussions are still in the early stages, the people familiar said. Reuters earlier reported some exchanges between Chinese officials and companies.