For the past year, the pitch from Chinese AI labs has been simple. Take our models. They’re free, they’re open-weight, they’re nearly as good as what Anthropic and OpenAI sell, and they cost a fraction as much. American companies listened. Airbnb built its customer service chatbot on Alibaba’s Qwen. Cursor built Composer 2 on Moonshot’s Kimi. Nvidia’s Jensen Huang stood on a stage in Beijing and called DeepSeek, Alibaba and MiniMax “world class.” Lindy moved 100% of its traffic off Claude to DeepSeek and watched its cost curve crash.Now Beijing is thinking about taking the offer back. Chinese authorities have held meetings with Alibaba, ByteDance and Z.ai over the past month about restricting overseas access to the country’s most advanced AI models, including ones not yet released, Reuters reported, citing three people familiar with the discussions. The talks were led by the Ministry of Commerce. Officials floated making the leak or theft of proprietary AI technology an offence under China’s national security law, and raised the idea of restricting who can fund domestic AI startups. Nothing is decided. But the direction is unmistakable.
Beijing’s silicon curtain: why China is rethinking free AI models
The logic here is the same logic Washington has been running for months. Beijing has been watching the fight over Anthropic’s Mythos, a model built to automatically find software vulnerabilities. In June, the White House barred foreign nationals from accessing Mythos and Fable, and Anthropic responded by pulling both offline for everyone. Fable came back after new safeguards. Mythos is still limited to a handful of trusted US organisations.That episode landed hard in China. Two of Reuters’ sources said Chinese authorities are worried Washington could point Mythos at Chinese systems. Zhou Hongyi, founder of cybersecurity firm 360, has publicly said China needs a Mythos of its own. State media has said much the same.Which puts Beijing in an odd spot. Open-sourcing was the strategy. It was soft power. It was how Chinese models got into American workflows in the first place, and it worked spectacularly.
Alibaba Qwen, DeepSeek and GLM: how Chinese AI models won over Silicon Valley
The numbers are not subtle. Chinese models have sat above 30% of tokens used by US companies on OpenRouter each week since February, peaking at 46%. The 12-month average before that was 11%. Vercel says DeepSeek’s usage share on its platform jumped from 1% in April to 23% in June, while DeepSeek’s share of AI spending stayed in the low single digits. That gap is the whole story.Z.ai’s GLM-5.2, released in June, landed within a percentage point of Anthropic’s Opus 4.8 on a closely watched agentic benchmark at roughly a fifth of the cost. DoorDash routes hard tasks to Anthropic’s Fable and cheap ones to Moonshot’s Kimi K2.6. Harvey does the same split. Alibaba’s Qwen crossed 700 million Hugging Face downloads by January.
What a China AI export control would mean for US companies and costs
If restrictions land, they will likely only apply to future models. Weights already published cannot be recalled. So Airbnb’s chatbot keeps running. Cursor keeps shipping.The pain is prospective. It means no next GLM, no next Kimi, no next cheap escape hatch when Anthropic and OpenAI raise token prices again.And it lands on companies already sandwiched. US House committees are probing Airbnb and Cursor over their Chinese model use. Beijing has told firms to drop Claude Code, flagging a “security back-door vulnerability.” Both capitals now treat the other’s AI as a liability.
Leave a Reply