Two frontier AI model developers, OpenAI and Google, have reportedly been supplying Chinese tech giants in violation of a Pentagon blacklist.
Both companies have been selling access to their AI models to Singaporean subsidiaries of Alibaba, Baidu, and Tencent, three of China’s leading technology companies. It suggests that US government efforts to slow China’s AI development are not being effectively enforced, and that stronger legislation than a simple blacklist may be needed to prevent Chinese companies from accessing these services.
OpenAI told the Financial Times, which first reported the apparent infringement, that it had suspended API access for users affiliated with Alibaba. Anthropic also suspended Alibaba last month, accusing the company of distilling its Claude model using 28 million responses.
Google has said that its AI services are available in both Hong Kong and Singapore, but that the company prohibits distillation, a process in which a smaller model is trained to replicate the outputs of a larger, more advanced model.
How Chinese AI companies benefit from US models
Alibaba, Baidu, and Tencent are all investing billions in AI model training and inference, but the costs for consumers and their own bottom lines are much lower than those faced by their US counterparts. Anthropic, OpenAI, and others have said this is partly because Chinese AI model developers routinely distill US models, reducing the cost of training their own systems.
While these Chinese companies are actively using and learning from US models, Chinese consumers cannot access most of them because they are blocked by the country’s Great Firewall. This has left many of China’s major technology companies facing less foreign competition, allowing TikTok developer ByteDance, Alibaba, and DeepSeek to establish themselves as market leaders.
The Chinese government has also considered preventing outsiders from accessing AI models developed by Chinese companies. This comes in response to developers increasingly shifting from Opus, GPT, and Gemini during the development stage to cheaper Chinese models for day-to-day operations.
US government taking some tips from China
Across the Pacific, the US government is also tightening its grip on domestic AI model developers. The Trump administration recently halted exports of Anthropic’s two most powerful AI models, Fable 5 and Mythos 5, prompting the company to shut them down. It has since been permitted to distribute them in select countries.
OpenAI received similar treatment with the launch of GPT-5.6, with the US government asking it to stagger the release through a verified group of partners before making it more widely available. This has reportedly led to internal discussions at OpenAI about giving the US government a small stake in the company when it goes public, potentially helping to accelerate future approval processes.
For businesses looking to deploy AI, the potential escalation of restrictions could increase the costs of training and deployment, as cheaper models may become inaccessible in the future. It could also harm other businesses in countries with murky allegiances to the US or China, which may face restrictions to prevent the other from gaining access.
Read more: OpenAI’s new ChatGPT Work agent can create documents, analyze workplace data, and complete multi-step business tasks, but its access to connected systems raises new questions about oversight and security.