Microsoft is quietly building a massive AI business in China by reselling OpenAI’s models to the country’s biggest tech firms. According to Bloomberg, ByteDance has become Microsoft’s single largest AI customer, with annual spending on Microsoft AI and cloud services expected to exceed $1 billion. Ant Group, Meituan, and Tencent are also buying big through Azure. Much of this spending goes toward supporting their overseas expansion — and conveniently sidesteps restrictions on direct sales of frontier AI models to China.
The revenue from Microsoft’s China AI business is growing at breakneck speed. Former chief commercial officer Judson Althoff revealed in a July 2025 internal meeting that Azure’s AI revenue in China is growing faster than anywhere else in the world — after a 400% surge in fiscal 2024, fiscal 2025 revenue nearly tripled. Althoff described it as building a bridge between the US West Coast and China’s East Coast, with Microsoft as the connector. That’s quite the contrast to the public posture of both the US government and OpenAI, who have been vocal about guarding against China’s AI rise.
The deals rely on a carefully constructed workaround. Since OpenAI and Anthropic don’t sell models directly to China over security concerns, Microsoft uses its special agreement with OpenAI to set its own resale policy bypassing the ban. To protect intellectual property, Microsoft doesn’t deploy GPT models in local data centers in Beijing or Shanghai. Instead, Chinese customers access the models over the internet from servers in Singapore and other overseas nodes.
But that Singapore routing hasn’t eased OpenAI’s worries. While Microsoft uses automated monitoring to prevent customers from building competing products, it hasn’t applied any extra surveillance specifically for Chinese clients. OpenAI has privately complained to Microsoft, accusing the company of not doing enough to stop Chinese tech giants from using distillation — training their own models on outputs from OpenAI’s models — to boost their capabilities. In practice, distillation using synthetic data is notoriously difficult to block at the technical level.