The cloud-API loophole

OpenAI won't sell its models in China, citing theft and misuse. ByteDance, Tencent, Ant Group and Meituan use them anyway — through Microsoft's Azure cloud, which reaches GPT-4 and GPT-4o as a metered service. A leaked transcript of a July 2025 Microsoft sales meeting, reported by Bloomberg, puts numbers on a hole everyone suspected: Azure's AI revenue in China roughly tripled in the year to June 2025 after quadrupling the year before, making China the company's fastest-growing AI territory, with ByteDance alone on pace to spend more than a billion dollars a year.

Microsoft says it sells only to established firms and monitors usage — yet sources said Chinese customers face no heightened oversight.

The whole thing turns on a definition. US export controls block the physical chips that train and run frontier models, but renting a model over an API isn't an export: the customer never takes possession of the weights or the code, only the answers that come back. So the same regime that stops a Chinese firm buying the hardware lets it pay for the output by the token. The figures grow off a small base — China is about 1.5% of Microsoft's revenue — but the direction is the point.

The tell is geography. These models don't run in Microsoft's Beijing or Shanghai regions; they run offshore, in Singapore and Hong Kong data centres, because Microsoft fears its partner's intellectual property would leak on the mainland. Microsoft trusts China enough to sell to it, not enough to host there. And OpenAI sits on both sides at once — refusing China directly while profiting as its cloud partner serves it, even as it privately presses Microsoft to stop Chinese firms distilling its models. The number China hawks lacked now has a name attached to it, which is what makes a quiet loophole into a policy fight.