Microsoft Sells OpenAI Models to ByteDance While OpenAI Stays Out

Published by James Harris on June 18, 2026June 18, 2026

Microsoft is selling OpenAI’s GPT models to ByteDance, Ant Group, Tencent, and Meituan through Azure, generating AI revenue in China that roughly tripled in the fiscal year ending June 2025, following a 400% surge the year before, all while OpenAI itself refuses to operate there.

The arrangement works because Microsoft’s contract with OpenAI gives it independent authority to set terms for selling GPT models abroad, a carve-out that lets Microsoft occupy a position no other Western company holds. Chinese customers access the models through data centers in Singapore and nearby countries rather than servers on Chinese soil, which is Microsoft’s primary technical safeguard. ByteDance alone is on track to spend over $1 billion annually on these services. OpenAI’s core concern is distillation: Chinese firms feeding GPT outputs into smaller models to replicate capabilities without paying for the underlying research, a technique that is genuinely difficult to police because synthetic training data is hard to distinguish from original work. The rapid growth in Azure’s China AI segment suggests that whatever monitoring Microsoft runs has not meaningfully slowed demand.

Despite the headline numbers, Brad Smith told U.S. lawmakers this business represents roughly 1.5% of Microsoft’s total 2024 revenue. That fraction matters because it calibrates how much political exposure Microsoft is absorbing for a relatively contained financial upside.

The pressure is building from multiple directions simultaneously. Anthropic’s models are excluded from Microsoft’s China offerings entirely, and U.S. Commerce Secretary Howard Lutnick ordered Anthropic to suspend exports of its Mythos and Fable models worldwide over concerns about military intelligence users in China and Russia. Goldman Sachs and JPMorgan Chase both removed Claude from approved tools for Hong Kong staff in 2025. While American AI companies navigate these restrictions, DeepSeek and comparable Chinese models are pulling U.S.-based developers in the opposite direction: an hour of coding on Claude costs roughly $10 versus under 50 cents on DeepSeek, a price gap that makes the competitive anxiety behind OpenAI’s distillation concerns feel less abstract.

What this arrangement actually demonstrates is that the US-China AI divide is less a wall than a membrane, and Microsoft is currently the most profitable gland in it. That position becomes harder to defend as Washington’s export control apparatus tightens and OpenAI’s own commercial interests grow more distinct from Microsoft’s. The question is not whether scrutiny of this arrangement increases, but how quickly the two companies’ incentives diverge when it does.