Microsoft quietly becomes China's gateway to OpenAI and the route runs through Singapore

ByteDance alone is on track to spend more than US$1 billion a year on Azure AI, even as OpenAI and Anthropic refuse to sell into China directly, leaving Microsoft to broker the models its own partners won't.

Microsoft has done what its closest AI partners will not: sell American frontier models into China. And for the region's channel, the most telling detail is where those transactions physically land. Not in Beijing or Shanghai, but in offshore data centres reached over the public internet from hubs that include Singapore.

According to Bloomberg, which first reported the arrangement this week, Microsoft has built a substantial business supplying AI models to Chinese technology firms through its Azure cloud, even as Washington and Beijing escalate their contest over the technology. ByteDance has been its largest AI customer in recent years, running largely on OpenAI models, and is on track to spend upward of US$1 billion a year on Microsoft's AI and cloud services, people familiar with the matter told the news agency.

Ant Group, Meituan and Tencent Holdings are also named as significant buyers, though Ant says it builds its own models and that its core products do not depend on external ones. The growth curve has been steep. Azure's AI revenue in China expanded faster than in any other Microsoft territory, roughly tripling in the financial year to June 2025 after surging about 400% the year before, then-chief commercial officer Judson Althoff told staff at a July 2025 sales meeting, per a transcript reviewed by Bloomberg.

Althoff cast Microsoft as the one company "bringing those two places together," referring to the AI hubs on the US west coast and China's east. For scale, president Brad Smith has told US lawmakers the China operation accounted for roughly 1.5% of group revenue in 2024. What makes the position unusual is structural.

Microsoft's singular partnership with OpenAI lets it set its own terms for selling GPT-series models in China, a market OpenAI and Anthropic have stayed out of entirely, citing intellectual-property theft and misuse risks. Anthropic's models are pointedly absent from Microsoft's China line-up. In effect, Microsoft operates as the middleman for models their makers refuse to sell directly.

For partners and customers across Asia, the plumbing is the point. Microsoft runs Azure regions near Beijing and Shanghai, but under its OpenAI agreements it does not host those models on Chinese soil, wary of IP leakage. Customers instead access them across the internet from facilities elsewhere, Singapore among them, placing the city-state's cloud capacity squarely inside the China-access supply chain.

The catalogue is widening in ways that cut both directions. Microsoft has carried DeepSeek's models on Azure since early 2025 and, by mid-2026, was reported to be trialling DeepSeek-V4 for its own Copilot tooling as inference costs from US labs climbed, a sign the platform's commercial logic increasingly favours whichever model is cheapest, wherever it is built.

That openness sits awkwardly against the risks. OpenAI has privately pressed Microsoft to do more to stop Chinese firms "distilling" its models, Bloomberg reported. Microsoft points to automated monitoring and a policy of selling only to established companies, not individual developers, as guardrails, yet sources told the agency that Chinese customers face no heightened oversight, and synthetic-data training is hard to police.

The takeaway is a familiar one sharpened by AI: the most consequential US-China technology flows increasingly run through cloud contracts and regional data centres, and the next policy shift on model exports could redraw that map overnight.