Vivold Consulting

Microsoft has become the lone conduit for OpenAI's models in China, profiting where their makers won't go

Key Insights

Microsoft has quietly become the main supplier of OpenAI's GPT models in China, selling them through Azure to the country's biggest internet firms even though OpenAI and Anthropic both refuse to sell there directly on IP and misuse grounds. Per Bloomberg, ByteDance is Microsoft's largest AI customer and is on track to spend over US$1 billion a year. A unique OpenAI contract lets Microsoft set its own overseas terms - and it's simultaneously testing a Chinese DeepSeek model for Western customers, taking margin on both sides of the trade.

Stay Updated

Get the latest insights delivered to your inbox

The China deal OpenAI itself won't make

Microsoft has become the main channel through which OpenAI's models reach China, selling the GPT series via Azure to the country's largest internet companies even as OpenAI and Anthropic keep their own models out of the market on intellectual-property and misuse grounds. The arrangement, detailed this week by Bloomberg, gives Microsoft a position no other US AI vendor holds: it profits from selling models whose own creators won't deal with Chinese firms directly.

The scale

- ByteDance has been Microsoft's largest AI customer, running largely on OpenAI models, and is on track to spend more than US$1 billion a year on Microsoft's AI and cloud services, per Bloomberg's sources.
- Ant Group, Meituan, and Tencent also buy models through Azure, though Ant says its core products rely on its own models.
- Azure's AI revenue in China grew faster than in any other territory - roughly tripling in the year to June 2025 after climbing about 400% the year before, a then-senior Microsoft executive told staff. Microsoft's president has separately told lawmakers China was about 1.5% of 2024 revenue.

Why it all runs through Microsoft

The key is Microsoft's singular contract with OpenAI, which lets it set its own terms for selling GPT models abroad. Both OpenAI and Anthropic have declined to sell into China directly, and Anthropic's models are absent from Microsoft's China line-up entirely - leaving Microsoft as the intermediary for models their makers consider too risky to serve there. Risk is the running tension: OpenAI has privately pressed Microsoft to do more to stop Chinese customers from "distilling" its models, the practice of using one model's outputs to train another. Microsoft points to automated monitoring and a policy of selling only to established companies, and it doesn't host the models on Chinese soil - customers reach them over the internet from data centres elsewhere, including Singapore. Even so, sources say Chinese buyers face no extra scrutiny and synthetic data is hard to police.

Playing both sides

The contradiction sharpens alongside what Microsoft hosts: it added DeepSeek's R1 to Azure in January 2025 and is now testing a fine-tuned, Azure-hosted DeepSeek-V4 as a cheaper option for its Copilot Cowork agent. So Microsoft is selling a Chinese model into Western businesses while selling American models into Chinese ones, earning margin on both legs - a balancing act that's increasingly contentious in Washington, where lawmakers cast China's AI push as a threat and OpenAI's private objections could grow louder.

Related Articles

HSBC's multi-year Google Cloud deal targets 200+ AI use cases, some worth $100M+ each

HSBC has signed a multi-year partnership with Google Cloud to build and deploy AI across wealth management, financial-crime risk, and internal decision support, using Gemini models and the Gemini Enterprise Agent Platform. The bank expects more than 200 AI use cases over two years, with selected ones each potentially returning over US$100 million. It builds on a deep existing base - 600-plus AI use cases and a Google-built financial-crime system screening 1.2 billion transactions a month.

The EU's new AI-labelling code previews Article 50 duties that become law on August 2

The EU has published a voluntary Code of Practice for marking and labelling AI-generated content, a practical playbook for meeting transparency duties that become law on August 2, 2026 under Article 50 of the AI Act. From that date, deepfakes and AI-generated text on public-interest matters must be labelled, and users must be told when they're talking to an AI. The code splits the work - model builders mark output in machine-readable form, deployers handle visible labelling - and companies have under two months to prepare.

Fresh off its IPO, SpaceX is buying AI coding startup Cursor for $60B to rescue its AI division

SpaceX has agreed to acquire AI coding startup Cursor in a $60 billion all-stock deal, just days after its historic IPO and less than two months after the two first tied up. The purchase is meant to revive SpaceX's struggling AI division - built around xAI, which it merged with earlier this year - after a restructuring marred by controversy. SpaceX told IPO investors it sees a $26 trillion addressable AI market; the deal is expected to close in Q3 2026.