Microsoft's China AI business exposes the weak spot in Washington's export-control strategy: you can restrict chips and still leave powerful models moving through cloud channels.
Microsoft sits in the middle of a problem Washington hasn't solved. The US has spent years trying to slow China's access to advanced AI hardware, but American models can still reach Chinese customers through cloud arrangements that look far less dramatic than a shipment of Nvidia chips.
Start with TikTok. As The Verge summarized reporting from The Information in July 2024, TikTok was paying Microsoft almost $20 million a month for access to OpenAI's models, making it one of Microsoft's largest cloud AI customers at the time. ByteDance, TikTok's Beijing-based parent, had already drawn scrutiny after The Verge reported in December 2023 that it had used OpenAI's technology to help build its own large language model. OpenAI suspended ByteDance's account while it investigated a possible breach of its terms.
You don't need to be an export-control lawyer to see the tension. The Guardian reported that OpenAI moved to block API access from China from July 9, 2024, a decision that sent Chinese developers toward domestic alternatives from SenseTime, Baidu, Tencent Cloud and Zhipu AI. Yet Microsoft Azure has been the commercial wrapper around much of OpenAI's enterprise distribution, and Azure's China services operate through 21Vianet, the local partner that runs Microsoft's public cloud inside the country.
That structure isn't new. It has been part of Microsoft's China setup for years. The harder question is whether policy has caught up with what cloud distribution now means.
Washington has mostly focused on the metal. The Biden administration restricted exports of Nvidia's A100 and H100 chips to China, and the Trump administration has kept advanced AI hardware near the center of the US-China technology fight. The logic is straightforward enough: if Chinese companies can't get the most capable GPUs, they have a harder time training frontier systems. But compute is only one door. If a Chinese customer can buy access to a leading model through an approved enterprise cloud path, the policy starts to look half-built.
Microsoft hasn't acted like a company retreating from AI infrastructure. Business Insider reported in April that Microsoft projected $190 billion in capital expenditures for calendar 2026, largely to expand data-center capacity. CFO Amy Hood told investors the company remained confident in the return on those investments, even as near-term cloud growth would still be constrained by supply and power. This is the part you should watch: the same company helping Washington's favorite AI lab scale is also trying to serve as much global demand as its infrastructure can hold.
Frankly, that's what Microsoft is supposed to do as a public company. It sells software and cloud capacity wherever the law lets it sell them. The awkwardness belongs to the policymakers who treat export controls as if the AI stack stops at chips.
The OpenAI relationship has also changed under Microsoft's feet. On April 27, 2026, Windows Central and Tom's Hardware reported that Microsoft and OpenAI ended their exclusive cloud arrangement, while Microsoft kept a non-exclusive license to OpenAI models through 2032 and a revenue-share relationship through 2030. A day later, Axios noted that the revised deal opened the door for OpenAI to sell through other clouds, including Amazon and Google. TechRadar reported on April 29 that OpenAI models, including GPT-5.5, were moving into Amazon Bedrock.
That undercuts one of Microsoft's cleanest strategic advantages. Azure may still be OpenAI's primary cloud partner, but it is no longer the only serious route to OpenAI distribution. If you're Microsoft, that makes every existing reseller channel more valuable in the near term and less defensible over the long term. You harvest while you still have the channel. You build your own models while you still have the cash.
DeepSeek shows the other side of the same trade. The Verge reported in January 2025 that Microsoft made DeepSeek's R1 model available through Azure AI Foundry and GitHub after safety and security reviews. The model mattered because it offered strong reasoning performance at a fraction of the training and inference cost that investors had come to expect from US frontier labs. For enterprise buyers, cost has a way of making geopolitics feel abstract very quickly.
So Microsoft is not just a conduit for American AI into China. It is also willing to package Chinese AI for enterprise developers when the model is useful, cheap and can be wrapped in Microsoft's controls. That's not hypocrisy. It's platform behavior.
The policy problem is sharper than the corporate one. If US officials believe model access itself creates national-security exposure, chip controls won't be enough. If they don't believe that, then the public anxiety around Chinese access to American AI is overstated. What doesn't work is pretending the risky part of AI trade lives only in GPU shipping manifests while the model layer moves through cloud contracts, partner regions and enterprise accounts.
Microsoft has found the rational commercial position: stay legal, keep selling, diversify the model catalog, and avoid being trapped by one partner or one country. Washington still hasn't found the matching policy position. Until it does, Microsoft will keep doing what large software companies do best, turning every gray zone into a revenue line.
Also read: Nuclear power is becoming the defining infrastructure bet behind the AI buildout • Enterprise AI is entering its ROI reckoning and the startups that survive will have the numbers to prove it • Epic Games is handing Unreal Engine 6 the keys to generative AI and the model that wins the default slot wins the industry
