OpenAI loosens Microsoft ties

OpenAI’s real business is no longer just models. It’s compute — huge amounts of it, locked in for years, financed somehow, and delivered fast enough to keep ChatGPT and future models running. That is why the Microsoft relationship mattered so much. It bundled money, cloud capacity, product distribution, and a tight grip on where OpenAI could run. In late April, that bundle got looser. OpenAI and Microsoft changed the deal, and the change is bigger than a normal contract tweak. (blogs.microsoft.com) ### What actually changed? The headline changes are pretty stark. Microsoft’s license to OpenAI’s IP is now non-exclusive. Microsoft will no longer pay a revenue share to OpenAI. OpenAI says it can now serve all of its products across any cloud provider. But Microsoft still isn’t out of the picture — it remains OpenAI’s “primary cloud partner,” and OpenAI products still ship first on Azure unless Microsoft can’t or won’t provide what OpenAI needs. (blogs.microsoft.com) ### Why is cloud freedom the big deal? Because OpenAI has outgrown the idea that one company can neatly host all of its ambition. Training frontier models and serving hundreds of millions of users are different jobs, and both are getting more expensive. The old setup already started to bend in January 2025, when Microsoft and OpenAI moved from hard exclusivity on new capacity to a right-of-first(blogs.microsoft.com)et enough of it. (blogs.microsoft.com) ### Where does Stargate fit in? Stargate is OpenAI’s attempt to secure that compute at infrastructure scale. When it launched on January 21, 2025, the plan was a new company that intended to invest $500 billion over four years in U.S. AI infrastructure, with $100 billion deployed immediately. Since then, OpenAI has kept widening the buildout — first with Abilene, then a 4.5-gi(blogs.microsoft.com) explicit: bring new capacity online faster, with multiple partners, not one captive pipe. (openai.com) ### So is Microsoft losing OpenAI? Not exactly. Microsoft gave up some exclusivity, but it kept a lot. It remains a major shareholder. It still has rights to OpenAI IP through 2032. It still gets revenue share payments from OpenAI through 2030, subject to a cap. And Azure still gets first shot at hosting OpenAI products. This looks less like a breakup and more like a reset — Microsoft stops trying to be the only lane, and instead tries to stay the most important lane. (blogs.microsoft.com) ### Why would Microsoft agree to that? Probably because the alternative was worse. If OpenAI needs more compute than Azure can supply on the right timeline, forcing exclusivity risks slowing OpenAI down and poisoning the partnership. Microsoft seems to have chosen a narrower but more durable position: keep the commercial relationship, keep product priority, keep long-dated IP access, and stop fi(blogs.microsoft.com)rms both sides published. (blogs.microsoft.com) ### What’s the catch for OpenAI? Freedom is expensive. A captive partner can subsidize you because it knows the workloads stay home. A multi-cloud buyer has more optionality, but it also has to coordinate more suppliers, more financing, and more physical buildouts. OpenAI’s own infrastructure posts read like a company racing to secure power, land, chips, and construction capacity all at once. That is not just a technical problem. It is a balance-sheet problem. (openai.com) ### Why does this matter beyond the two companies? Because this is what the AI stack looks like once models stop being the scarce thing. The scarce thing becomes the ability to finance and deliver gigawatts. OpenAI is trying to turn itself from a lab with a patron into a platform with multiple industrial backers. Microsoft is trying to stay central without being exclusive. Everybody else in cloud now has a clearer opening. (openai.com) ### Bottom line? The April rewrite says OpenAI wants less dependence and more capacity. But the trade is simple — less dependence also means less shelter. Microsoft loosened its grip. OpenAI gained room to move. Now it has to prove it can turn that room into actual compute.