Microsoft walked away from talks to lease cloud capacity from Oracle for roughly $3 billion, and according to media reports the reason wasn’t price but technical — Oracle didn’t meet FedRAMP certification, the mandatory security framework for hosting U.S. government data. When Oracle $ORCL hesitated over the massive engineering burden and the time needed for quick fixes, Microsoft $MSFT left the table. Oracle is downplaying it and says the details in the article are inaccurate and that the two companies continue to maintain a cooperative relationship.

For me, the more important story is what’s behind this. Demand for AI capacity is growing faster than datacenters can be built, and even giants like Microsoft are being forced to negotiate sharing infrastructure with competitors — and are now discovering how fragile such deals are. At the same time Microsoft is pursuing its own aggressive expansion and plans an unprecedented $190 billion in capital expenditures for fiscal 2026. This is another signal that the bottleneck for AI is no longer the models but physical capacity, power, and compliance. Whoever can build and certify their own infrastructure has a long-term advantage — and Microsoft is clearly betting it’s better off relying on itself.