Microsoft is winding down its arrangement to rent AI computing from Oracle, Bloomberg reported on Friday — and Microsoft has already been switching the capacity off for months. The software giant declined to renew the contract, part of a larger reassessment of billions of dollars in cloud partnerships.
The reason, according to people familiar with the matter, is mundane in a way that should worry every GPU landlord: the Oracle hardware ran an older generation of Nvidia chips that no longer met Microsoft's needs. The 2023 deal gave Microsoft extra compute for AI-assisted Bing; its terms were never disclosed. Three years later, the chips that were good enough then are the ones Microsoft is walking away from.
This is not an isolated breakup. A day earlier, Bloomberg reported Microsoft was tightening oversight of data center projects run by neocloud providers IREN, Nebius, and Nscale amid delays and rising costs. IREN said Thursday it had delayed some infrastructure deliveries after Microsoft requested network design changes, pushing its $4 billion capacity target from year-end to January — possibly late February. CNBC reported that Microsoft is shrinking its roster of specialist neocloud partners and has not yet decided which ones it will stop using. Microsoft declined to comment; a spokesperson told CNBC only that it regularly reviews these investments against customer demand.
Oracle is pushing back. "Demand for OCI's AI infrastructure remains strong, including for earlier-generation GPUs," a spokesperson told Reuters, and the company said in September that contracts for four-year-old GPUs were renewing at a 20% premium. The market's first reaction was harsher than Oracle's line: shares fell as much as 5.2% after the report before recovering to roughly flat by mid-afternoon.
The timing is what makes this more than a contract dispute. Two days ago the Nasdaq entered correction and the AI trade had its worst day in 19 months. Four days ago Elon Musk warned 200 neoclouds were about to starve after SpaceX borrowed $40 billion for GPUs. The pattern is now visible: the companies that spent 2024 and 2025 renting any rack they could find are now auditing every one of them — and the ones with yesterday's silicon are the first to get cut. What to watch: which neoclouds Microsoft drops outright, and whether Oracle's premium renewals hold now that its biggest showcase customer has left. The AI boom isn't ending — it's just stopped paying for old hardware.
— via Reuters