Microsoft has been quietly pulling back the AI computing capacity it rents from Oracle — though "quietly" is maybe the wrong word, since the story got out on November 7 via Wallstreetcn and Cailianshe. People familiar with the matter say Redmond has spent recent months switching off Oracle-leased capacity, having decided not to renew, as part of a broader re-evaluation of its cloud partnerships.
The chip generation is the crux. Oracle's hardware runs on older-generation Nvidia chips, which no longer match what Microsoft needs and have lost value accordingly, the same people said. Oracle, for its part, insists demand for OCI's AI infrastructure remains strong — including for earlier generations of GPUs.
Nor is Oracle the only one feeling the chill. A day before the report surfaced, AI cloud operator IREN said part of its infrastructure — capacity originally due online by year-end, capable of supporting more than $4 billion in annualized revenue — had been pushed to January after Microsoft asked for network design changes. It could slide further, possibly to late February. Microsoft is also tightening oversight of projects run by neocloud providers like IREN, Nebius and Nscale amid delays and rising costs. CoreWeave's shares fell sharply on the news.
Some history here. The Oracle arrangement goes back to 2023, when the two companies signed a multi-year agreement for Microsoft to use OCI's AI infrastructure to power the AI-assisted version of Bing. The financial terms were never disclosed.
One analysis put it plainly enough. GPUs age, and the world's largest renter of computing capacity no longer wants to pay for old silicon.