Microsoft Winds Down Its Oracle AI Cloud Deal and Puts Outside GPU Suppliers Under Review

Microsoft has begun winding down an agreement to rent AI computing capacity from Oracle, according to a November 6 Bloomberg report citing people familiar with the matter, after deciding not to renew a contract that involved older-generation Nvidia chips that no longer met its requirements. The rented Oracle resources have been switched off in recent months as Microsoft's technical specifications moved on.

The arrangement is old by AI-infrastructure standards: it dates to 2023, when Oracle agreed to supply servers that supported Microsoft's AI-powered Bing search engine. Bloomberg could not determine the full size of the contract, and Microsoft declined to comment. Oracle did not say which customers absorbed the capacity Microsoft walked away from.

The retreat fits into a broader rethink. According to Bloomberg, Microsoft has committed more than $80 billion to renting capacity from outside cloud providers, including CoreWeave, Nscale and Nebius, to support its AI operations. Several executives have since grown frustrated with high costs and delivery delays, and the company has tightened oversight of data center construction, withdrawn from projects it considers risky, and discussed whether to sign further agreements with these providers. A separate November 5 Bloomberg report said Microsoft had even sent employees to supplier sites to check construction standards and compliance, with facilities built by Nebius, IREN and Nscale receiving closer attention.

CNBC reported the same day that Microsoft is evaluating a smaller group of outside AI-computing suppliers, known as neoclouds, as it prioritizes its own data center footprint. No final decision has been made on which providers could lose business. The review is not a spending pullback: Microsoft reported $45.2 billion in capital expenditures and finance leases for its latest quarter, and finance chief Amy Hood said in October that spending would accelerate in the second half of the fiscal year even as some GPU-related contracts were delayed into the first half of 2027, weighing on Azure growth.

Several supplier decisions have already landed. Two people told CNBC that Microsoft opted not to renew a deal with Oracle. Microsoft previously walked away from a planned Nscale site in England, which its UK and Ireland chief Darren Hardman attributed in October to difficulties securing electricity, and it has declined an option to expand its existing Texas agreement with Nscale, according to Reuters.

Oracle, for its part, pushed back publicly. In a November 6 statement on X, the company said demand for OCI's AI infrastructure remains strong, including for earlier-generation GPUs, and pointed to its latest earnings call: GPU capacity that came up for renewal in its fiscal first quarter was renewed or resold at prices 20 percent above prior agreements under long-term contracts, and most of those GPUs were at least four years old. The company framed that pricing as evidence its existing hardware still attracts customers even as newer chips arrive.

The stakes for the neoclouds are enormous. CoreWeave disclosed when it went public that Microsoft accounted for 62 percent of its 2024 revenue. CoreWeave, Nebius and IREN have all lost roughly half their market value over the past month. IREN, which delivered data center capacity to Microsoft in August, said a customer-requested networking upgrade has pushed the expected start of operations into January or February, and moved its target for more than $4 billion in annual recurring revenue from December into early 2027. IREN co-CEO Daniel Roberts told analysts that capital markets had become more selective, shifting attention from contract size to whether projects can be delivered, financed and operated profitably.

What is changing is not whether hyperscalers spend, but whom they pay. A large contract is no longer the prize; proving a supplier can secure power, deliver equipment on schedule and bring capacity online is what keeps it. For a generation of GPU landlords built almost entirely on hyperscaler demand, that is a much harder test to pass.