Microsoft is switching off rented AI computing capacity from Oracle after deciding not to renew a 2023 agreement, Bloomberg reported on Friday. The servers ran older Nvidia chips that no longer matched Microsoft needs. The move matters because Microsoft is reassessing more than $80B in rental contracts with outside cloud providers.
Why Microsoft let the Oracle contract lapse
The agreement, announced in 2023, gave Microsoft access to Oracle servers to run its AI-powered Bing search. Bloomberg could not determine its full size. In recent months Microsoft has been switching off that rented capacity rather than extending it. Separate talks to lease additional Oracle capacity collapsed in June, according to Business Insider reporting cited in the coverage.
The stated reason is hardware age. The rented systems used an older generation of Nvidia chips, and people familiar with the matter said they no longer fit Microsoft requirements. Microsoft declined to comment on the decision. Oracle said demand for its AI infrastructure remains strong, including for older chips, even as its shares fell about 3% after the report.
The episode fits a wider concern about the pace of chip turnover. Gartner has warned that much AI infrastructure spending could face large write-downs as newer chips arrive. Ageing hardware loses value quickly when next-generation accelerators change performance and efficiency expectations. For operators, capacity signed two or three years ago can become a cost burden rather than an advantage.
What the pullback means for AI capacity buyers
Microsoft has committed more than $80B to rent capacity from other cloud firms such as CoreWeave, Nscale and Nebius. Its leaders have grown frustrated with costs and delays under those arrangements, Bloomberg said. This week Microsoft reportedly turned down an option to expand a Nscale deal in Texas, weeks after pulling out of a Nscale site in the United Kingdom. For enterprise buyers, this signals tighter scrutiny of rented AI capacity on price, delivery dates and chip generation.
Smaller firms feel the effect through availability and contract terms, while large firms face portfolio decisions. A small company renting a few dozen accelerators needs guarantees on chip type, replacement timelines and exit clauses. A large platform with multi-billion commitments must compare outside rental against building its own data centers. Microsoft pattern of pausing expansions shows that even signed options are now conditional on economics and readiness.
Several points still need verification before drawing broad conclusions. The full size and pricing of the Oracle deal were not disclosed, so its direct financial effect is unclear. Oracle position on demand for older chips contrasts with Microsoft assessment, and neither side detailed utilization rates. Buyers should ask vendors which Nvidia generation is supplied, how upgrades are handled, what delays trigger compensation, and whether renewal depends on performance thresholds. This report alone does not show a retreat from outside cloud use.
A concrete marker to watch is what happens to the remaining outside rental portfolio. Further non-renewals, pauses in Texas or the United Kingdom, or new terms with CoreWeave, Nscale and Nebius would confirm a sustained shift. Continued expansion would point to selective pruning instead. Oracle financial indicators, including free cash flow, credit standing and share price after a fall of more than half since a June peak, will show how the market prices older AI capacity.
