Oracle troubles rise as Microsoft scraps deal; software giant rethinks cloud deals

Microsoft is winding down a partnership with Oracle that supplied rented computing capacity for Microsoft's AI-powered Bing search engine. The agreement began in 2023, but Microsoft has chosen not to renew it. The decision shows how quickly AI infrastructure needs can change as newer chips become available.

The rented machines were hosted by Oracle and used older generations of Nvidia chips. Bloomberg reported that Microsoft had steadily deactivated this server capacity in recent months. The article does not disclose the contract's total value, and Bloomberg could not determine its full scale.

Microsoft is now reviewing more than $80 billion in outside cloud commitments. Executives have reportedly become unhappy with rising costs and delays in third-party arrangements. Oracle still says demand for its AI infrastructure remains strong, including for older hardware, but losing Microsoft's usage reduces capacity utilization and adds pressure to its finances.

Cloud computing means using computing resources hosted in remote data centers through a provider. Instead of buying and operating every server, a company pays to use processing power, storage, and related systems when needed. AI workloads require especially large amounts of computing capacity, making cloud access valuable.

Microsoft can rent capacity from Oracle or Nscale when its own facilities cannot supply enough suitable machines. This approach can provide additional Nvidia chips and data-center space without waiting to build everything internally. The Oracle arrangement supported Microsoft's AI-powered Bing search engine, while Nscale agreed to deploy about 200,000 Nvidia AI chips across facilities in Europe and the United States.

Renting also creates trade-offs. Microsoft is reviewing more than $80 billion in outside cloud commitments because executives are reportedly dissatisfied with escalating costs and operational delays. The company is therefore reassessing how much capacity to obtain, from whom, and under what terms.

Microsoft's outside cloud commitments exceed $80 billion, according to Bloomberg's report. That figure covers arrangements with multiple providers, including CoreWeave, Nscale, and Nebius Group. Reviewing such a portfolio shows that AI capacity is not a small technology purchase; it is a major infrastructure and financial decision.

Nscale agreed in 2025 to deploy around 200,000 Nvidia AI chips under an expanded Microsoft agreement covering data centers in Europe and the United States. A chip deployment of that size requires large facilities, power supplies, construction programs, and financing. It also shows why Microsoft needs more than one source of computing capacity.

The scale does not mean every chip is operating at once or that every planned expansion will happen. Microsoft recently declined an option to more than double its Texas arrangement with Nscale. It also withdrew from a planned Nscale site in the United Kingdom, reflecting the difficulty of matching demand with timely, economical capacity.

If Microsoft and other customers stop using or renewing rented servers, Oracle has fewer paid workloads running on its infrastructure. Lower utilization can make expensive data centers less productive and weaken the returns on the chips, buildings, and operating systems Oracle has funded. It can also make future expansion harder to justify.

The article links Oracle's AI data-center construction spending to negative free cash flow. Its credit rating is only one level above junk status, while its stock has fallen by more than half since reaching a June high. Microsoft's non-renewal therefore adds pressure to an already difficult financial position, although the contract's value was not disclosed.

Oracle says customer demand remains robust, including for earlier-generation hardware. That claim suggests the effect is not that all older equipment becomes useless immediately. However, losing capacity utilization from a major customer can still compound financial headwinds while Oracle continues building AI infrastructure.

A working chip is not automatically a competitive chip. If newer Nvidia generations provide more computing power or better efficiency, customers may prefer them for demanding AI workloads. Older chips can then earn less because they require more equipment, energy, or time to produce the same result, even though they remain operational.

The Oracle-hosted machines used older generations of Nvidia chips. Microsoft decided that this hardware no longer aligned with its current computing requirements and steadily deactivated the rented capacity. Oracle says some customers still want earlier-generation hardware, so obsolescence is a matter of commercial value, not instant failure.

Rapid improvement creates accounting risk. Companies may spend heavily on buildings and equipment expecting years of strong demand. If newer hardware arrives sooner, the older assets may generate less income than forecast. Gartner has warned that aggressive AI infrastructure spending could lead to significant write-downs as faster, more efficient chips enter the market.

A mixed strategy spreads Microsoft's dependence across its own data centers and several outside providers. That can help the company obtain capacity from different locations and hardware fleets. It also gives Microsoft more freedom to adjust when a provider becomes expensive, delayed, or technically unsuitable.

The article describes commitments involving CoreWeave, Nscale, and Nebius Group, alongside Oracle. Microsoft is reviewing more than $80 billion in outside cloud commitments, while ending the Oracle arrangement and declining to expand an Nscale deployment in Texas. These changes show the company actively comparing providers rather than treating every contract as permanent.

The trade-off is complexity. Managing multiple agreements can create coordination challenges and uneven availability. Still, recent problems make diversification attractive: Microsoft reportedly faced escalating expenses and operational delays, pulled out of an Nscale project in the United Kingdom, and did not exercise a Texas expansion option.

AI data centers need several inputs at the same time. Providers require large chip supplies, reliable electricity, suitable buildings, construction work, and financing for expensive projects. A shortage or delay in any one of these can leave purchased equipment unused or postpone promised computing capacity.

The Nscale example makes the mechanism clear. Microsoft withdrew from a planned site in Loughton, Essex, after power connection delays pushed its expected supply date into the 2030s. It also declined an option to more than double an existing Texas agreement. These decisions show that contracted demand does not automatically become available capacity.

Financing adds pressure because construction and equipment require major upfront spending. Oracle's substantial AI data-center outlays have pushed its free cash flow negative, while its credit rating sits one level above junk status. Meanwhile, rapidly changing chips can reduce the value of earlier investments, making expansion both slower and financially riskier.