Microsoft winds down Oracle AI rental deal despite capacity shortages

Company-reported FY27 Q1 figures, visualized by Magica. Capital expenditure includes finance leases; neither figure measures the Oracle agreement. Source: Microsoft.

Microsoft is still buying and building

Microsoft’s October 28 earnings call supplies an important limit to any interpretation of the Oracle decision as a general infrastructure retreat. The company said customer demand continued to exceed available supply, even after it added 24 data centers across five continents in its fiscal first quarter of 2027.

Microsoft reported $45.2 billion in capital expenditure for that quarter, with roughly three quarters going to short-lived assets, primarily CPUs and GPUs, including higher memory and storage costs. The total includes finance leases and is distinct from the $39.4 billion of cash paid for property, plant and equipment. These companywide figures establish the scale of investment; they do not measure the Oracle contract or the savings from ending it.

The company also said some third-party capacity deliveries expected in that quarter had slipped. In its outlook, it said some third-party GPU service contracts had been delayed into the fiscal second half. It expected capital expenditure above $50 billion in the second quarter and approximately $170 billion for calendar 2026, with spending accelerating in the fiscal second half partly as delayed capacity arrived. Those are forecasts, not completed spending.

Building more is only one response to the shortage. Microsoft said engineering improvements had freed capacity in its existing GPU fleet and enabled it to monetize underused CPU capacity. It described allocating infrastructure to the highest-value opportunities across its services. Neither those efficiency gains nor the count of new data centers establishes how much Oracle capacity Microsoft can replace.

Outside suppliers remain part of that strategy. The separate account said Microsoft continued spending heavily on third-party AI servers because it could not build its own data centers fast enough. It also reported that Microsoft had recently cancelled a data center deal with cloud provider Nscale and delayed payments to AI cloud company IREN after delivery delays at those providers.

IREN’s difficulties show how infrastructure requirements can affect timing as well as supplier selection. In the account relaying the broader partnership review, IREN said on November 5 that Microsoft had requested network-design changes, delaying some deliveries. IREN moved its target for bringing more than $4 billion in annualized revenue capacity online from year-end to January, with a possible further delay to late February. That is prospective revenue capacity, not revenue already earned or the value of the Oracle agreement.

What Oracle can earn from the released capacity

Oracle says demand for older equipment persists. A spokesperson told the news agency carrying the broader review that demand for OCI AI infrastructure remained strong, including for earlier-generation GPUs. Microsoft had not immediately responded to that agency’s request for comment.

In a statement carried in the separate account, Oracle’s spokesperson also pointed to the company’s September disclosure that it had recently renewed GPU-rental contracts at a 20% premium to their original long-term pricing.

That company claim concerns other renewals. It does not establish the price Oracle can obtain for the capacity Microsoft is relinquishing, or show that those contracts involve comparable equipment and terms.

The next evidence needed is the amount of capacity leaving, when the wind-down finishes, and whether Oracle can rent it to other customers at comparable or higher prices. With the original contract’s terms undisclosed, Microsoft’s savings and Oracle’s revenue exposure cannot be calculated. The financial test is whether capacity that no longer fits one buyer can keep earning for its supplier.