Microsoft lets Oracle data-center lease expire amid capacity review

By Elena Vasquez

Microsoft is letting a 2023 Oracle data-center lease expire. It is also reviewing other data-center contracts.

The expired agreement covered computing capacity based on older Nvidia chips, leased for AI features in Microsoft’s Bing search engine. Bloomberg cited insiders who said, “These no longer meet Microsoft’s current requirements,” explaining why the company did not renew.

Why did Microsoft let the Oracle lease expire?

The 2023 Oracle deal was one of the first signs that data-center computing capacity could become scarce. When Microsoft signed it, the company planned to run Bing’s AI features on the available hardware. Letting the contract lapse now reflects a change in workload requirements, not a retreat from external computing.

Oracle disclosed the commercial pressure around legacy systems in September: “Customers paid a 20 percent premium for contracts with older hardware,” Oracle said. For every $100 of base contract value, that premium adds $20, yet Microsoft still let the agreement lapse.

For Microsoft, renewal would have preserved contracted capacity, but not a hardware profile matching its present requirements. The decision separates the value of a data-center commitment from the usefulness of the accelerators attached to it.

What changes in the Nscale contract?

Microsoft is also examining agreements with other data-center operators. It declined an expansion of leased Nscale capacity in Texas despite holding an option to increase the deployment. Datacenter Dynamics reported the decision, while Bloomberg said Microsoft also remains frustrated by construction and commissioning delays at several centers.

The two decisions differ in their immediate trigger. The Oracle lapse concerns the age of contracted hardware; the Nscale refusal follows Microsoft’s frustration with construction and commissioning delays. Neither case removes Microsoft’s need for third-party computing.

Contract options preserve commercial flexibility, not delivery certainty. Microsoft’s Nscale decision shows how a valid option can remain unused while broader construction and commissioning delays complicate new capacity.

Microsoft is weighing hardware fit against delivery timing. Some contracted capacity runs on chips that no longer fit current workloads, while planned additions may arrive late because construction and commissioning have slipped.

How broad is the data-center construction slowdown?

RegLayer, a compliance service, recorded 47 large U.S. AI data-center projects delayed, canceled or blocked this year. The count spans delays, cancellations and blocks, showing that Microsoft’s constraints sit inside a broader U.S. development slowdown.

At 47 projects, the figure shows the scale of disruption across the U.S. pipeline, rather than a single operator’s setback. For data-center operators, that mismatch limits how quickly new leased capacity becomes available.

The count does not indicate how much computing capacity each project would have supplied. It does show that project-execution risk extends across the market rather than remaining inside Microsoft’s supplier base.

For Microsoft specifically, the delays complicate its attempt to expand leased Nscale capacity in Texas. A contractual option does not guarantee commissioned computing when projects slip.

How will Maia change Microsoft’s Nvidia exposure?

Separately, Microsoft plans to run a larger share of AI inference for its own products on its in-house Maia accelerators. The aim is to reduce dependence on Nvidia by moving more internal workloads onto Microsoft-designed chips.

The emphasis is on inference for Microsoft’s own products, not a wholesale replacement of every Nvidia-based workload. That distinction would put more internal inference on Maia while third-party capacity stays in the near-term mix.

Maia will not cover the company’s full AI-computing requirement in the near term. Microsoft therefore plans to expand internal inference capacity while continuing to assess external contracts against workload requirements and delivery performance.