Microsoft let an Oracle GPU lease expire because old accelerators stopped earning their keep

Microsoft has declined to renew the cloud-server agreement it struck with Oracle in 2023. The servers were originally emergency capacity: Microsoft needed somewhere to run its own AI applications, including Bing features, while preserving room in Azure for paying customers. A person familiar with the decision told The Information that the Oracle machines contained older chips that became less useful over time. Microsoft was not relying on them for a significant share of its compute.

This is not evidence that Microsoft has stopped leasing third-party infrastructure. It continues to spend heavily on outside capacity because its own data-centre construction could not match demand. The distinction is between compute in general and a particular generation, location and delivery schedule. Microsoft recently cancelled a planned Nscale lease after power delays and postponed payments to Iren when capacity arrived late. Here, the problem was not delivery but ageing.

Oracle says demand for its AI infrastructure remains strong, including for earlier-generation GPUs, and points to contracts renewed in September at prices 20% above their original levels. Both statements can be true. A chip that no longer satisfies Microsoft’s preferred workloads may still serve inference, fine-tuning or a customer with lower performance requirements. Resale value depends on the work, electricity price, network and software environment around it—not merely the model number.

The episode matters because so much AI infrastructure finance treats long leases as if they turn rapidly changing equipment into stable utility assets. The tenant’s creditworthiness is only one variable. Renewal risk rises when newer accelerators deliver enough additional throughput or energy efficiency to overwhelm switching costs. A server can remain busy and still become uneconomic for the customer whose contract originally financed it.

That makes the renewal decision a small but useful depreciation test. Microsoft needed Oracle when compute was scarce enough that almost any suitable GPU was valuable. Three years later it could discriminate. As more early boom-era leases mature, the industry will learn whether older accelerators find lower-value second lives—or whether lenders and infrastructure owners assumed useful lives that customers will not accept.