Microsoft is reportedly letting a 2023 compute agreement with Oracle run out, according to Bloomberg, which cites people familiar with the matter. The contract covered older-generation Nvidia hardware that no longer fits Microsoft’s needs, and Microsoft has been switching off that capacity in recent months. The deal’s value was not disclosed, though the chips were meant to power an AI-assisted version of Bing.
Oracle answered quickly. A spokesperson told DCD that demand for its AI infrastructure remains strong, even for earlier-generation GPUs. The company says GPU capacity up for renewal in its first quarter was renewed or resold at a 20% premium for long-term contracts, though most of those chips were four years old or older. Oracle presents that as proof that older hardware holds its value. That is Oracle’s own claim, and it is worth watching whether the pattern lasts. The two companies also have newer contracts that appear unaffected.
History between them adds some tension. In June 2026, reports said a $3 billion deal fell apart because Oracle lacked FedRAMP, the framework that clears cloud services to handle US government data. Oracle denied it, called the reports inaccurate and described Microsoft as both an OCI partner and a customer.
Bloomberg also describes a broader pattern. Microsoft is reportedly reviewing contracts with several neocloud providers. Estimates put its neocloud leases above $60 billion a year ago and above $80 billion now. People familiar with the matter say the company is tightening oversight of some data center projects with Nebius, Iren and Nscale, and walking away from ones it considers risky.
Recent events fit that picture. Iren said one project slipped by one to two months after Microsoft asked for a networking change. In the UK, Microsoft pulled out of Nscale’s Loughton project because the grid couldn’t connect the site before the 2030s. Even so, Microsoft remains a major customer of these providers and holds ongoing large contracts.
Microsoft told Bloomberg that it manages relationships with infrastructure partners against its requirements and customer demand, and that monitoring delivery and adjusting plans when circumstances change counts as normal, responsible practice.
Both readings can hold at once. A company can behave carefully with a vast budget and still depend on partners. What bears watching is whether trimming turns into a retreat, and whether the used-GPU market really holds up.

