Overview
- Microsoft has begun switching off rented Oracle Cloud Infrastructure GPU capacity after choosing not to renew agreements for older Nvidia hardware, according to multiple reports on Nov. 6, 2026.
- The rental ran as a cross‑cloud burst: Azure workloads ran on Oracle bare‑metal GPU servers over a private interconnect so models could train on OCI and results could sync back to Azure.
- Unwinding the setup requires moving model checkpoints and datasets, reconfiguring networking and identity controls, and requalifying performance to avoid breaks in serving latency.
- Oracle says demand for its AI infrastructure remains strong and that capacity up for renewal was renewed or resold at roughly a 20% premium, and it disputes some reporting about the deal.
- Microsoft says it will pursue other leasing partners and expand its own capacity while tightening oversight of external suppliers after reviewing more than $80 billion in outside commitments.