Particle.news

Microsoft Winds Down Oracle AI Capacity Agreement

The move reflects a push to newer GPUs, tighter oversight of large outside deals and faster buildout of company-owned AI infrastructure.

Overview

  • Microsoft has begun deactivating rented Oracle GPU servers and is declining renewals for older-generation capacity, reports said on Friday.
  • Company sources told reporters the Oracle machines used earlier Nvidia chips that no longer met Microsoft’s evolving performance and cost targets.
  • Oracle disputed parts of the coverage and said capacity that came up for renewal was either renewed or resold under long-term contracts at about a 20% premium.
  • Unwinding the cross-cloud setup requires moving model checkpoints and datasets, rewiring private interconnects and identity access, and requalifying performance to avoid breaking latency.
  • The change comes as Microsoft reviews more than $80 billion in external capacity commitments and shifts toward other leasing partners and expanded in-house data center and custom-hardware plans, a shift that could tighten supply for third-party GPU providers.