How they use Nebius AI Cloud

Microsoft buys dedicated GPU infrastructure from Nebius under a five-year commercial agreement. SemiAnalysis reports a contract value of $17.4 billion, expandable to $19.4 billion. The September 2025 filing calls for capacity to be delivered in tranches from Nebius's Vineland, New Jersey data centre. Nebius reported delivering the first tranche in November 2025. In February 2026, it said remaining tranches were scheduled throughout 2026, with more than half due in the second half. In July 2026, it reported delivery of the latest planned tranche, with remaining tranches on the contracted schedule. Its Q2 2026 shareholder letter subsequently reported delivery of all contracted Microsoft capacity tranches, including two over the preceding month.

At Inflection 2026, Nebius grouped Microsoft with users of its bare-metal infrastructure. It described supplying Microsoft and Meta with custom racks and servers, the latest GPUs with InfiniBand, and multi-tier storage for their internal ecosystems. In its March 2026 conference transcript, Nebius also described Microsoft and Meta as customers buying basic bare-metal services and using their own software stacks. A Nebius announcement connects the Microsoft agreement to NVIDIA Blackwell systems. In November 2025, SemiAnalysis linked Nebius's anticipated adoption of GB300 NVL72 rack-scale systems to customers including Microsoft. Nebius said it secured the memory and storage components needed for the full scope of its Microsoft and Meta contracts in the second half of 2025, before price increases.

The London Institute for Mathematical Sciences describes Nebius as assisting Microsoft with training and deploying frontier models at scale. The companies have not named the Microsoft models or internal services using the capacity.