Meta signed a five-year agreement with Nebius valued at up to $27 billion for AI infrastructure services. The deal, announced this week, positions the former Yandex cloud division as a principal supplier in Meta's accelerated compute buildout. Nebius shares moved 40% in early trading before settling 28% higher by close.
The contract follows a $17 billion arrangement Nebius closed with Microsoft in late 2024. Both deals structure capacity commitments across distributed GPU clusters, primarily NVIDIA H100 and H200 configurations with optional liquid-cooled deployments. Meta's agreement includes performance milestones tied to inference latency and training throughput, with payment tranches released quarterly against verified uptime metrics. Nebius will deliver the first 120,000 GPU equivalents by Q3 2025, scaling to 450,000 by year-end 2027.
The arrangement matters because it confirms the emergence of a third-party compute layer between hyperscalers and foundation model operators. Meta already operates 600,000 GPUs internally but now treats external capacity as a variable cost lever, preserving balance sheet flexibility while maintaining training velocity. Nebius benefits from the capital-light broker model: it secures long-term purchase orders from chip manufacturers, then resells capacity at 12-18% gross margins without owning data centers. The Microsoft and Meta contracts alone represent $44 billion in forward revenue against an estimated $6 billion in capital outlays, a leverage ratio traditional cloud providers cannot match.
The deal also signals cooling in custom silicon development. Meta paused its MTIA chip roadmap in December, citing vendor risk and scaling friction. Committing $27 billion to third-party compute suggests the company now views proprietary accelerators as a hedge, not a primary path. That reversal benefits NVIDIA, which underwrites the majority of Nebius capacity, and pressures AWS, Google Cloud, and Oracle to either match the broker model or accept margin compression on infrastructure services.
Operators should watch Nebius capacity delivery through Q3 2025, particularly whether it meets the 120,000 GPU commitment without supply-chain delays. Meta's earnings call in late April will clarify whether this spend replaces or supplements internal buildout. The company guided $60-65 billion in total 2025 capital expenditures; if unchanged, the Nebius deal likely offsets planned data center construction. Watch also for similar announcements from Anthropic, xAI, or OpenAI, any of which would establish Nebius as the de facto compute utility for the foundation model tier.
Nebius now holds $44 billion in contracted revenue against a market capitalization of $22 billion, implying the equity prices in execution risk but not the platform value of becoming the compute clearing layer for AI.