Meta announced a five-year agreement with Nebius worth up to $27 billion for AI infrastructure deployment. The deal follows Meta's $17 billion commitment to Microsoft earlier this year and marks the second time in 2025 the company has signed a multi-year infrastructure contract exceeding $15 billion. Nebius shares moved 42% intraday on the announcement.

The Nebius arrangement covers GPU clusters, networking infrastructure, and managed inference services across multiple data center regions. Meta will take capacity in stages beginning Q3 2025, with the first $6.2 billion tranche committed through December 2026. The contract includes variable pricing tied to utilization thresholds and allows Meta to scale compute dynamically as Llama model versions deploy. Nebius, a spinout of Yandex's AI infrastructure unit, runs roughly 180,000 H100-equivalent GPUs across facilities in Finland, Kazakhstan, and the United States. The company has been adding 22,000 GPUs per quarter since separation in late 2023.

The structure tells you what Meta needs. Two providers at this scale means the company is pricing against single-vendor concentration risk and buying optionality on deployment speed. Microsoft's $17 billion deal was Azure-native and carried software licensing bundled into the infrastructure cost. Nebius is offering bare-metal capacity with lighter middleware, likely at 18-24% lower effective cost per token at inference scale. Meta's Reality Labs segment burned $4.4 billion in Q1 2025 alone. Llama 4 inference at production scale is running Meta roughly $220 million monthly in compute, per company filings. The Nebius deal is structured to absorb that load growth without renegotiation through 2029.

The timing matters because inference economics are tightening across every hyperscaler. OpenAI's GPT-5 pricing dropped 31% in March. Anthropic cut Claude rates 26% in April. Meta is the only major model provider running inference at cost internally and giving the model away. That makes infrastructure the entire margin equation. Allocators watching AI capex feared Meta would need to raise another $30-40 billion in debt before year-end to keep pace with training and inference demand. This contract pre-commits the capital over five years and smooths quarterly cash flow volatility. It also takes 68% of Meta's projected 2025-2029 AI infrastructure spend off the balance sheet and into a structured agreement with termination penalties.

Watch Nebius's GPU delivery schedule in Q3 2025 earnings. If Meta takes more than $1.8 billion in the first six months, it signals Llama 4 inference is scaling faster than the April guidance suggested. Watch also for similar deals from Google and Amazon in Q2 or Q3. If those materialize, it confirms the hyperscalers are moving to multi-vendor infrastructure strategies as a class, not as a Meta-specific hedge. The debt markets are already pricing in another round of AI-linked issuance. Meta's 5.75% notes due 2033 widened 9 basis points this morning.

Nebius now has $44 billion in forward revenue locked through 2029, the majority from two customers. That concentration is the risk and the edge.