Meta Platforms committed $27 billion over five years to Nebius Group for dedicated AI infrastructure, with $12 billion earmarked for Nvidia Vera Rubin GPU capacity. The agreement, structured as a direct procurement rather than elastic cloud consumption, represents the largest known AI compute lock-up by a hyperscaler and pushes Nebius—formerly Yandex's cloud arm—into the top tier of enterprise GPU providers. Meta gets guaranteed access to scarce chip supply. Nebius gets balance-sheet certainty that changes its cost of capital overnight.
The deal architecture matters. Nebius will deploy Nvidia's Vera Rubin architecture, the GB200-based platform scheduled for volume production in the second half of 2025. Meta's commitment predates general availability, meaning the social platform is paying for allocation priority and deployment exclusivity during the critical ramp period when every other hyperscaler will be fighting for the same silicon. The $12 billion GPU tranche alone exceeds the annual capital expenditure of most cloud providers. The remaining $15 billion covers networking fabric, storage, power infrastructure, and cooling systems—the unsexy plumbing that determines whether GPU clusters run at 85% utilization or 40%.
This is not a cloud lease. Meta is locking capacity with contract guarantees that remove variability from its inference scaling roadmap through 2030. The five-year horizon aligns with the expected deployment cycle for Llama 4, Llama 5, and whatever comes after—models that will require 10x to 50x the inference compute of today's frontier systems as Meta pushes multimodal understanding and real-time video generation into its 3.9 billion monthly active users. Nebius gains the revenue visibility to raise debt against contracted cash flows, a structural advantage in a market where competitors are still selling on-demand capacity with 30-day cancellation windows.
The Nebius selection is a calculated divergence from AWS, Google Cloud, and Microsoft Azure. Meta already operates its own data centers and has custom silicon in production, but those assets are optimized for training and near-term inference. The Nebius arrangement offloads the capital intensity and lead-time risk of scaling next-generation inference infrastructure while preserving optionality on where workloads actually run. Nebius, spun out of Yandex in late 2023 and now headquartered in Amsterdam, has eight data centers across Europe and the U.S. with direct peering into Meta's backbone. The deal converts Nebius from a regional player into a strategic counterweight to the hyperscale trio, with a single customer representing more contracted revenue than the company's entire 2024 run rate.
Watch three follow-on moves. First, Nvidia's shipment schedule for Vera Rubin in Q3 and Q4 2025—any delay cascades directly into Meta's inference roadmap and forces reallocation among existing H100 and H200 clusters. Second, whether Nebius raises debt or equity against the contracted $27 billion within the next 90 days; the capital structure choice will signal how much margin cushion is baked into the deal. Third, how AWS, Google, and Microsoft respond to a tier-one customer going external for GPU capacity at this scale—either with price concessions on their own Nvidia allocations or accelerated deployment of their custom AI chips (Trainium, TPU v6, Maia).
Meta just paid $27 billion to remove the single biggest variable in its AI strategy: whether the chips will be there when the models are ready.