Meta Platforms signed a $27 billion, five-year infrastructure agreement with Nebius Group, with $12 billion earmarked exclusively for Nvidia's Vera Rubin GPU architecture. The contract represents the largest single AI capacity commitment disclosed this cycle.
Nebius Group—the AI infrastructure spinoff from Yandex—will provision compute capacity across multiple data center facilities through 2030. The Vera Rubin allocation is meaningful: Nvidia's next-generation GPU line won't ship in volume until late 2025, and Meta has secured delivery priority for approximately 40% of Nebius's planned Rubin deployment. The remaining $15 billion covers Hopper H100 and H200 systems, networking infrastructure, and co-location services. Nebius disclosed the contract represents roughly 60% of its forward revenue visibility through 2029.
This is a pre-buy on unprecedented scale. Meta is locking tomorrow's capacity at today's pricing, insulating itself from the spot-market chaos that defined 2023 and early 2024 when H100 lease rates spiked to $4.20 per chip-hour. By contrast, the Nebius agreement prices Rubin capacity at a fixed rate reported near $2.80 per chip-hour—a 33% discount to current Hopper spot. The economics assume Rubin delivers the 2.5x training efficiency Nvidia projects, but Meta is hedging execution risk across its Llama model roadmap regardless of whether that materializes.
The deal also clarifies the infrastructure winner in the Yandex breakup. Nebius retained the data center assets, the hyperscale relationships, and the capital allocation discipline. Yandex kept the consumer internet properties and the geopolitical exposure. Nebius now has contracted revenue equal to 11x its 2024 run rate, and the Meta anchor gives it the credit profile to raise another $8-10 billion in project finance by mid-2025. Brookfield Asset Management already holds a 19% stake and is expected to lead the next funding round.
The GPU supply chain tightens further. Nebius's Rubin allocation implies TSMC's CoWoS-L packaging line is already oversubscribed into Q3 2026. Meta, Microsoft, and Google now control approximately 70% of the Rubin wafer starts Nvidia has reserved through the end of 2026. For late-stage AI companies without similar contracts—Anthropic, Cohere, Mistral—the compute access gap widens. Expect those firms to either raise at lower multiples to afford spot pricing or accept capacity-sharing arrangements that complicate their independence narratives.
Watch Nebius's construction velocity in Finland and Kazakhstan, where it operates carrier-neutral facilities with direct fiber to Frankfurt and Singapore. The company disclosed 18 MW of new power under contract in Q1 2025, with an additional 60 MW expected online by December. Meta's Llama 4 training runs are scheduled to begin in Q4 2025, which means the first Rubin racks must be live by September. Any delay in TSMC's packaging yields or Nvidia's firmware certification creates a $400-600 million quarterly revenue hole for Nebius and pushes Meta's model launch into 2026.
The five-year horizon is the tell. Meta isn't buying compute. It's buying certainty in a market where certainty is the scarcest asset.