Meta signed a five-year agreement worth up to $27 billion with Nebius for AI data center capacity, announced this morning without advance disclosure. The deal follows Microsoft's $17 billion commitment to the same provider in late 2025, concentrating $44 billion of long-term compute procurement with a single Netherlands-domiciled infrastructure operator spun out of Yandex.
Nebius operates GPU clusters across Finland, Kazakhstan, and the United States, with expansion planned into Sweden and the United Arab Emirates. The Meta contract specifies capacity delivery beginning Q3 2026, with ramp milestones tied to chip availability from NVIDIA and potential alternative suppliers. Nebius publicly trades under ticker NBIS after completing its separation from Russian parent Yandex in November 2024, immediately attracting Western hyperscaler attention previously blocked by sanctions exposure. The company reported $890 million in trailing revenue as of year-end 2025, making the Meta contract alone worth 30 times current annualized sales.
The procurement pattern signals constrained optionality among hyperscalers seeking non-Chinese compute at scale outside Amazon's internal buildout. Microsoft and Meta collectively represent roughly 65 percent of Western enterprise AI workload by training parameter volume, and both now rely on the same third-party infrastructure provider for material portions of their 2026-2028 capacity roadmap. Nebius lacks the vertical integration of CoreWeave or Lambda Labs but offers geographic diversity and power access in jurisdictions with stable regulatory frameworks for AI compute. The Finland and Kazakhstan sites draw direct cooling from ambient air systems, reducing operating expenses by an estimated 22 percent compared to liquid-cooled facilities in Virginia or Ireland.
Allocators should track three developments. First, whether Nebius converts contract commitments into bonded construction financing for the Sweden and UAE sites, expected by Q2 2026. Second, NVIDIA's Blackwell chip allocation between Nebius and direct hyperscaler orders, visible in Nebius earnings calls starting May. Third, energy procurement announcements from Nebius tied to nuclear or renewable baseload sources, necessary to sustain power draw above 1.2 gigawatts by 2027. Any delay in chip supply or power contracting compresses Nebius margin and creates schedule risk for both Meta and Microsoft inference deployment.
Meta's commitment arrives four months after its Llama 4 training cycle began, suggesting the purchased capacity serves inference scaling rather than next-generation foundation model work. The company already operates 600,000 NVIDIA H100 equivalents internally, enough for parameter counts beyond 2 trillion, but lacks distributed inference infrastructure for consumer-facing deployment at WhatsApp and Instagram scale across non-US markets. Nebius provides that reach without requiring Meta to navigate Finnish or Kazakh data residency law directly, a structure worth noting for其他 hyperscalers evaluating third-party compute partnerships in jurisdictions where wholly owned subsidiaries face elevated compliance costs.