Meta Platforms signed $41 billion in combined AI infrastructure agreements with Nebius Group and CoreWeave over a span that appears to be measured in weeks, not quarters. Nebius disclosed $27 billion in deals on March 16—split into a $12 billion initial tranche and $15 billion in options—while CoreWeave announced a separate $14 billion commitment the same day. Both contracts run five years. Both are compute-capacity purchase agreements, not equity stakes or joint ventures.

The deals formalize what has been visible in capex guidance since November: Meta is no longer building all its AI infrastructure in-house. Nebius, the Amsterdam-based spin-out from Yandex's cloud division, will deliver capacity starting in the second half of this year. CoreWeave, the GPU-cloud firm that went from crypto mining to enterprise AI in eighteen months, is already supplying Meta and is now locked in through 2030. Neither vendor disclosed margin structure, but both emphasized that Meta is pre-committing to fixed capacity at fixed periods, which means Meta is accepting utilization risk in exchange for supply certainty.

This matters because hyperscalers rarely outsource compute at this scale unless internal timelines have slipped or demand models have moved faster than construction schedules. Meta's capex guidance for 2025 sits between $60 billion and $65 billion, up from $38 billion in 2024. The $41 billion in vendor commitments represents nearly two-thirds of the lower bound, which suggests a meaningful shift in the build-versus-buy calculus. Llama 4, expected mid-year, will require training clusters in the hundreds of thousands of GPUs. Inference workloads for AI Studio, WhatsApp AI, and Ray-Ban Meta are already live and scaling faster than internal deployment can match.

The vendor split is also notable. Nebius brings deep ties to European data sovereignty frameworks and access to power grids in the Nordics and Central Europe, where energy is cheaper and cooling is structural. CoreWeave brings speed: the company can provision 10,000 GPUs in weeks, not quarters, by leveraging colocation partnerships and pre-negotiated power agreements with utilities in the U.S. and Canada. Meta is effectively buying geographic and operational optionality, hedging against single-region power constraints and permitting delays that have already pushed some U.S. data center timelines into 2027.

Operators and allocators should watch three things over the next ninety days. First, whether Nebius or CoreWeave announce additional debt or equity raises to fund the infrastructure these contracts require—neither company is big enough to self-fund $41 billion in buildout without external capital. Second, whether Meta adjusts its own capex guidance upward in the April earnings call, which would signal that these vendor deals are additive, not substitutive. Third, whether other hyperscalers follow with similar outsourcing moves, which would confirm that the internal build model has hit a ceiling.

Meta just turned $41 billion in multi-year vendor contracts into a public signal that it cannot wait for its own data centers. The contracts are five years, but the capacity need is immediate, and the vendors are unproven at this scale. That is not caution. That is urgency with a purchase order.