On March 16, 2026 — the same day GTC opened — Nebius Group N.V., the Amsterdam-based AI cloud company listed on Nasdaq as NBIS, announced a new long-term AI infrastructure supply agreement with Meta. The contract is worth up to roughly $27 billion over five years: $12 billion of dedicated capacity across multiple locations, plus a commitment by Meta to buy up to $15 billion of residual capacity on Nebius clusters yet to be built. Delivery begins in early 2027.
It is one of the largest customer contracts a neocloud — the new generation of GPU-specialist cloud providers — has ever landed, and it marks the moment Meta, famous for building its own data centers, formally pulls external GPU clouds into its core compute portfolio. Only a week earlier, NVIDIA announced a $2 billion strategic investment in Nebius. Two mega-transactions in one week have pushed the company to the center of the AI infrastructure market.
Deal Structure: $12 Billion Dedicated, $15 Billion Residual
The first component is straightforward: $12 billion of dedicated AI compute capacity for Meta, spread across multiple sites over the five-year term. The second component is the unusual one. Meta has committed up to $15 billion to purchase leftover capacity on Nebius’s upcoming AI clusters — Nebius will offer that capacity to its own AI cloud customers first, and Meta takes whatever remains.
For Nebius, this works as an anchor contract: regardless of how much external demand materializes, Meta guarantees the floor, which directly underwrites the capital expenditure for next-generation data centers and makes long-term revenue predictable. Founder and chief executive Arkady Volozh said the deal helps secure the large, long-term capacity contracts that accelerate the build-out and growth of the company’s core AI cloud business. Nebius added that its 2026 financial guidance is unchanged.
Among the First Large-Scale Vera Rubin Deployments
The hardware substrate of the deal is NVIDIA’s Vera Rubin platform, unveiled the same day in the GTC keynote — and Nebius will be among the first to deploy it at scale. Data Center Knowledge’s analysis notes what the platform implies: higher-bandwidth memory, next-generation interconnects, and liquid-cooled racks drawing well over 100 kW each, clusters sized to power some of the largest AI training systems of the late 2020s.
For Meta, locking in next-generation platform capacity nearly three years ahead is insurance on the training needs of its next model generation: the early-2027 delivery window lines up neatly with the next training cycle.
A Turning Point for the Neocloud Model
Matt Kimball of Moor Insights & Strategy put the market significance plainly: there are a lot of neocloud providers fighting for attention and headlines, and Meta’s commitment amounts to major validation for Nebius. Hyperscaler demand for AI infrastructure right now, he argued, is “insatiable.” He pointed to CoreWeave as proof that hyperscalers renting from specialized cloud providers is an established pattern — and that “the neocloud opportunity isn’t just a North America–centric play anymore.”
In other words, this is more than one contract; it is a business model being confirmed. Neoclouds have been upgraded from “spare compute for startups” to “long-term suppliers to giants,” and the presence of an anchor customer is what makes multibillion-dollar build-outs financeable.
What It Means for the Compute Market
Data Center Knowledge situates the deal inside Meta’s broader multi-supplier posture: an earlier large AMD chip commitment, a Corning fiber agreement, and now Nebius capacity — Meta is spreading bets across external suppliers alongside its own construction, in the same direction as February’s reports of Meta renting TPUs from Google. For hyperscalers, external cloud capacity is now a routine instrument, not an emergency measure.
Two practical effects for developers and startups. First, when contracts in the $27 billion class lock up neocloud capacity almost three years ahead, the secondary market’s available supply and pricing get squeezed; if you are planning long-horizon training capacity, treat “big-customer priority” as a line item in your risk model. Second, the residual-capacity structure means part of what future Nebius clusters don’t sell is already spoken for — open neocloud capacity is becoming scarce, and the negotiating window will not stay open forever.
Sources
- Nebius signs new AI infrastructure agreement with Meta — Nebius Newsroom
- Meta, Nebius sign $27B deal to power Nvidia Vera Rubin deployments — Data Center Knowledge
- Nebius, Meta Agree to $27 Billion AI Infrastructure Pact — The Wall Street Journal
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.