Summary: Nebius announced a five-year AI infrastructure agreement with Meta on March 16, 2026, worth up to about $27 billion. Meta receives $12 billion of dedicated capacity and commits up to $15 billion for residual capacity on clusters Nebius has yet to build, with first deliveries in early 2027. The deal moves external GPU clouds into Meta’s core compute portfolio and supports Vera Rubin deployments.
Nebius Group N.V. — an AI cloud company headquartered in Amsterdam and listed on Nasdaq under the ticker NBIS — picked the day GTC opened, March 16, 2026, to announce a fresh long-term AI infrastructure supply agreement with Meta. The five-year contract is worth up to approximately $27 billion, split between $12 billion of dedicated capacity spread over multiple locations and a Meta commitment to purchase up to $15 billion of residual capacity on clusters Nebius has yet to build. First deliveries start in early 2027.
Measured against every customer contract a neocloud — the new wave of GPU-specialist cloud providers — has ever signed, this one ranks among the largest, and it pinpoints the moment Meta, long known for erecting its own data centers, formally brings external GPU clouds into its core compute portfolio. A week before that, NVIDIA had put a $2 billion strategic investment into Nebius. Landing two mega-transactions inside seven days has moved the company to the heart of the AI infrastructure market.
Breaking Down the Deal: $12 Billion Dedicated, $15 Billion for Leftover Capacity
The opening piece is the $12 billion: Meta receives AI compute capacity dedicated to it across several sites for the length of the agreement. The piece that breaks the mold is the second. Meta has pledged as much as $15 billion toward leftover capacity on the AI clusters Nebius is about to construct — the order of operations matters, because Nebius first offers that capacity to customers of its own AI cloud, and Meta then picks up whatever goes unsold.
From Nebius’s side, the arrangement behaves like an anchor contract: however much outside demand ends up materializing, Meta stands behind the floor, and that floor is what directly supports capital expenditure on next-generation data centers while keeping long-term revenue in predictable territory. Founder and chief executive Arkady Volozh said large, long-term capacity contracts of this kind speed the build-out and growth of the company’s core AI cloud business. Nebius also noted that its financial guidance for 2026 stands unchanged.
An Early, Large-Scale Home for Vera Rubin
Underneath the whole agreement sits NVIDIA’s Vera Rubin platform, introduced that same day in the GTC keynote, and Nebius counts among the earliest operators set to run it at scale. The analysis at Data Center Knowledge spells out what the platform carries with it: memory with higher bandwidth, next-generation interconnects, and liquid-cooled racks each pulling well beyond 100 kW — clusters built to carry some of the largest AI training systems of the late 2020s.
From Meta’s chair, reserving capacity on a next-generation platform almost three years before it arrives works as insurance for the training demands of its coming model generation: an early-2027 delivery window that dovetails with the next training cycle.
Where the Neocloud Business Model Turns
The market reading comes from Matt Kimball at Moor Insights & Strategy: the neocloud field is crowded with vendors chasing eyeballs and headlines, so a commitment carrying Meta’s name stands as weighty validation for Nebius — and hyperscaler appetite for AI infrastructure is, in his words, “insatiable.” He reached for CoreWeave as evidence that hyperscalers renting capacity from specialized cloud providers is already an established pattern, adding that “the neocloud opportunity isn’t just a North America–centric play anymore.”
Put differently, calling this one contract undersells it — what has been confirmed is a business model. Neoclouds move up from “spare compute for startups” to the rank of “long-term suppliers to giants,” and it is the anchor customer attached to the deal that turns multibillion-dollar build-outs into something financeable.
Implications for the Compute Market
Data Center Knowledge frames the agreement as one piece of Meta’s wider multi-supplier posture: the sizable AMD chip order from before, the fiber agreement with Corning, and now capacity from Nebius — alongside its own construction, Meta keeps distributing wagers across outside vendors, a direction that matches February’s reports of Meta renting TPUs from Google. For hyperscalers, external cloud capacity has become a routine instrument rather than a measure reserved for emergencies.
For developers and startups, two effects follow. One: once deals of the $27 billion class tie up neocloud capacity nearly three years in advance, available supply and pricing on the secondary market both come under pressure — anyone planning long-horizon training capacity should carry “big-customer priority” as an explicit line in the risk model. Two: the residual-capacity structure means some of what Nebius’s future clusters cannot sell is already spoken for — openly available neocloud capacity is turning scarce, and the window to negotiate will not stay open indefinitely.