• Nebius's SEC filing separates the agreement into $12 billion of dedicated five-year GPU-capacity orders and a further order with potential value of up to $15 billion. The second amount is not another block of capacity that Meta immediately bought.
  • Nebius intends to sell the specified residual capacity to other AI-cloud customers. If it remains unsold, Meta must buy it for the remainder of the period ending five years after deployment. Deployments are due in tranches from early 2027 on NVIDIA Vera Rubin, at multiple locations that were not named.

The $27 billion headline contains two different commitments

Nebius announced the agreement on 16 March 2026 after its subsidiary Nebius, Inc. signed an Infrastructure Services Agreement with Meta Platforms, Inc. on 13 March. Its Form 6-K says the agreement and initial orders have a total contract value of up to approximately $27 billion, not that $27 billion had been paid, delivered or recognized as revenue.

The first part is precise: orders worth $12 billion for dedicated GPU clusters, each running for five years, with storage and connectivity. The old article's wording that Meta would spend 'at least' $12 billion was wrong. The filing states $12 billion; the upper-bound language applies to the combined structure and the residual order.

The $15 billion component is a residual-capacity backstop

A further order gives Meta access to capacity in specified upcoming Nebius clusters. Nebius says it intends to sell that capacity to third-party AI-cloud customers. Where the relevant capacity is not sold, Meta is obligated to purchase the remainder for the period ending five years from its initial deployment, up to a potential $15 billion in total.

That allocation matters. Third-party demand can take the capacity first; Meta provides a buyer for what remains. Calling the arrangement a simple discretionary option understates Meta's obligation, while adding $12 billion and $15 billion as if both were already earned overstates Nebius's present revenue. The SEC filing supports neither shortcut.

Meta is buying a service, not taking ownership of the servers

The dedicated clusters will use NVIDIA Vera Rubin and be deployed in tranches beginning in early 2027 across multiple locations. The filing does not disclose the sites, megawatts, GPU count, unit prices or tranche schedule. Nothing in the reviewed record shows that any of this new capacity was operating when the agreement was announced.

The filed master agreement describes GPU-as-a-service and related services under signed orders. Nebius or its affiliates retain ownership of the GPUs, servers and other infrastructure for tax purposes, and the contract says no tangible property passes to Meta. Onboarding, possible data-security assessment, service standards and acceptance criteria sit between a signed order and a functioning service.

The contract transfers delivery and utilization risk in different ways

The 6-K records termination rights, service-level commitments, discounted monthly fees for late delivery, warranties, indemnities and liability limits. Those provisions make the early-2027 timetable commercially meaningful, but the detailed orders are partly confidential. Public readers cannot calculate the fee path, financing proceeds, margins or revenue-recognition schedule from the headline values.

Nebius kept its 2026 guidance unchanged at announcement. In its May 2026 shareholder letter, it later described the structure as supporting financing and long-term visibility, while again separating the $12 billion purchase from the $15 billion residual arrangement. That is management's assessment; delivery, customer mix and accounting still have to be observed.

This is Meta's second disclosed Nebius contract, not an extension of one order

Nebius's November 2025 results disclosed an earlier Meta agreement of approximately $3 billion over five years. The March 2026 contract is a distinct second agreement. Reuters placed both alongside Nebius's large Microsoft contract as evidence that specialist AI-cloud suppliers are being used for long-duration hyperscaler capacity.

The strategic question is not whether the press-release maximum is large. It is how much Vera Rubin capacity becomes operational on time, how much residual capacity third parties absorb, what Meta ultimately purchases, and how the resulting fees, capital expenditure and concentration appear in audited disclosures.

What to watch

  • Named locations, power capacity, GPU counts and the first accepted deployment tranche.
  • Evidence that NVIDIA Vera Rubin systems have arrived, passed acceptance and entered service.
  • Third-party sales against the residual clusters and the amount, if any, left for Meta.
  • Late-delivery fee discounts, terminations or other service-level consequences disclosed by either party.
  • Nebius capital expenditure, financing and customer-concentration disclosures tied to the orders.
  • Reported revenue and remaining performance obligations, kept separate from maximum contract value.

Sources