Meta isn’t buying compute. It’s reserving it, five years out, on hardware that hasn’t shipped yet.
The social media giant signed a five-year AI infrastructure agreement with Nebius Group worth up to $27 billion, according to the official announcement from Nebius. The deal breaks into two tranches: $12 billion in dedicated AI infrastructure capacity and $15 billion drawn from Nebius cloud operations. Deal coverage describes it as Meta’s largest single contract for cloud and data center services.
The timeline is the detail that matters most. Deliveries are expected to begin in early 2027, according to deal coverage, which means Meta is committing capital now for infrastructure it can’t touch for the better part of a year. The agreement, described by the companies as leveraging “the first large-scale deployments of the Nvidia Vera Rubin platform,” ties Meta’s capacity pipeline directly to a GPU generation that’s still in its pre-ship phase. That language originates from the deal announcement itself, not from independent Nvidia verification.
What does Nebius bring that a hyperscaler doesn’t? That’s the quieter question here. Nebius is a publicly traded company, smaller than AWS, Azure, or Google Cloud by a significant margin. Meta’s choice of a specialist infrastructure partner over an established hyperscaler suggests either price advantage, flexibility in deal structure, or a deliberate move to diversify compute supply chain away from the major platforms.
For AI infrastructure strategists, this deal represents a data point in an emerging pattern: frontier AI development is shifting from spot-market compute purchasing toward multi-year, locked-in capacity agreements. Five-year contracts don’t exist in a market where demand is uncertain. They exist when the buyer is certain about the demand and willing to pay for the certainty.