Meta to spend up to $27 billion on Nebius Group's AI infrastructure over the next five years

Meta is prepared to spend up to $27 billion over five years to gain access to Nebius Group's AI infrastructure. It is one of the clearest signals yet that the largest tech companies now compete not only on models but on the compute without which those models cannot be trained quickly and shipped in products.

Scale of the agreement

In essence, this is long-term access to external infrastructure for artificial intelligence. Meta may put up to $27 billion toward it over five years. The wording matters: the company is not buying Nebius Group outright and is not announcing an acquisition, but reserving a large block of capacity for itself. For the market, it signals that even a big tech company can no longer simply build everything in-house when demand for training and serving AI systems grows faster than new capacity comes online.

The sum alone sets the scale. If the commitments are drawn down in full, this would be one of the largest known infrastructure contracts in the AI segment in recent times. Deals like this are usually signed not for experiments but for specific product and research workloads: training new generations of models, serving users, testing features and building a buffer for demand peaks. In other words, Meta appears to be insuring itself against a compute shortage just as the race for model quality is accelerating.

The five-year horizon also matters. Over such a period, companies usually buy not one-off capacity but predictability of supply. It helps balance between their own data centers and outside partners: keep part of the workload in-house and quickly move part to wherever the needed number of accelerators is available. For Meta, this approach lowers the risk that new AI releases or model updates will have to be delayed because of a plain lack of infrastructure.

Why Meta needs it

For Meta, infrastructure is no longer a supporting layer but a direct part of product strategy. The more actively the company develops AI services, the more it needs GPUs, networks, storage and delivery channels that can scale without a long pause to build its own data centers. Access to capacity years ahead provides not only speed but predictability: teams can plan model training and feature launches without checking hardware availability every quarter.

In practice, such a contract can give Meta several advantages at once:

• more resources for training large models;
• stable capacity for running AI features in mass-market products;
• spare throughput for periods of sharp load growth;
• less dependence on a scenario in which all infrastructure has to be built in-house.

It is a strong signal for Nebius Group as well. If the company wins a contract of this size from Meta, the market sees it not as a niche contractor but as an infrastructure partner able to handle one of the most demanding AI workloads in the world. In the AI market this matters as much as the money: such agreements raise the trust of other large customers who need not just GPU access but confidence in reliability, timelines and scale of delivery.

What it means

The AI race is shifting from demos and releases to energy, servers and long-term compute contracts. The news of Meta's plan to spend up to $27 billion shows a simple shift: the winners will be not only those who build strong models but those who locked in infrastructure in advance. For the market, this means continued growth in demand for independent AI providers and an even higher cost of mistakes in capacity planning.

*Meta is designated an extremist organization and banned in Russia.