$21 billion. That's how much additional computing spend Meta is committing to CoreWeave, the GPU cloud provider that went public in early 2025. The deal puts Meta among CoreWeave's largest customers and confirms that AI infrastructure costs are not leveling off.
CoreWeave built its business renting out Nvidia GPU clusters - specialized processors used for training AI models and running them in production (called inference). Meta needs this capacity to train and serve its Llama family of open-weight models, which power AI features across Facebook, Instagram, and WhatsApp, as well as the standalone Meta AI assistant.
The Broader Spend Picture
Meta has committed to spending between $60 and $65 billion on capital expenditures in 2026, a significant increase from prior years. AI infrastructure - data centers, networking, and GPU rentals - makes up the majority of that figure. The CoreWeave commitment is part of that already-announced budget, not a surprise addition on top of it.
The pattern across big tech is consistent: Microsoft, Google, Amazon, and Meta are all committing tens of billions annually to AI compute. Companies that can't match this spend - which is almost every company - will access AI through the APIs these players expose rather than running their own foundation models. That dynamic is not changing; this deal accelerates it.
For the AI tools that everyday users interact with, this spending race has a practical consequence: foundation models get more capable as more compute goes into training them. The competition is moving to the product layer, not the infrastructure layer. That's where most AI productivity tools actually compete - and where a $21 billion compute advantage doesn't automatically translate into a better product.