Meta lines up long AI chip deals and plans to sell extra capacity

In short: Meta is locking in large, long-term supplies of AI computing power and is planning a cloud business to sell any extra capacity.

What's going on

Meta has been signing very large, multi-year agreements to secure AI computing power. This means the chips and data centers needed to run and train modern AI systems. You can think of it like reserving years of electricity and factory space in advance because you expect shortages.

One reported deal involves buying enough AMD AI chips to support up to 6 gigawatts of data center capacity over five years, valued at more than $100 billion. Meta has also committed to buying several million Nvidia chips, in a package expected to cost tens of billions of dollars.

Meta is also using outside suppliers for capacity. It has contracts with CoreWeave totaling up to $35 billion for AI cloud capacity through 2032, and a $27 billion agreement with Nebius over five years. It is also linked to roughly 1.6 gigawatts of capacity at two Crusoe data centers in Texas and Missouri.

At the same time, Meta is planning a new cloud business to sell access to AI computing power when it has more than it needs. That could mean renting out raw computing capacity, or hosting Meta’s own AI models for other companies to use through an API (a simple way for software to request a service, like ordering from a menu).

What to watch

These moves point to a broader reality that high-end AI computing power is scarce and expensive. The key question is whether Meta can actually “overbuild” and then attract outside customers, while competing with established cloud providers like Amazon, Microsoft, and Google.

Source: NYTimes