How much water does Meta's AI use?

Meta trains Llama in the US and runs AI features for three billion users, but publishes no per-query water or energy figure. Its Talavera de la Reina campus in Spain shows how much cooling design matters.

~600,000 L/yr

projected cooling water at Talavera de la Reina, Spain, after a switch to dry-air cooling (down from ~200M L/yr)

Meta's AI water story has two halves. The first is scale-by-default: Llama models train on Meta's own US data centers, and AI features ship to roughly three billion accounts, which makes even tiny per-query numbers add up fleet-wide.

The second half is siting and design. In 2023, Meta's planned campus in Talavera de la Reina, Spain was projected to need about 200 million liters of water a year for cooling (660 million for the whole development), drawn from the Tagus basin in a drought-prone region, and the plan drew scrutiny. The project won environmental approval in 2024 after switching to dry-air cooling, which cut its projected cooling water to about 600,000 liters a year. Same company, same kind of building: a design choice moved the water bill by more than two orders of magnitude.

Open weights complicate attribution in an interesting way: a Llama model running on your own GPU uses your electricity and your grid's water, not Meta's. The calculator's Llama entries assume typical cloud serving; if you self-host on a clean grid, your Mid number is genuinely lower.

Your number

What does your AI actually use?

Pick your model, set your usage, get the number, with sources.

Try the calculator

Sources

Cited on this page, in order of appearance.