WATT

What does AI actually cost in energy and water? Every number here is a sourced estimate with a range you can adjust.

Energy
…
……
Water
…
……
CO₂
…
……
Assumptions: every number is contested. Drag to explore.
Wh per query …
PUE (datacenter overhead)
Uptime Institute Global Data Center Survey (2023) 1.2
Water use (L/kWh)
Shehabi et al., LBNL (2024) 1.8
Grid carbon (g CO₂/kWh)
Ember / EPA eGRID (2024) 400
For scale

Compared with Training GPT-3 once (~1,287 MWh): —

Patterson et al., Carbon Emissions and Large Neural Network Training (2021). Estimated ~1,287 MWh to train GPT-3. Illustrative; modern frontier runs differ widely and are poorly disclosed.