Electricity, water, carbon & cost of your AI usage — sourced, adjustable estimates
Pick a model and how much you use it — everything else has sensible, sourced defaults
Longer, harder tasks use more energy per request
Prompts / generations in the period below
The window your request count covers
Carbon depends on where the data center's power comes from
Water use varies enormously by site and season
US residential average is about 19¢/kWh (EIA)
Spread the one-time training run across all queries the model ever serves
One tap to load a typical usage pattern
Everything in this table is an energy comparison — apples to apples
| Activity | Energy each | Your AI energy equals | Per AI request |
|---|
Each bar shows a single request as a share of a familiar baseline — same units within every bar
Estimated energy per typical request — your model highlighted
Cumulative electricity if this usage continues for 12 months
The per-request estimates behind this calculator — sortable, downloadable, and honest about uncertainty
Every figure is energy first — water, carbon and cost are derived from it using the assumptions you pick
The best public anchors:
Data centers shed heat, often by evaporating water. UC Riverside researchers estimate roughly 500 mL per 5–50 prompts depending on site and season. Efficient hyperscalers run near 1.1 L/kWh; hot-climate evaporative sites plus the water behind power generation can reach ~4 L/kWh, while closed-loop liquid cooling cuts direct use by 70–90%.
An image costs about as much as a long text answer (1–3 Wh). Video is another world: independent measurements put an 8–12 s clip anywhere from ~90 Wh to ~1,300 Wh (Sora 2 Pro at 1080p). OpenAI shut its Sora social app down partly because of exactly these compute costs.
Training a frontier model is a one-time run estimated at tens of GWh — but amortized over the ~trillion requests a popular model serves, it adds only ~0.05–0.1 Wh per query. Inference (usage), not training, now dominates AI's total footprint.