How much water does ChatGPT use?
About 2.20 mL for a typical prompt, counting its share of training. Only 0.55 mL of that evaporates cooling the data center; the other 1.65 mL is used up at the power plants that make its electricity. That's a half-liter bottle every 227 prompts or so, and about 937.4 mL for a month of everyday use.
One prompt
- Data center cooling
- 0.55 mL
- Power plants
- 1.65 mL
- Water consumed
- 2.20 mL
A month at 10 prompts a day
- Data center cooling
- 234.1 mL
- Power plants
- 703.4 mL
- Carbon, for comparison
- 167 g CO₂e
- Water consumed
- 937.4 mL
How I got this
I start from the energy: about 0.3 Wh to answer a prompt (Epoch AI), plus its share of training the model, which is about 40% of machine learning energy (Patterson et al. 2022). Google's 2025 measurements of Gemini imply about 1.1 liters of cooling water per kWh of server power. Power plants on the US grid consume about 3.14 liters per kWh, mostly in cooling towers (Li et al., Making AI Less Thirsty), and 5% of the power is lost on the way.
These are liters consumed, meaning evaporated or otherwise not returned, not liters withdrawn and given back. Where the water comes from matters more than the amount, and I assume about 67% of data center water and 40% of power plant water comes from water-stressed areas. When I price it, a liter from a stressed basin counts in full and a liter from a well-watered one counts 20%, which brings a prompt to 1.26 mL of scarce water.
Every factor and source is on How it's calculated.
What it costs to cover
The water in a month of everyday use costs under 1¢ to cover, at $0.30 per 1,000 scarce gallons, which funds paid water-saving work in dry regions. Covering all of a month's AI use, with its carbon and air pollution, comes to under 1¢ in donations.
That's below the $1.00 smallest payment, so most people pay once: $2.57 with fees covers past use and every year ahead.
Measure your own use and see what it would cost to cover.