Hong Life CalculatorsUnderstanding the AI water footprint calculator — concept and principle
Data centres that run AI chatbots use water, both to cool servers and indirectly through the power plants that supply their electricity. This calculator estimates the water per request and your total over a period using the model of Li, Yang, Islam and Ren, giving an order-of-magnitude figure rather than a precise measurement.
How it is calculated
Water per request (mL) = E × (WUE + PUE × EWIF)
- E is server energy per request in Wh (equivalently mL when multiplied by L/kWh).
- WUE is on-site water use effectiveness: litres evaporated for cooling per kWh of server energy.
- PUE is power usage effectiveness: total facility energy ÷ server energy (1 or more).
- EWIF is the electricity water intensity factor: litres consumed by power generation per kWh.
- Total = water per request × requests per day × days; on-site share = WUE ÷ (WUE + PUE × EWIF).
The principle
Computing uses electricity, and nearly all of that electricity becomes heat. Many data centres remove the heat with evaporative cooling, which consumes water on site. Generating the electricity also consumes water at power plants, for example in cooling towers. The facility uses more energy than the servers alone because of cooling and power conversion, which is why PUE multiplies the off-site part. Multiplying energy by these water intensities converts kilowatt-hours into litres. Because Wh × L/kWh gives mL, the formula works directly in millilitres.
Worked example
Default inputs: 20 requests per day for 365 days, 4 Wh per request, WUE 0.55, PUE 1.17, EWIF 3.142 L/kWh (the study's U.S. averages).
- Off-site factor = 1.17 × 3.142 = 3.676 L/kWh; total factor = 0.55 + 3.676 = 4.226 L/kWh.
- Water per request = 4 × 4.226 = 16.9 mL, shown rounded as 17 mL.
- Your total = 16.9 × 20 × 365 ÷ 1,000 = 123.4 litres.
- Requests per 500 mL bottle = 500 ÷ 16.9 = 29.6.
- On-site cooling share = 0.55 ÷ 4.226 = 13%; electricity = 4 ÷ 1,000 × 7,300 = 29.2 kWh.
| Server energy per request (Wh) | Water per request (mL) |
|---|---|
| 1 | 4.2 |
| 2 | 8.5 |
| 3 | 13 |
| 4 | 17 |
| 5 | 21 |
| 6 | 25 |
| 8 | 34 |
How to read the result
The study's estimates range from about 7 to 48 mL per request across the U.S. locations it lists, depending on local climate and the power grid. Energy per request is not published by most providers and varies with the model, reply length and hardware, so treat results as millilitres to tens of millilitres per request. Most of the water in this example is off-site, at power plants, so cleaner and less water-intensive electricity reduces the footprint more than cooling changes alone.
Common mistakes
Frequently asked questions
Why does AI use water at all?
Data centres evaporate water to cool servers, and power plants consume water to generate electricity.
What is PUE?
Total facility energy divided by IT equipment energy; 1.0 would mean no overhead at all.
Does location matter?
Yes. Climate affects cooling water, and the local power mix affects off-site water.
Sources: P. Li, J. Yang, M. A. Islam and S. Ren, "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models", Communications of the ACM, 2025 (arXiv:2304.03271)