Things That Took MoreWater Than Building This Site

How we got one big bucket

This book was made in one long conversation with an AI coding assistant: a first draft and several rounds of redesign, about sixteen messages from the person in all. Nobody can meter that exactly, so we built the estimate from published measurements and leaned toward the thirsty side.

Electricity. Climate scientist Zeke Hausfather logged eight weeks of his own AI coding. It came to about 150 watt-hours for each message he typed, counting all the work the AI did in reply, with an uncertainty range of roughly 60 to 290. Sixteen messages at 150 Wh is about 2.4 kWh.

Water per kWh. The data center evaporates some water to stay cool: Google's own measurement works out to about 1 liter per kWh, and the US average is lower. The power plant evaporates more: about 3 liters per kWh on the average US grid, the figure used in the "Making AI Less Thirsty" paper. That makes about 4 liters per kWh.

Training. Before it answers anything, a model is trained, and that water counts too. Google reported that over three years about two-fifths of its machine-learning energy went to training and three-fifths to answering. So each kWh of answering carries about two-thirds of a kWh of training with it, and we multiply by 5/3.

2.4 kWh × 4 liters per kWh × 5/3 ≈ 16 liters (4.2 gallons), rounded up to one 5-gallon bucket (19 liters).
Most pessimistic: 4.7 kWh × 5 liters per kWh × 5/3 ≈ 39 liters. Every thing in this book took more than that.

Not counted: water used to make the chips, build the data center, or run the screen you're reading on. The footprints for jeans, phones and burgers do count their whole supply chains, so keep that in mind when comparing.

Tap water and footprints

Some pages count water straight from the tap: flushes, showers, sprinklers. Others count a water footprint, every liter it took to grow and make something, including rain that fell on the fields. Rain on a wheat field isn't taken from anyone's river, so the two kinds aren't perfectly comparable. Each page's small print says which kind it is.

The answer doesn't change if you only count tap water. One person's flushes for a single day, or one five-minute shower, beat even the most pessimistic count for the book.

One question to an AI

Google measured a median text prompt to its Gemini apps in May 2025 at 0.24 Wh and 0.26 mL of water, about five drops. That counts the data center, not the power plant. Adding about 3 liters per kWh for the power plant gives roughly 0.7 mL more, and the 5/3 training share brings the total to about 1.7 mL, or thirty-odd drops. Long or complicated requests use more, and other companies' models differ. Mistral's own analysis put a 400-token answer from its assistant at 45 mL, counting more of the supply chain, so treat these as rough.

2,400 liters ÷ 1.7 mL ≈ 1.4 million questions. At one a minute, that's about 2.7 years.

Training, and models at home

Mistral's life-cycle analysis of Mistral Large 2 found that training it and running it for 18 months used about 281,000 cubic meters of water (281 million liters), with training the largest part. The "Making AI Less Thirsty" paper estimated that training GPT-3 used about 5.4 million liters, counting the power plants. Frontier models today are larger, and their makers haven't published comparable numbers.

Open models run on home hardware skip the data center's cooling water entirely. A 2026 benchmark on a home graphics card (RTX 4060 Ti) measured about 2.4 joules per token for 7-billion-parameter models, measuring the graphics card alone, and about a tenth of that for the smallest models. A 500-token answer from a mid-sized model is about 0.33 Wh, similar to Google's data-center figure, and about 1 mL of power-plant water on the average US grid. Home computers don't batch work the way data centers do, so local AI doesn't automatically save energy. What it changes is where the water goes, and clean electricity can take it to nearly zero.

How soon? Epoch AI finds that open models have recently trailed the best closed models by only a few months, and that models small enough for a single consumer graphics card have reached the frontier's benchmark scores within about 6 to 12 months.

Every number in the book

Sources

1. Zeke Hausfather, The real energy use of agentic AI, Skeptical Science (2026).
2. Simon P. Couch, Electricity use of AI coding agents (2026).
3. Google's measurement of a median Gemini text prompt (August 2025), as reported by The New Stack.
4. Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren, Making AI Less "Thirsty" (2023).
5. Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report.
6. GCSAA, Golf courses reduce water usage by 31 percent (survey of 2024 water use). Facility count from the National Golf Foundation: about 14,000 US golf facilities at the end of 2024.
7. Arjen Y. Hoekstra, The water footprint of food (2010), and the Water Footprint Network: bread, apple, coffee, egg, chocolate, beef.
8. WWF, The impact of a cotton T-shirt.
9. UN Environment Programme, Putting the brakes on fast fashion.
10. Friends of the Earth and Trucost smartphone estimate (2015), via the Water Footprint Calculator.
11. US EPA WaterSense: toilets, faucets, showerheads and outdoor watering.
12. Water Research Foundation, Residential End Uses of Water (2016), summarized by the Alliance for Water Efficiency: flushes per person per day and gallons per flush.
13. David Patterson et al., The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink (2022): the training and inference shares of Google's machine-learning energy.
14. Mistral AI's life-cycle analysis of Mistral Large 2 (2025), as reported by The Decoder.
15. Energy Efficiency of Locally Deployed LLMs: A Preliminary Quantitative GPU Power Benchmark on Consumer Hardware (2026).
16. Epoch AI, the gap between open and closed models and the best model you can run on a single consumer GPU.