Which water. Two different waters are on the meter. The first is used at the data center to cool the racks; the second is used at the power plant to make the electricity. The power plant share is usually the bigger one, and it is the one most public per-query numbers leave out.
Consumed, not withdrawn. The meter counts water that evaporates and does not return to the basin. Cooling towers consume 50 to 80% of what they withdraw; once-through cooling returns almost all of it, warmer; closed-loop chip cooling consumes close to none but shifts load to electricity. Withdrawal figures would be larger, and most of that water goes back.
What the companies publish. Google: a median Gemini text prompt uses 0.26 mL of water at the data center, about five drops, power plant water excluded. Microsoft: fleet cooling intensity 0.27 liters per kWh in 2025, and it reports replenishing more water than it withdraws. Sam Altman: about one fifteenth of a teaspoon per ChatGPT query; his later "38,000 queries per almond" line was fact-checked as overstated. Anthropic has published no water figure, so this page estimates from the others.
Where it happens matters. A liter evaporated in a water-stressed basin competes with farms and taps; the same liter in a wet region often does not. Roughly two in five US data centers sit in high water-stress areas. The places on this page are on the public record; the API does not say where a given request ran.
How to read your number. A single message is a few drops. A month of heavy work is a few showers. Both are true, and the gauge shows both so neither side gets to win the argument by picking a unit.