No, One AI Image Doesn’t Drain a Bottle of Water. Here’s the Real Number.

You’ve seen the claim. Every AI image you generate supposedly guzzles a bottle of water, or ten gallons, or enough to fill a bathtub. It’s the kind of stat that travels fast because it feels right. The problem is that almost none of it survives contact with the actual research.

Here’s the single most important thing to know, and it’s the part every viral post skips: there is no study that tells you how much water one AI image uses. Not one peer-reviewed study has measured it. So any specific water figure you’ve read, ours included, is calculated backward from an energy number. It’s an estimate, not a measurement, and anyone quoting you an exact amount is guessing.

What researchers have actually measured is energy. The most-cited study, Luccioni, Jernite, and Strubell’s “Power Hungry Processing”, found that generating 1,000 images averaged about 2.9 kilowatt-hours, or roughly 2.9 watt-hours per image. That’s about what it takes to charge a smartphone, and that’s the high end. Smaller, lower-resolution models have been measured at around 0.3 watt-hours per image, ten times less. The number swings enormously based on the model, the resolution, and how many steps it runs.

Now here’s where the water estimate comes from, and why you should hold it loosely. Take that measured energy, multiply by a rough cooling-plus-generation factor of about 5 liters per kilowatt-hour, and you land at roughly 15 to 60 milliliters per image, a shot glass or two. But read that as a back-of-the-envelope calculation we ran, not a finding anyone published. No study established it. It rests on two assumptions, the energy figure and the conversion factor, and both move with location, model, and time of day.

So where do the scary numbers come from? Mostly from one study about something else. The famous “bottle of water” figure traces to UC Riverside’s “Making AI Less Thirsty,” which found that running 10 to 50 GPT-3 text prompts could use about 500 milliliters of water. That’s text, across dozens of prompts, on an older model. It gets misquoted as the cost of a single image or a single query. The “10 gallons per image” claim has no academic source at all. For scale on the company-disclosed side, OpenAI’s Sam Altman said in 2025 an average ChatGPT query uses about 0.3 milliliters, and Google put a median Gemini text prompt at about 0.26 milliliters. Drops, not bottles.

None of this means AI is free of a water footprint. It means the footprint sits somewhere other than where the viral posts point. The genuinely thirsty step is upstream, in the factory. Making a single semiconductor can take 28 to 100 liters of water, far more than any number of inferences. And the cooling story is shifting fast. Microsoft said in 2024 its new data center designs will use zero-water evaporative cooling, and Google runs a closed seawater loop in Finland, which quietly undermines the “every image drains a lake” narrative for newer facilities.

The honest takeaway isn’t “AI uses no water.” It’s that no one has measured exactly what one image costs, the best estimate is a shot glass at most, the real water goes into building the chips, and anyone quoting you a bottle per image is recycling a number that was never about images in the first place.