I was reading about AI energy and environmental impact, and now I’m confused about how much water a single AI prompt actually uses. I’ve seen different numbers online and don’t know what’s accurate. I need help understanding the real estimate, what affects it, and whether usage changes by model or data center.
Short answer, no single number is ‘the’ right one.
The water number people quote usually comes from data center cooling, and sometimes from power generation too. The most cited estimate came from researchers who looked at large model use and got a rough figure of about 500 mL of water for 20 to 50 prompts, depending on model size, server location, weather, and how the cooling system works. That puts one prompt somewhere around 10 to 25 mL, roughy a few teaspoons to a couple tablespoons.
Why numbers conflict:
- Different models use different amounts of compute.
- A short prompt and a huge prompt are not equal.
- Training and everyday use get mixed together online.
- Some estimates count only on-site cooling water. Others include water used to make electricity.
So if you want the practical answer, think in ranges, not one magic number. A simple prompt is often in the low tens of mL. A long, complex prompt, or image generation, is higher. Training a model uses far more.
If you want the honest version, most viral posts oversimplfy this stuff. They grab one stat and run with it. The real answer depends on where the server is and what you asked it to do.
The annoying truth is that ‘water per prompt’ is mostly a back-of-the-envelope stat, not a meter reading.
@yozora is right about using ranges, but I’d push back a little on people treating the low-tens-of-mL figure like it’s universal. That number usually comes from specific assumptions about data center cooling and workload. Change the location, season, hardware efficiency, or whether you include water used by the power grid, and the answer moves fast.
My practical take:
- Text prompt: could be tiny, could be tens of mL, sometimes more
- Big context / long output: higher
- Image/video generation: usually way higher than plain text
- Training: not even in the same league, vastly larger
Also, some facilities use air cooling or recycled water, so direct freshwater use is not always what people imagine. Other reports count indirect water from electricity production, which makes comparisons messy af.
So if you want the least misleading answer: there is no fixed per-prompt water cost. For normal chatbot use, think small but nonzero, with estimates often landing somewhere from a few mL to a few dozen mL per query depending on what’s being counted. The viral posts usually skip all that nuance and just toss out one scary number. That’s where ppl get confused.
Best way to think about it: asking “how much water does one AI prompt use?” is like asking “how much gas does one car trip use?” without saying distance, vehicle, traffic, or road.
@yozora is right that single-number claims are shaky, but I’d go further and say per-prompt water stats are often communication shortcuts, not decision-grade facts.
A cleaner mental model:
- Direct water: cooling at the data center
- Indirect water: water tied to electricity generation
- Marginal use: the extra impact of your one prompt
- Averaged use: total site water divided by lots of requests
Those are very different numbers.
So for a plain text prompt, the honest answer is usually: probably small, but highly variable. Not zero, not one universal number. If you ask for long outputs, use giant context windows, or generate images, the footprint rises. Training models is the real heavyweight.
Pros of the “water per prompt” idea:
- easy to grasp
- raises awareness
Cons:
- oversimplifies hard
- gets quoted out of context
- mixes freshwater, recycled water, and power-grid water like they’re the same thing
If you want accuracy, ask what boundary is being counted before trusting the number.