I was reading about AI data centers and saw claims that AI systems use huge amounts of water, but the numbers were all over the place. I’m trying to understand how many gallons AI actually uses and what affects that amount, because I need clear info for a discussion on AI’s environmental impact.
The short answer is, there is no single gallons number for ‘AI.’
Most water use comes from cooling data centers. If the site uses evaporative cooling, water use climbs. If it uses air cooling or seawater cooling, freshwater use drops.
A few rough figures people cite:
Training a large model once, estimates range from hundreds of thousands of gallons up to a few million gallons, depending on the model, chip count, location, and weather.
Per user query, some reports estimate a tiny fraction of a liter up to a few ounces of water when you average cooling use across many requests. The number moves a lot.
What affects it:
- Data center design.
- Local climate.
- Time of year.
- Chip efficiency.
- How big the model is.
- How long training runs.
- Whether you count only onsite water, or also electricity generation water.
Best way to read claims, check if they mean:
Direct water at the data center.
Indirect water from power plants.
Training only, or daily use too.
So yeah, the numbers look all over the place becuase people count diff things. If you want, I can break down one famous claim and show how they got the gallons.
The annoying but honest answer is that ‘AI’ does not have one water number. People mash together training, daily chatbot use, and power plant water, then act like it’s one neat stat. It isnt.
I mostly agree with @sterrenkijker, but I’d push back on one thing: per-query gallons can be so tiny that quoting it alone can be kinda misleading. The real impact shows up at scale, across millions of users and nonstop workloads.
Useful way to think about it:
- Small unit: one prompt or image request, often a very small amount of water on average
- Big unit: one large training run, potentially hundreds of thousands to millions of gallons
- Biggest unit: an entire data center campus over a year, which can be millions or even billions of gallons depending on cooling setup
What really drives it is not just the model. It’s where the servers are, how hot the weather is, whether the operator uses potable water, reclaimed water, or almost none onsite, and whether you count offsite electricity generation too.
So when you see a claim, ask: gallons for what exactly? One query? One model training run? One whole facility? Thats why the numbers are all over the place.
I’d add one sanity check to what @sterrenkijker said: a lot of scary numbers quietly mix water withdrawn with water consumed. Those are not the same thing.
- Withdrawn = taken in for cooling, then often returned
- Consumed = evaporated or otherwise not returned locally
That distinction can change the headline a lot.
A practical rule of thumb:
- One chat prompt: usually tiny on its own
- One image/video generation job: higher than text, often meaningfully so
- One giant model training run: can be very large
- One AI-heavy campus: the yearly total is what really matters for local water stress
I slightly disagree with the idea that the answer is always hopelessly fuzzy. If a company publishes:
- facility water use
- local climate
- cooling type
- PUE/WUE
- % reclaimed vs potable water
…you can get pretty close.
Big factors people miss:
- Air cooling vs evaporative cooling
- Region and season
- GPU utilization
- Whether inference runs 24/7
- Grid electricity water intensity
Pros for ': can make dense sustainability data easier to scan if used as a content heading or comparison label.
Cons for ': too vague alone, not helpful unless paired with clear units like gallons per query, per training run, or per year.
So the best question is not “how many gallons does AI use?” It’s “how many gallons, for which workload, at which site, over what time period?”