AI Data Centres Are Thirsty. But How Thirsty, Exactly?
Protests are spreading across the American Southwest over water use at AI data centres. The real numbers are murky, and tech companies aren't making it easier to find them.
Key points
- Protesters in Texas and Arizona, and across New Mexico, are demonstrating against the water use and energy consumption of AI data centres.
- Data centre processors can reach internal temperatures of up to 176 degrees Fahrenheit, requiring active cooling to stay functional.
- Some claims online say a single chatbot query uses 500 ml of water; others say AI's water footprint is near zero. Neither figure is well supported.
- Major tech companies publish water consumption figures that are often incomplete and inconsistent, making independent verification difficult.
Signs reading "Water for people not AI" are appearing at protests from Texas to Arizona. The anger is real. Whether the scale matches the fury is harder to say.
What is actually happening inside a data centre?
These buildings are enormous rooms packed with specialised processors, the chips that run AI software. Those chips get hot. Very hot. Processors can reach up to 176 degrees Fahrenheit, for the same reason your laptop gets warm when it's working hard. Left unchecked, that heat destroys equipment.
Many data centres use water-based cooling systems. The principle is the same one your body uses when you sweat: water absorbs heat, then carries it away as it evaporates. Depending on the facility, that process can drain millions of gallons from local water supplies each year. We looked at one unconventional cooling approach in our 20 August story on recycled-water systems.
Should local communities be worried?
Probably yes, especially in drought-prone regions. The Southwest already faces chronic water scarcity, so any large new demand on local aquifers matters. The data, though, is thin.
Figures circulating on social media range wildly. Some articles, as Ars Technica AI reported, claim a single chatbot query consumes 500 ml of water, roughly the contents of a standard water bottle. Other commentators dismiss the concern entirely. Neither camp offers clean sourcing.
The companies building these facilities publish annual environmental reports, but those figures are often incomplete, use different accounting methods, and don't break out AI workloads specifically. That inconsistency makes it nearly impossible for outside researchers or regulators to build a reliable picture.
| What we know | What we do not know |
|---|---|
| Data centres use water for cooling | Exact litres used per AI query |
| Processors reach up to 176°F | How much is AI versus other computing |
| Protests active in TX, NM, AZ | Verified per-facility water draw |
| Tech firms publish some data | Whether that data is comparable across companies |
What happens next?
Pressure is building from two directions. Local communities and state officials in water-stressed regions are asking harder questions before approving new facilities. Some tech companies are also exploring air cooling and on-site water recycling, though neither is yet standard practice. Where those facilities end up is its own fight, one we covered on 27 August.
For people who live near a proposed or existing data centre, the most practical step is checking whether your state's environmental permitting process requires water-use disclosures. Many don't.
What this coverage keeps running into is a transparency problem more than a technology one: companies could publish consistent, facility-level water data but don't, which means communities are left arguing from estimates while the buildings go up.
Common questions
Does using ChatGPT or another AI chatbot actually cost water?
Most likely yes, in a small amount, because the servers running those tools use water-cooled systems. Reliable figures for how much per query don't currently exist.
Can anything be done about it?
Some facilities already use air cooling or treat and recycle water on site. Regulators can require those approaches as a condition of planning permission, but few currently do.
