AI does use a significant amount of water, but the scale depends heavily on where and how the underlying data centers are built and powered. Most of this water is not in the model itself, but in cooling the servers and generating the electricity they consume.

Where the water is used

Most AI-related water use comes from:

  • Cooling data centers that train and run AI models.
  • Power plants (especially fossil-fuel and some nuclear) that use water for steam and cooling to generate the electricity those data centers need.

Unlike irrigation, much of this water is withdrawn as clean, treated water and then lost to evaporation rather than being returned to rivers or groundwater in usable form.

How much water are we talking about?

Estimates vary widely, but they show that AI can be very thirsty at scale:

  • One analysis projected U.S. AI alone could require on the order of hundreds of billions of gallons of water annually for cooling if growth continues aggressively, comparable to the indoor water use of tens of millions of households.
  • Case studies of large tech companies show their data-center water use increasing by double-digit percentages in a single year as AI workloads expand.
  • At the user level, some research has suggested that a multi-question chat session with a large model may indirectly consume the equivalent of roughly a small bottle of drinking water once you fold in both data-center and power-plant cooling, though there is methodological debate and uncertainty around headline numbers.

The key point: a single query is negligible, but billions of queries and large training runs add up to a notable water footprint regionally and globally.

Why location and design matter

AIā€™s water impact is not uniform; it is highly place-dependent :

  • Many data centers are sited in already water-stressed regions, where extra industrial withdrawals compete with households, agriculture, and ecosystems.
  • Some cities have discovered that a small number of large data centers can account for a sizable share of their municipal water use, triggering public concern and even legal or political pushback.
  • Where grids rely heavily on thermal power plants (coal, gas, nuclear), the indirect water footprint from electricity use can rival or exceed the water used at the data center itself.

By contrast, siting AI infrastructure in cooler climates, using air or seawater cooling, and running primarily on renewables can sharply reduce freshwater withdrawals.

Are the scary statistics exaggerated?

There is an active debate around how AI water numbers are presented:

  • Critics argue that some popular statistics either ignore training costs (making AI look too efficient) or count huge upstream flows at power plants in ways that make AI look worse than it is in practice.
  • Researchers and sustainability advocates stress that while some headlines over-simplify, the broader concernā€”large new loads in already stressed basins and limited transparency from companiesā€”is real.

So the short answer is: yes, AI can use a lot of water at scale, but the exact ā€œper promptā€ number people share online is often a rough, context-dependent average rather than a precise, universal figure.

What can be done about it?

Several levers can reduce AIā€™s water footprint without abandoning the technology:

  • Technical efficiency
    • More efficient chips and model architectures reduce energy per computation, which in turn cuts cooling needs and associated water use.
* Advanced cooling (liquid cooling loops, heat reuse, hybrid airā€“water systems) can shrink water withdrawals per unit of compute.
  • Infrastructure and policy
    • Locating new AI data centers in regions with ample water and resilient ecosystems, instead of arid or drought-prone basins.
* Reusing non-potable water (like treated wastewater) instead of relying primarily on drinking-quality supplies.

* Requiring disclosure of site-level water use so communities and regulators can see the impact and set limits or efficiency standards.
  • Usage choices
    • Prioritizing AI for applications with strong social or environmental value and avoiding wasteful, always-on or novelty workloads that chew up compute for little benefit.

TL;DR

AI does use a lot of water when you add up all the data centers and power generation behind it, and this is becoming a serious environmental and local- planning issue in some regions. The impact is not fixed, though: better hardware, cleaner energy, smarter cooling, and transparent governance can meaningfully shrink the water footprint of AI over the next few years.