AI infrastructure myths
Does AI really consume water, or is that a false claim?
AI can consume water, but the viral version of the claim is often too simple. The real answer depends on where the model runs, how the data center is cooled, what electricity powers it, and how much utilization the hardware gets.
Published . Updated . 7 min read
Key takeaways
- The claim is not completely false: some data centers use water directly for cooling, and some electricity generation uses water indirectly.
- Per-prompt water numbers are usually estimates, not universal facts. Location, season, grid mix, cooling design, and model size all change the answer.
- RightOne.ai’s answer is efficiency: avoid sending easy tasks to overpowered routes when a smaller model can do the work well.
When people say “AI consumes water,” they are usually pointing at the physical infrastructure behind inference and training. AI does not drink water. Data centers may use water to remove heat, and the power plants supplying electricity may also consume water. That makes the claim directionally real, but it is easy to exaggerate.
Two kinds of water use
Direct water use
Some data centers use evaporative cooling or cooling towers. In those facilities, water can be consumed directly as heat is removed from servers. Other facilities rely more on air cooling, closed-loop liquid systems, seawater cooling, or local climate advantages. The same model can have a different water footprint depending on where it runs.
Indirect water use
Electricity generation can also consume water, especially in thermal power plants. If a data center draws from a grid with water-intensive generation, the indirect water footprint can be higher. If the grid mix is different, the estimate changes.
Why viral per-prompt numbers can mislead
- They often mix training, inference, cooling, and electricity into one number without explaining assumptions.
- They may assume a specific region or cooling method and then present it as universal.
- They rarely account for batching, hardware utilization, model size, prompt length, or whether the answer used a small or frontier model.
- They can ignore the alternative: a human workflow, a traditional search workflow, or an inefficient AI workflow may also consume energy and water indirectly.
So is it a false claim?
No. It is not false that AI infrastructure can consume water. The false part is pretending there is one universal amount per question. Water use is a system property: facility design, grid mix, model route, token count, and utilization all matter.
Where routing helps
If every user turn goes to a maximum-effort frontier model, the infrastructure burden grows. If simple tasks use lighter routes and hard tasks get stronger routes only when needed, the same product can provide better quality-per-resource.