How Much Water Does an AI Prompt Use? Why Estimates Differ

There is no single water-use number that accurately describes every AI prompt. A short text reply, a long reasoning task and image generation are different workloads, and studies can count different parts of the infrastructure supporting them.

The right question is what an estimate measures before deciding whether another estimate contradicts it.

What Google measured

In its 2025 inference study, Google reported a median Gemini Apps text prompt using 0.24 watt-hours of energy and 0.26 millilitres of water under its methodology.

That is a company-reported result for a defined service and measurement approach. It is not a universal conversion factor for every AI model, nor an estimate of the full lifetime impact of every activity associated with AI. “Median” also describes a middle observation, not necessarily the total footprint divided by the number of requests.

Why a different study can produce a different answer

The research paper Making AI Less “Thirsty” examines water consumption and withdrawal associated with AI, including distinctions between direct and indirect water use. Its modelling should be read with its assumptions, systems and dates in view.

A comparison becomes misleading if one number includes water associated with electricity generation and another counts a narrower operational boundary. The same problem occurs when training a model is compared with running a single inference request.

Before sharing a dramatic number, ask whether it refers to one prompt, an entire conversation, a training run or a whole data centre.

Five details every useful estimate needs

These questions do not imply that narrow measurements are worthless. They establish which comparison the measurements can legitimately support.

Efficiency and total demand can move differently

A service can use less energy per task while doing enough additional tasks to increase total demand. The International Energy Agency’s updated Energy and AI analysis projects data-centre electricity use roughly doubling from 2025 to 2030.

That is a projection about data centres, not a measured water bill for one chatbot. It is useful context for why a small per-request number cannot settle the infrastructure question by itself.

How to evaluate a provider’s environmental claim

Look for a published method, the reporting year, the workload and the exclusions. Check whether a claim about improvement uses the same basis in both periods. Treat a target as a target, and a forecast as a forecast.

For a reader, a sensible response is to use the service deliberately and ask providers for better disclosure. There is no need to pretend that multiplying every prompt by one viral statistic produces a precise personal water footprint.

Our cloud-computing comparison and Alphabet score provide broader company context. OpenAI’s score and Anthropic’s score address conduct beyond the environmental cost of a single request.

Sources checked on 24 September 2026. Ownership and policies can change after this review.