How much water and electricity does AI use per prompt?

The published per-prompt numbers

Only a few AI companies have published per-prompt figures, and each measured something slightly different. Put side by side:

• Google, median Gemini Apps text prompt, May 2025 data: 0.24 watt-hours, 0.26 milliliters of water, 0.03 grams of carbon dioxide equivalent. Google says the energy and carbon per median prompt fell 33 and 44 times over the prior 12 months.
• OpenAI, average ChatGPT query, stated by Sam Altman in a 2025 blog post: about 0.34 watt-hours and 0.000085 gallons of water, roughly 0.32 milliliters. No methodology was published with it.
• Mistral, one 400-token Le Chat response, lifecycle study published July 2025: 1.14 grams of carbon dioxide equivalent and 45 milliliters of water, including upstream impacts such as server manufacturing.

Why the figures disagree

Method changes the answer more than the model does. Google shows this with its own data: counting only the active AI chips gives 0.10 watt-hours and 0.12 milliliters per median prompt, while its fuller method, which adds idle capacity held for traffic spikes, host CPUs and memory, and data center cooling and power overhead, gives 0.24 watt-hours and 0.26 milliliters. Google calls the narrow version an optimistic scenario that underestimates the real footprint.

Water figures diverge further because they depend on what is counted. Google estimates water per prompt from its fleet-wide average water use per unit of energy at its data centers. Mistral follows a lifecycle standard that also counts water used upstream, which is one reason its figure is more than a hundred times higher. Neither number is wrong; they answer different questions.

All three are self-reported. Google states its data and claims have not been verified by an independent third party, and the OpenAI figure has no published method at all.

What the per-prompt numbers leave out

The published figures describe a median or average text prompt. They do not tell you the cost of a long reasoning run, an image or video generation, a coding agent that calls the model dozens of times for one task, or a document upload with a very long context. Those workloads use more compute, so a per-prompt number is the floor for heavy use, not the typical cost of everything you might do.

They also exclude training. Mistral is one of the few to publish it: training and 18 months of use of Mistral Large 2 accounted for about 20.4 thousand tonnes of carbon dioxide equivalent and 281,000 cubic meters of water as of January 2025.

Small per prompt, large in total

The reason people still worry is scale. The International Energy Agency estimates data centers used about 415 terawatt-hours in 2024, around 1.5 percent of world electricity, with the United States the largest share at 45 percent. It projects data center use to more than double to about 945 terawatt-hours by 2030, slightly more than Japan uses today, with AI the most important driver.

Local impact can be sharper than the global share suggests, because data centers cluster. Site-level data is also hard to get. On September 30, 2026, a Nebraska TV station reported that Google marked electricity and water use in its state data center reports as trade secrets, and that the redaction could be copied out: the Lincoln site reported 52.65 megawatts at peak and about 13.3 million gallons of water in a year. That is a single news report on one site, not a national figure, but it shows why per-prompt averages do not settle local questions about water and power.

How to use these numbers sensibly

• Compare like with like: check whether a figure counts only chips, the whole data center, or the full lifecycle before comparing it with another.
• Treat viral per-prompt claims without a named source and method as unverified, whether they are very high or very low.
• For your own use, the volume and type of work matter more than any single prompt. Heavy agent loops, long reasoning modes and media generation are where the cost adds up.
• For builders, the levers are the same ones that cut your bill: smaller models where they are good enough, shorter prompts, caching, and stopping agents from looping.
• For local AI, you can measure directly: a plug-in power meter on your own machine shows what your workload actually draws.

A useful next action

If you need a number for a report or a decision, cite the source and its method, for example the Google median of 0.24 watt-hours for a text prompt, and say what it excludes. If you want to reduce your footprint as a builder, start by checking whether a smaller model handles your task.