Live
AI & ML

How much energy does one AI chat use? What Google, OpenAI and Mistral have actually disclosed

Abstract illustration: a lightning bolt inside a water drop on a dark green background

Every few weeks a new number goes viral about how much electricity or water one AI chat burns. Most of those numbers are estimates by outsiders. Only a handful of AI companies have published their own figures, and they measured different things in different ways. Here is what Google, OpenAI and Mistral have actually disclosed, and how to read it.

Google: 0.24 watt-hours for a median Gemini prompt

Google’s figure is the most detailed. For the median text prompt in the Gemini Apps, measured in May 2025, it reported 0.24 watt-hours of energy, 0.03 grams of CO2-equivalent emissions and 0.26 millilitres of water, about five drops. Google compared the energy to watching TV for less than nine seconds.

The method counts more than the AI chips themselves. It includes idle machines kept ready for traffic spikes, the host CPUs and memory, and data-centre overhead such as cooling. Measure the chips alone and the figure falls to 0.10 Wh, which Google itself calls an optimistic scenario that underestimates the real footprint. Google also said the energy per median prompt fell 33-fold between May 2024 and May 2025. Two caveats from Google: the figures haven’t been independently verified, and they won’t hold as models and usage change. The full method is in a technical paper on arXiv (2508.15734).

Chart: Google median Gemini prompt 0.24 Wh and 0.26 mL water; OpenAI average ChatGPT query 0.34 Wh and about 0.32 mL; Mistral Le Chat 400-token response 1.14 g CO2e and 45 mL water on a life-cycle basis
Company-reported figures. Methods differ, so they are not directly comparable. Graphic: prompt/power

OpenAI: 0.34 watt-hours, one sentence long

OpenAI’s number comes from a June 2025 blog post by chief executive Sam Altman, “The Gentle Singularity,” which said the average ChatGPT query uses about 0.34 watt-hours, “about what an oven would use in a little over one second,” and about 0.000085 gallons of water, roughly a fifteenth of a teaspoon. Altman didn’t publish a method, so we don’t know what was counted.

Mistral: a full life-cycle analysis

French developer Mistral took a different approach in July 2025. With the consultancy Carbone 4 and the French ecological transition agency ADEME, it published a life-cycle analysis that includes making the servers, not just running them. Its marginal figure for one 400-token reply from its Le Chat assistant was 1.14 grams of CO2-equivalent and 45 millilitres of water. Training Mistral Large 2 over 18 months produced 20.4 kilotonnes of CO2-equivalent and consumed 281,000 cubic metres of water. Mistral called the study a “first approximation,” because the environmental cost of manufacturing GPUs had to be estimated.

How to read the numbers

Mistral’s water figure is much larger than Google’s because its method counts much more, including manufacturing and other upstream impacts, not only data-centre cooling. That is why these figures can’t be lined up on one scale.

Some things the numbers do tell you. A single text chat is small next to everyday uses of electricity. The total is not small, because these services handle billions of prompts. Long reasoning sessions, image and video generation, and agents that run for minutes are not “median prompts” and almost certainly cost more. None of the three companies has published per-prompt figures for those heavier tasks.

Disclosure is still voluntary in the U.S. and Canada, and full published methods, like Google’s and Mistral’s, are the exception. If you want better data, the most useful question to ask any provider is simple: what did you count?

More from prompt/power: Data centre ballot measures: what voters decided · Canada’s Budget 2026 AI spending · SoftBank’s power-equipment loss

// Author
Jessie Isberg

Latest from prompt/power

  1. Jensen Huang’s resignation: what Nvidia’s 8-K says and leaves outNov 16
  2. Quebecor’s offer for TELUS says nothing about TELUS’s AI factories. Here’s what’s at stakeNov 16
  3. Can an AI chatbot be your therapist? Where the rules stand in the U.S., Canada and beyondNov 16
  4. Nvidia’s first earnings without Jensen Huang: the numbers to watch on TuesdayNov 16
  5. SoftBank wrote off 39% of its AI power equipment spendingNov 16

Leave a Reply

Your email address will not be published. Required fields are marked *