How much electricity does one AI prompt really use?

ecogpt.coPublished Oct 19, 2026 · 3 min read

Google measured 0.24 Wh for a median Gemini prompt; OpenAI says 0.34 Wh for ChatGPT. Here is what those numbers include, what they leave out, and why a long document can cost a hundred times more.

Google's 0.24 Wh is a fleet-wide measurement that counts idle capacity and overhead; its narrower active-chip method gives 0.10 Wh.
OpenAI's 0.34 Wh is one sentence in a CEO blog post with no methodology. Epoch AI's independent estimate for GPT-4o is about 0.3 Wh.
A typical prompt costs about as much as running a 10 W LED for two minutes. A 100,000-token input costs around 40 Wh.
Only three providers have published any per-prompt figure. The rest of the index is estimated, and labelled as such.

The two numbers everyone quotes

In August 2025 Google published the most detailed figure the industry has: a median Gemini Apps text prompt used 0.24 watt-hours of electricity, emitted 0.03 grams of CO₂e and consumed 0.26 millilitres of water, measured across its fleet in May 2025. Two months earlier Sam Altman had written that an average ChatGPT query uses about 0.34 watt-hours, "about what an oven would use in a little over one second".

The numbers look similar. They are not the same kind of number.

What Google's method includes

Google's "comprehensive" method counts the energy of the chips doing the work, the host machines around them, the idle capacity kept warm for traffic spikes, and the data-centre overhead (cooling, power conversion). It is the way you would measure a fleet if you wanted the total bill divided by the number of prompts served. Google also reports a narrower "active chip" figure, 0.10 Wh, which is how most academic benchmarks measure. The difference between the two, 2.4×, is overhead that nobody serving at scale can avoid.

Critics at The Verge and elsewhere pointed out what the figure omits: it is a median, not a mean, so the long tail of heavy prompts is invisible, and the water number counts only on-site cooling. Those caveats are fair. The measurement is still the best disclosed figure in the index.

What OpenAI's number includes

Nobody knows. The 0.34 Wh figure comes with no definition of "average", no model, no date and no boundary. Epoch AI's independent estimate for a typical GPT-4o query, built from the model's likely size and the hardware it runs on, lands at about 0.3 Wh, which is reassuringly close. Epoch also showed the catch: a 10,000-token input costs about 2.5 Wh and a 100,000-token input about 40 Wh. Median prompts are cheap. Pasting in a contract is not.

Mistral counts something else again

Mistral's life-cycle assessment reports carbon and water per 400-token response (1.14 gCO₂e and 45 mL) rather than energy, and it counts the whole life cycle: hardware manufacturing, training amortised over the model's life, and the water used to generate the electricity. The index records 0.6 Wh for Mistral as an estimate, because a 123-billion-parameter dense model producing a 400-token answer sits above a median Gemini prompt and well below an image.

What a watt-hour means in practice

0.24 Wh runs a 10 W LED bulb for about 90 seconds.
1,000 median prompts use about as much electricity as 20 smartphone charges.
40 Wh, the cost of a 100,000-token prompt, is a smartphone charged three times.

The per-prompt number is small. The reason data centres used 415 TWh in 2024 is the number of prompts, the training runs behind the models, and everything else those buildings do.

How the index handles the rest

Twenty-seven of the thirty companies in the index publish nothing per prompt. For each of them we estimate a figure from the closest measured analogue (0.3 Wh for a closed frontier model, less for mixture-of-experts designs that activate fewer parameters, 0.04 Wh for small models, 1.6 to 3.6 Wh for an image) and print the working next to the number. The methodology page lists every rule. An estimate never raises a score: the efficiency pillar rewards disclosure, not our arithmetic.