How much does a ChatGPT prompt use?
An average ChatGPT query uses 0.34 Wh of electricity and 0.32 mL of water, the only figure OpenAI has ever published, stated by Sam Altman on June 10, 2025. That is two minutes of an LED bulb and about six drops of water. It is also the thinnest of the three vendor figures on record: one sentence in a blog post, with no method note, no boundary and no model named.
Figures published on June 10, 2025, per average query (a stated figure, with no published method) · Sources: OpenAI / Sam Altman : « The Gentle Singularity » · Epoch AI : énergie par requête ChatGPT · OpenAI / TechCrunch : ChatGPT ~2,5 milliards de prompts/jour (2026) ; part de l'IA générative ~80 % (Demandsage / Statista). Calcul : ChatGPT ~900 Md/an ÷ 0,8 ≈ 1 100 Md pour toute l'IA générative. · NBER / OpenAI : « How People Use ChatGPT » (Chatterji et al., 2025) : 700 M d'utilisateurs hebdomadaires, ~18 Md de messages/semaine, part professionnelle 47 % → 27 % (juin 2024 → juin 2025)
Where does the 0.34 Wh figure come from?
From a blog post by Sam Altman, “The Gentle Singularity”, published on June 10, 2025. One sentence: the average query uses about 0.34 watt-hours and 0.000085 gallons of water, which works out at 0.32 mL. Nothing has followed since: no method note, no model tested, no measurement window, no split between short and long queries. Epoch AI had reached a nearby order of magnitude in February 2025, roughly 0.30 Wh for a typical GPT-4o query, worked out from model size and hardware. The two agree. That makes the order of magnitude credible, not proven.
What does the figure cover, and what does it leave out?
The word “average” does all the work here. A ChatGPT query can be a one-word hello or a hundred-page document to summarise. A reasoning model writes intermediate steps before answering and uses an order of magnitude more. Without the distribution of real usage, nobody can tell what that average hides. Three items sit outside the count, exactly as they do at Google: model training, server manufacturing, and the water used upstream to generate the electricity. Mistral's life-cycle assessment with ADEME, which counts all of it, lands at 45 mL of water per response against OpenAI's 0.32 mL. Both can be right: they are not measuring the same thing.
Does a ChatGPT prompt use more than a Gemini prompt?
On paper yes: 0.34 Wh against 0.24 Wh for Gemini's median text prompt (Google, August 2025). The 40% gap says less than it looks. Google publishes a median measured across its own fleet, with a technical paper setting out the boundary. OpenAI publishes an average, with nothing behind it. A median and a mean do not compare, least of all on a distribution where a few very long queries pull the mean up. The useful point sits elsewhere: three independent methods land between 0.24 and 0.34 Wh. The order of magnitude for a text prompt holds, even though none of these numbers is directly comparable to the next.
What does a year of ChatGPT prompts add up to?
ChatGPT handles about 2.5 billion prompts a day (OpenAI figure, July 2025), so roughly 900 billion a year. Multiply by 0.34 Wh and you get about 310 GWh of electricity; by 0.32 mL, about 292,000 m³ of water. For scale: the yearly electricity of 72,000 French homes (4,300 kWh each, ADEME), and enough water to fill 5,800 private swimming pools (50 m³ each, US EPA). OpenAI publishes no CO₂ figure at all. Converted at the average grid carbon intensity (480 g/kWh), the total lands near 150,000 tonnes of CO₂e a year, the annual footprint of 16,500 people in France. This covers text inference only: no training, no images, no video, no idle servers.
A ChatGPT prompt in everyday actions
The figures published for ChatGPT, converted with the site's reference actions (every action value is sourced, see the Sources page):
- 1 hour of LED light = 29 ChatGPT prompts
- 1 hour of TV = 206 ChatGPT prompts
- 1 glass of water = 781 ChatGPT prompts
- One 5-minute shower = 156,250 ChatGPT prompts
The figure prices a query. It says nothing about what the query is worth. Two minutes of a light bulb tells you nothing about the time saved, or about where AI belongs in an organisation. That is a different trade: the GHIS studio works with small and mid-sized companies on exactly that question.