The key figures of an AI prompt
The reference page on generative AI's consumption: electricity, water and CO₂ per prompt, per type of use and per model. Every figure is a published value, dated and linked to its source.
Figures updated on June 30, 2026
Date of the site's latest cross-checked data review. Publication dates specific to each figure are given in the answers.
How much electricity does an AI prompt use?
An average text prompt uses about 0.3 Wh of electricity: the equivalent of an LED bulb on for 2 minutes. Google publishes 0.24 Wh per median Gemini text prompt (August 2025), Epoch AI estimates ~0.30 Wh for ChatGPT (February 2025) and OpenAI reports 0.34 Wh (June 2025). We use 0.3 Wh on a range of 0.24 to 0.4 Wh, inference only, training excluded.
Sources: Google Cloud : impact environnemental de l'inférence IA · Epoch AI : énergie par requête ChatGPT · OpenAI / Sam Altman : « The Gentle Singularity »
How much water does an AI prompt use?
An average text prompt uses about 0.3 mL of direct cooling water, i.e. five drops. Google publishes 0.26 mL per median Gemini prompt (August 2025), OpenAI 0.32 mL (June 2025). Counting the upstream water of electricity generation, the estimate rises to around 1.5 mL; Mistral's life-cycle assessment counts 45 mL per response with the whole life cycle included. Scopes differ: every figure states what it includes.
Sources: Google Cloud : impact environnemental de l'inférence IA · OpenAI / Sam Altman : « The Gentle Singularity » · Mistral AI : analyse de cycle de vie de Le Chat, avec l'ADEME (juillet 2025)
How much CO₂ does an AI prompt emit?
An average text prompt emits about 0.2 g of CO₂e. There is no official meter: we convert the electricity used with the grid's average carbon intensity (~480 g/kWh), i.e. ~0.15 g in operation and up to ~1 g over the full life cycle. Published reference points: Google reports 0.03 g per median Gemini prompt (largely decarbonised power mix), Mistral 1.14 g per response in a full life-cycle assessment.
Sources: Epoch AI : énergie par requête ChatGPT · EcoLogits / GenAI Impact · Mistral AI : analyse de cycle de vie de Le Chat, avec l'ADEME (juillet 2025)
How much does an AI-generated image use?
Generating one image uses about 6 to 12 Wh, i.e. 20 to 40 times a text prompt. The figure comes from public benchmarks (“How Hungry is AI?”, Hugging Face's AI Energy Score): no vendor publishes an official value for image generation to date.
Sources: « How Hungry is AI? » (arXiv 2505.09598) · Hugging Face : AI Energy Score / S. Luccioni (énergie image & vidéo)
How much does a reasoning prompt use?
A “reasoning” prompt (the model generates intermediate steps before answering) uses about 15 to 30 Wh, i.e. 50 to 100 times a simple text prompt. These estimates come from public benchmarks on open models; no vendor publishes an official figure for its reasoning modes.
How much does an AI-generated video use?
Generating a video clip of about 5 seconds uses on the order of 940 Wh, the equivalent of about 12 hours of video streaming. It's the heaviest item in consumer AI, and the cost grows faster than the clip's length. Measurements: Hugging Face (S. Luccioni) and public benchmarks.
Sources: Hugging Face : AI Energy Score / S. Luccioni (énergie image & vidéo) · « How Hungry is AI? » (arXiv 2505.09598)
How many prompts does AI handle every day?
ChatGPT handles about 2.5 billion prompts a day (OpenAI figure, July 2025), i.e. roughly 900 billion a year. Assuming ChatGPT holds about 80% of the market, global text generative AI handles on the order of 1.1 T prompts a year. It's that scale, not the single prompt, that makes the consumption significant.
What do vendors publish, model by model?
Only three players have published per-request figures so far: Google (Gemini), Mistral (Le Chat) and OpenAI (one global figure). Scopes differ, from the full-stack median prompt to the complete life-cycle assessment: details and comparison caveats are on the dedicated pages.
- Mistral(Mistral AI)1.14 g CO₂e · 45 mLper ~400-token response (full life-cycle assessment)Read the details
- Le Chat(Mistral AI)1.14 g CO₂e · 45 mLper ~400-token response (full life-cycle assessment)Read the details
- Gemini(Google)0.24 Wh · 0.26 mL · 0.03 g CO₂eper median text prompt (full-stack inference)Read the details
- Claude(Anthropic)No official figure publishedMethodological estimate: 0.24 to 0.34 Wh per text promptRead the details
Citing these figures
These figures can be freely reused with a link to this page; they are dated, sourced and refreshed at every data review. For the methodology converting them into everyday actions, see the Methodology page.