Ask how much a single chatbot query costs the planet and the answers scatter wildly: anywhere from a tenth of a watt-hour to several watt-hours, and from five drops of water to roughly 45 millilitres. None of those figures is invented. Each measures a different thing, on a different model, inside a different data centre — and that is exactly the problem.
- Google’s own numbers: a median Gemini text prompt uses 0.24 Wh, emits 0.03 g of CO₂ equivalent and consumes 0.26 ml of water, roughly five drops.
- Mistral’s lifecycle study: a 400-token reply in Le Chat costs 1.14 g of CO₂ equivalent and 45 ml of water, because it counts manufacturing and upstream impacts Google leaves out.
- The aggregate is what scales: data centre electricity demand is set to more than double to about 945 TWh by 2030, more than Japan consumes today.
What the labs actually publish
Google remains the most forthcoming. In a report dated 21 August 2025, backed by a technical paper, it calculated that the median text request to its Gemini app draws 0.24 Wh — framed as the energy of watching television for under nine seconds — alongside 0.03 g of CO₂ equivalent and 0.26 ml of water. The company also claims that between May 2024 and May 2025, energy per median prompt fell 33-fold and its carbon footprint 44-fold. It concedes that no independent third party has verified any of it.
OpenAI has published nothing comparable as of October 2026. The closest thing to a disclosure is a line in Sam Altman’s personal blog from June 2025:
The average query uses “about 0.34 watt-hours” of energy and “about 0.000085 gallons of water” — roughly a fifteenth of a teaspoon.
No methodology accompanies the figure, which is why Google’s technical paper notes that without a published method such numbers cannot be compared with anything.
Mistral took a third route. In July 2025, working with the consultancy Carbone 4 and the French agency ADEME and reviewed by two external firms, it published a lifecycle analysis of Mistral Large 2: 1.14 g of CO₂ equivalent and 45 ml of water for a 400-token response in its Le Chat assistant. It reports no watt-hours and describes the work as “a first approximation”.
Four reasons the figures never converge
1. Where the boundary sits. Google demonstrates this with its own data. Count only the AI chips while they work and a median prompt uses 0.10 Wh and 0.12 ml of water. Add standby machines, CPU, memory and building cooling and the same prompt reaches 0.24 Wh and 0.26 ml — a 2.4-fold swing. Google’s paper admits that the absence of consensus on scope makes estimates for similar tasks differ “by an order of magnitude”.
2. Which water gets counted. Google measures the water evaporated to cool its data centres. Mistral follows ISO 14040/44 lifecycle rules and folds in earlier impacts such as server manufacturing. Pengfei Li, Shaolei Ren and colleagues at the University of California, Riverside and the University of Texas at Arlington go further and include the water burned by power stations, using a US average of 3.14 litres per kWh generated.
3. Which model, which task. Google measured median text prompts in May 2025, in the use phase only, excluding training. Mistral measured one 400-token answer from one model. Li’s team studied GPT-3, estimating a 500 ml bottle per 10 to 50 medium-length replies. Independent testing also shows that lab benchmarks running a single request at a time can overstate inference energy by 4 to 20 times compared with production systems that batch and cache; reasoning models can draw roughly 30 times more than smaller ones, and agentic requests can fire off 5 to 50 model calls at once.
4. Where and when it runs. Efficiency “varies as much in space as in time”, as Li and his co-authors put it, depending on outdoor climate and the grid’s generation mix. Hold a prompt’s electricity constant and its emissions vary 34-fold between a coal-heavy grid and a hydropower-heavy one — around 100-fold once accounting choices diverge. US grids hosting data centres average about 307 g of CO₂ equivalent per kWh today.
Small per prompt, large in aggregate
Both statements are true at once. The International Energy Agency calculated in April 2025 that global data centre electricity consumption will more than double to roughly 945 TWh by 2030, “more than Japan’s total electricity consumption today”, with AI-optimised capacity more than quadrupling. On water, Li’s study projects that global AI demand will require between 4,200 and 6,600 million cubic metres withdrawn in 2027, and 380 to 600 million cubic metres actually consumed.
The bill is not only in queries. Mistral estimates that training and serving its model over 18 months to January 2025 produced 20,400 tonnes of CO₂ equivalent and 281,000 m³ of water. Corporate-level frameworks suggest training can account for 32% to 56% of total AI emissions in a given case, with a possible range of 5% to 83%. Data centres are expected to absorb 9% to 17% of US electricity by 2030, up from about 5% in 2025.
The per-prompt figure has become a rhetorical weapon for both camps: companies cite it to argue the impact is negligible, critics cite older numbers to claim every query costs a bottle of water. Neither argument survives contact with the data. Per-query efficiency is improving fast, but query volumes and data centre construction are growing faster still. What is missing is not another estimate but a shared standard and independent audit. Mistral has shown that a methodologically reviewed disclosure is possible; OpenAI’s silence and Google’s lack of external verification say a great deal about who benefits from keeping the arithmetic murky.