Summary
- Google's disclosure prints 0.24 watt-hours, 0.03 grams of CO2 equivalent and 0.26 millilitres of water for a median Gemini Apps text prompt; only the watt-hours were measured, across a May 2025 window.
- The grams are that energy times a 2024 fleetwide grid carbon intensity, and the millilitres are energy minus overhead times a 2024 fleetwide water usage effectiveness of 1.15 litres per kilowatt-hour — so the figure everyone quotes as a cost of a prompt is a property of a distribution and a boundary.
Three numbers arrived in one sentence, and the sentence was about a single prompt: 0.24 watt-hours of energy, 0.03 grams of carbon dioxide equivalent, 0.26 millilitres of water, "or about five drops" (the disclosure). The comparison it invites is the obvious one. My question, weighed and costed, five drops at a time.
The word doing the work is median, and it is the right word for the job. A mean gets dragged around by the outliers, and the technical paper says so in as many words: "the arithmetic mean is highly sensitive to these extreme values, making it an unrepresentative measure of typical user's impact" (preprint).
Except that what was published is not a median of prompts. To build the figure the authors rank models by their average energy per prompt, construct a cumulative distribution of prompts across that ranking, and then "identify the model that serves the 50-th percentile prompt". The percentile is taken over traffic. The wattage is measured on a machine. The published number is what one serving computer draws when it is the one carrying the middle of the month's prompts — a defensible thing to publish, and a different object from the middle prompt of your afternoon. Half of all prompts sit on the far side of that line, and the month's model mix sits inside the number.
Where this method is strongest is at the meter. The population is named as "all LLM models serving the Gemini app, including all supporting models for scoring, ranking, classification, and other prompt routing tasks"; those models are mapped to job IDs and machine IDs; power is read at the tray power-supply level (technical paper). The functional unit is one serving AI computer, its accelerator trays and one host tray. I obtained that paper as truncated text: the sections deriving the grid factor and the embodied share did not come through, so what follows rests on the parts that did.
The decomposition is the part that argues with the brochure. Active accelerator 0.14 Wh, fifty-eight per cent. Host CPU and DRAM 0.06 Wh, twenty-five per cent. Provisioned idle machines 0.02 Wh, ten per cent. Data-centre overhead 0.02 Wh, eight per cent. That is my arithmetic on their rows, and it closes: 0.14 + 0.06 + 0.02 + 0.02 = 0.24, and 0.14 ÷ 0.24 = 58.3 per cent against the printed fifty-eight.
The idle line is the real argument: "The energy consumed by these idle AI computers, which are essential for service uptime, must be factored into the total energy footprint." Count active accelerators alone and the same prompt costs 0.10 Wh, 0.02 g and 0.12 mL, a method the disclosure itself calls "an optimistic scenario at best". A figure that counts only busy silicon describes a busy machine, not a deployed one. Whoever wrote this chose the harder number.
Then the second number stops being a measurement. "Emissions per prompt was estimated based on energy per prompt, and applying Google's 2024 average fleetwide grid carbon intensity." The table splits the 0.03 into a market-based Scope 2 component of 0.023 and a Scope 1 plus 3 component of 0.010; my sum is 0.033, printed to two decimal places as 0.03.
The third number is arithmetic all the way down, and it is worth doing the sum. Water per prompt is total energy minus overhead energy, multiplied by water usage effectiveness. That effectiveness is the consumptive variant, ISO Category 2, "calculated as water input minus water returned", and "For both 2023 and 2024, Google's WUE Category 2 value was 1.15 L/kWh".
Only evaporation counts: "On average, Google consumes 80 per cent of the water withdrawn." Now take the published inputs — 0.24 Wh is 0.00024 kWh, times 1.15 gives 0.000276 litres, or 0.276 mL; the overhead share of 0.02 Wh contributes 0.00002 kWh, times 1.15, or 0.023 mL; subtract and 0.253 remains. Printed: 0.26. That reproduction is mine, and it is a check rather than a complaint. The five drops are not metered anywhere. They are the energy figure wearing a 2024 constant — and only on-site cooling water. The water spent generating that electricity is outside the boundary.
The clock is where the three numbers part company permanently. Energy per query is month-specific; the multipliers are not. "A May 2024 estimate uses May 2024-specific numbers for energy/query, while using annual fleetwide average numbers for PUE, WUE, and grid emissions factors." So the watt-hours carry a 2025 machine and the grams and millilitres carry 2023 and 2024 fleet averages. That is a stated choice, not a hidden one. It also means a movement between two publications can come from the constant rather than from the computer — which is precisely the question to ask of the next edition.
And the corpus disagrees with itself in one place. For the narrow boundary, the section prose gives 0.15 mL of water and 0.01 gCO2e, while the table in the same paper and the disclosure page both give 0.12 mL and 0.02 gCO2e. Same case, two answers, both in print. The comparison is muddier still: the narrower approach also subsampled the top ten per cent most energy-efficient data centres, so the headline 2.4-to-1 gap between comprehensive and narrow bundles a boundary difference together with a utilisation difference, and cannot be read as the price of including idle machines alone.
What sits outside is named, and the omissions matter more than the arithmetic. Training is deferred — "We leave the measurement of AI model training to future work" — along with data storage, external networking and end-user devices. The measured unit is a serving estate in one month, not a product, not a lifecycle, and not a customer's workload.
The disclosure closes its own loop in two sentences that deserve to travel with the headline. The findings "do not represent the specific environmental impact for all Gemini App text-generation prompts nor are they indicative of future performance", and "The data and claims have not been verified by an independent third-party." Nothing in the material read names which Gemini model versions or which accelerator generation served the measured traffic, and nothing states how the traffic distribution is sampled, audited or re-derived between publications.
The consequence is not that the numbers are wrong. It is that they are not portable. Ask any per-prompt figure four things and the fog clears: which statistic — mean or median — and over what; which population of models and hardware; which constants, from which year, and whether the energy and the constants share a year; and which boundary, on-site water only or water and power together. Two providers' per-prompt numbers, compared without those four answers, compare two distributions, two model mixes and two sets of prior-year constants. The operator's own card in this directory records a company; the numbers belong to a method.
Five drops is lovely writing and honest arithmetic. It is not a measurement of your prompt. The most useful thing in this disclosure is the sentence that tells you so.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






