Source library

453 sources cited by the hub, each read in the original document. For each one, the content that uses it. How we choose them is described in our methodology.

Institution 23Scientific study 11Company 9Association 6Dataset 6Press 2Other, no entry 19

The canonical reference

A single format for all sources, and a stable number: the one used in the hub's citations.

[14]stable number Elsworth et al. (Google),authors (publisher) Measuring the environmental impact of delivering AI at Google scale,title August 2025.date arxiv.org/abs/2508.15734link to the document

In the text:Google measures 0.26 mL for one question to Gemini [14] : the water evaporated to cool its data centres.

The structure

By topicBy typeMost cited

Electricity19 sources

Water9 sources

Institution 2

1IEA, Energy and AI, April 20257

56Sispea (Office français de la biodiversité), report on water services, 2024 data, June 20261

Company 4

14Elsworth et al. (Google), Measuring the environmental impact of delivering AI at Google scale, August 20257

15Microsoft, 2026 Environmental Sustainability Report, Data Fact Sheet, July 20267

17Mistral AI, with Carbone 4 and ADEME, Our contribution to a global environmental standard for AI, July 20256

21Sam Altman (OpenAI), The Gentle Singularity, June 20253

Scientific study 1

30Li, Yang, Islam and Ren, Making AI Less "Thirsty", 20232

Dataset 1

35FAO, SDG indicator 6.4.2 (level of water stress), via Our World in Data1

Press 1

31Washington Post, A bottle of water per email: the hidden environmental costs of using AI chatbots, September 20241

Climate28 sources

Hardware10 sources

Uses16 sources

Society28 sources, of which 19 have no entry

A source is listed under each topic of the content that cites it: no. 14 appears under Water, Climate and Uses. The figure on the right counts the hub's content that cites it.

14

Source entryCompanyGoogleAugust 2025

Topics:Water Climate Uses Electricity

Measuring the environmental impact of delivering AI at Google scale

Elsworth et al. (Google), August 2025arxiv.org/abs/2508.15734 ↗

Strengths

"The only published measurement over a full inference scope."
A measurement in production: active accelerators, processor and memory, idle machines, data centre overhead.
It also gives the narrow approach (0.10 Wh): the wider scope more than doubles the value.

Its limitations

Measurement published by the provider itself, not audited by a third party.
Valid for Gemini; it cannot be transposed as it stands to other assistants.
For water, the narrowest scope of the published figures: 170 times less than the life cycle assessment by Mistral AI.

The 8 values recorded in this document

Value	What it measures	Where in the document	Checked on
0.24 Wh	Electricity used by a text query, median question to Gemini, May 2025 · reference value	Results section, table 1	21 Oct. 2026
0.24 Wh	Energy of the median question to Gemini Apps (text), May 2025 · reference value	Results section, figure 3	21 Oct. 2026
0.10 Wh	The same query, narrow approach (accelerators only)	Section 4.1, table 1	24 Oct. 2026
0.26 mL	Water consumed per question, Google Gemini (median) · reference value	Section 4.2, table 2	21 Oct. 2026
1.15 L/kWh	Water consumed per kWh, Google data centres serving language models	Section 3.3	21 Oct. 2026
0.03 gCO2e	Emissions per median written question to Gemini	Section 4.2, table 2	21 Oct. 2026
94 g/kWh	Google electricity in 2024, according to its contracts	Section 3.3, p. 6	24 Oct. 2026
345 g/kWh	Google electricity in 2024, according to the grid	Section 3.3, p. 6	24 Oct. 2026

Where it is cited in the hub · 16 items

Aisle of servers in a Google data centre, a technician at work between the racks

AnalysisWhy do estimates of AI's impact differ?

A Google data centre: a long beige building fronted by a row of cooling towers, in front of hills

Where does this figure come from?"An email written by an AI is a bottle of water"

A developer at a lectern runs the Claude Code agent, projected on a large screen in front of a room

AnalysisWhat is the footprint of my AI use over a year?

AnalysisHow do you account for AI in a Bilan Carbone®?

A green circuit board carrying 4 chips under cooling plates, connected to red, blue, yellow and green pipes

AnalysisCan AI be decarbonised?

A lit filament LED bulb hanging in the centre of a metal frame, against a dark background

ViewpointShould we feel guilty about using AI?

A small desktop computer with an open case, its graphics card with 2 fans taking up almost the whole space, sitting on a wooden table next to a golden Nvidia DGX Spark box

AnalysisDoes AI running on your own computer use less?

The headquarters and data centre site of OVHcloud in Roubaix, under a blue sky, with heat extraction chimneys along the boulevard

AnalysisIs a French AI better for the climate?

Servers mounted in racks in a data centre, black fronts with blue and green indicator lights on

ViewpointDo we really know how much electricity ChatGPT uses?

A technician in a hard hat and safety vest walks between large white pipes marked "Industrial water" in a data centre

AnalysisDoes AI really use water, and how?

Google's data centre at Eemshaven, in the Netherlands, among wind turbines

AnalysisIs AI using more or less energy over time?ChartWater for one question, from 0.26 to 50 mL depending on scope ChartOne question, 4 ways of counting CO2e

Key figure0.24WhElectricity used by a text query to Gemini Key figure0.3WhElectricity used by a written question to an AI Key figure0.26mLWater for one question to Gemini

Cite this source

[14] Elsworth et al. (Google), Measuring the environmental impact of delivering AI at Google scale, August 2025. https://arxiv.org/abs/2508.15734

For software and AI

@misc{elsworth2025,
  author = {Elsworth and others},
  title  = {Measuring the environmental impact of delivering AI at Google scale},
  year   = {2025}, month = {aug},
  url    = {https://arxiv.org/abs/2508.15734}
}