UsesLevel 3 of 6 useUse, level 3 of 6Used to calculate 35 entries, levels 3 to 6Calculated with 5 entries, levels 1 and 2Database structure 
6Territories
5Organisations
4Occupations and activities
2Machine operations
1Calculation factorsAsk a quick question
A written question and its answer of about half a page, to a medium-sized assistant served by a large operator. Projet Celsius estimate.
- It takes 63 uses to emit as much as one phone charge (8.7 gCO2e).
- The selected value rests on 4 published readings with the same scope; the database calculation, detailed in the Calculation tab, is used to vary the parameters.
- Out of scope: the manufacturing of servers and chips, model training, the internet network and the user's device.
Sentence to citeAsk a quick question, Projet Celsius estimate: 0.14 gCO2e per question, for 0.3 Wh of electricity, servers worldwide (458 gCO2e/kWh).
Estimated breakdown of emissions by item
0.14 gCO2e in total: 0.3 Wh of electricity × 458 gCO2e/kWh (servers worldwide).
Inputs of the calculation
PUE of the data centre, Read 1,000 tokens, Write 1,000 tokens
Shares calculated by the database from its items: the selected value remains the published value, whose breakdown these shares do not measure.
Origin of the value
One production measurement (Google), two estimates (Epoch AI, Microsoft) and one statement (OpenAI) have agreed since 2025; reasoning and a simple "thank you" are variants of it.
Comparison with the other uses
↓ The heaviest use in the topic, "Deep research": 45 gCO2e, 330 times more
Emissions of one use, linear scale, 7 of the 21 uses in the topic. For calculated values, servers worldwide.
Characteristics
| Unit | 1 question : short question |
|---|---|
| Default location | world, 458 gCO2e/kWh (Ember 2025, life cycle); 11 other locations to choose from |
| Counted | the computation of the answer on the servers and data centre cooling |
| Not counted | the manufacturing of servers and chips, model training, the internet network and the user's device |
| Status | estimated, 4 readings with the same scope, 9 other published values, 1 set aside |
All the characteristics
| Identifier | IA-0042, projetcelsius.com/en/ai/database/IA-0042/ |
|---|---|
| Version | 0.3 of 2 November 2026 |
| Category | Uses, a request to AI |
| Choice of the selected value | production measurements and models with the same scope; the database calculation from its factors serves as a check |
| Sources | Google, Epoch AI, Microsoft Research, OpenAI, 2025 to 2026 |
| Licence | CC BY 4.0 |
Usage parameters
Advanced settings
Result for the chosen case
For a carbon footprint
| Item | purchases of services (GHG Protocol scope 3, category 1); regulatory BEGES, category 4 |
|---|---|
| Quantity | 1 question |
| Factor | 0.14 gCO2e per question, IA-0042 version 0.3, based on the local grid (location-based), servers worldwide |
| Total | 0.14 gCO2e · not counted: the manufacturing of servers and chips, model training, the internet network and the user's device |
Breakdown by item
Quantity, unit value, contribution and nature of the data for each item
| Item | Quantity | Unit value | Contribution | Share | Nature of the data |
|---|---|---|---|---|---|
| Write 1,000 tokens | × 0.4 | 0.96 Wh | 0.38 Wh | 84 % | convention |
| Read 1,000 tokens | × 0.1 | 0.32 Wh | 0.032 Wh | 7 % | convention |
| PUE of the data centre | × 1.10 on the rest | 0.042 Wh | 9 % | convention |
The database calculation gives 0.46 Wh; the selected value, 0.3 Wh, comes from published measurements. This calculation cross-checks it and varies it with the parameters.
Entries that use it (35)
Presentation of 10 slides, Application with an agent, End-to-end study, Document assistant, AI feature of a product, Vibe, per day, AI overview of a search, Question to Claude, Office worker, 1 year, Accountant, 1 year, Human resources, 1 year, Executive assistant, 1 year, Manager or executive, 1 year, Communications officer, 1 year, Journalist, 1 year, Consultant, 1 year, Lawyer, 1 year, Researcher, 1 year, Developer, 1 year, Data scientist, 1 year, Graphic designer, 1 year, Videographer, 1 year, Customer service officer, 1 year, Salesperson, 1 year, Field-based occupation, 1 year, Teacher, 1 year, Secondary school pupil, 1 year, Student, 1 year, Free personal use, 1 year, Claude Pro, 1 year, Claude Max, 1 year, Vibe Pro, 1 year, Perplexity Pro, 1 year, Microsoft 365 Copilot, 1 year, GitHub Copilot, 1 year and
Sources of the calculation
6 publications, for this entry and those used to calculate it, with the chosen parameters.
Energy and factors
The references of the selected values and of the entries used in this calculation.
Epoch AI · February 2025
Role and limits : Epoch AI
What the publication covers. the calculation from the hardware and the assumed size of the model
What it does not cover. actual measurements
Reading by the database. A public, transparent calculation
Reference of the selected value
Scope of the figure used. GPT-4o, 500 tokens written, full scope
Limit. A public calculation, from the hardware and the assumed size of the model
Role and limits : Google
What the publication covers. chips, host, standby machines, data centre, on-site water
What it does not cover. manufacturing, training
Reading by the database. The only production measurement published by a provider
Reference of the selected value
Scope of the figure used. Median of a text prompt; chips, host, standby machines, data centre
Limit. The only production measurement published by a provider, not audited by a third party
Microsoft Research · 2026
Estimation de l'énergie des requêtes en production (Joule) ↗ (new tab)
Role and limits : Microsoft Research
Author. Microsoft Research, Oviedo et al.
What the publication covers. H100 nodes, production
Reading by the database. Calibrated on production
Reference of the selected value
Scope of the figure used. Median, 500 tokens read and 300 written, H100 nodes
Limit. Not a measurement, but calibrated on production; judges public estimates to be 4 to 20 times too high
OpenAI · June 2025
Role and limits : OpenAI
Author. Sam Altman, OpenAI
What the publication covers. an average query
What it does not cover. method, scope
Reading by the database. One sentence, no method
Reference of the selected value
Scope of the figure used. "An average query", with no published method
Limit. No published scope; consistent with Google
Watershed · Aug. 2026
Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement ↗ (new tab)
Role and limits : Watershed
Author. Bistline, Ulissi, Davis, Glidden, Joyce, Li, Mohsenin and Suh (Watershed)
What the publication covers. factors per 1,000 tokens read and written at the server before PUE (host and standby machines included), PUE separate (1.10 for a large operator, 1.56 for an unknown operator)
What it does not cover. clean measurements, the model and the hardware
Reading by the database. Default factors proposed for review (public data as of June 2026); for a large operator, taken from Google's measurement published in 2025 (Elsworth et al.), which already contains the data centre overhead; constant per token, they underestimate long contexts and agents according to the authors
Reference for the variant "Medium model, Watershed default factor"
Scope of the figure used. Default factor for 1,000 tokens read, medium model, large operator, at the server before PUE (Watershed table 2, public data as of June 2026)
Limit. A default value that Watershed derives from Google's measurement (0.24 Wh for the median text query of the Gemini app), with about 500 tokens per query and a written token costing 3 times a read token; Watershed does not give the split between read and written tokens. The 0.24 Wh already include the overhead of Google's data centre (0.02 Wh, 8%): with 375 tokens read and 125 written, the simplest reading, the factor contains it, and the PUE of 1.10 applied afterwards counts it a second time (about 9%). Values proposed for review, final at the next version of the framework
Reference for the variant "Medium model, Watershed default factor"
Scope of the figure used. Default factor for 1,000 tokens written, medium model, large operator, at the server before PUE (Watershed table 2, public data as of June 2026)
Limit. 3 times the reading factor, based on the ratio between writing and reading selected by Watershed (Patel et al., Splitwise, 2024). Same derivation and same caveat as for reading: Google's 0.24 Wh already include about 8% of data centre. Values proposed for review, final at the next version of the framework
Quick question → : Data centre PUE →
The data centre, convention C3
Scope of the figure used. building cooling and losses
Limit. This figure covers part of the calculation; it does not measure the whole entry.
Reference of the selected value
Scope of the figure used. Default factor for a large operator, from the server to the data centre; 1.56 for an unknown operator
Limit. A default value, not a measurement
Conversion into carbon
The source of the electricity factor of the selected location.
Worldwide: 458 gCO2e/kWh, life cycle.
Role and limits : Ember
Author. Ember, Our World in Data
What the publication covers. electricity generation over the life cycle, by country
What it does not cover. guarantees of origin
Reading by the database. The same source for all countries
Conversion of electricity into carbon: worldwide
Scope of the figure used. the production of electricity over the life cycle, including power plant construction and fuels
Limit. An average for the chosen country, not the electricity contract or the data centre of a specific service.
Published values
24 figures for the same operation, each within its own scope.
Electricity
0.3 Wh selected, between 0.24 Wh and 0.34 Wh for published values of the same scope
| Source | Date | Electricity | What the figure measures | Scope | Status |
|---|---|---|---|---|---|
| Quick questionselected by the database, v0.3 | 2026 | 0.3 Wh | the range of readings with the same scope: from 0.24 Wh to 0.34 Wh | same scope | estimated |
| Google, Gemini Apps ↗Measuring the environmental impact of delivering AI at Google scale | August 2025 | 0.24 Wh | Median of a text prompt; chips, host, standby machines, data centre | same scope | measured |
| Epoch AI ↗How much energy does ChatGPT use? | Feb. 2025 | 0.3 Wh | GPT-4o, 500 tokens written, full scope | same scope | estimated |
| Microsoft Research ↗Estimation de l'énergie des requêtes en production (Joule) | 2026 | 0.31 Wh | Median, 500 tokens read and 300 written, H100 nodes | same scope | estimated |
| OpenAI, Sam Altman ↗The Gentle Singularity | June 2025 | 0.34 Wh | "An average query", with no published method | same scope | published |
| MIT Technology Review ↗We did the math on AI's energy footprint | May 2025 | 0.032 Wh to 1.86 Wh | An answer from Llama 3.1, from 8 to 405 billion parameters, to varied questions; GPU measured then doubled for the rest of the server | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 0.052 Wh | Llama 3.1 8B, short question (100 tokens read, 300 written), data centre included | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 0.207 Wh | GPT-4.1 nano, short question (100 tokens read, 300 written), data centre included | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 0.423 Wh | GPT-4o of March 2025, short question (100 tokens read, 300 written), data centre included | other scope | estimated |
| EcoLogits and compar:IA ↗EcoLogits and compar:IA | 2026 | 0.5 Wh to 1.6 Wh | Per model, 400 tokens written, tokens read ignored | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 0.871 Wh | GPT-4.1, short question (100 tokens read, 300 written), data centre included | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 0.95 Wh | Claude 3.7 Sonnet, short question (100 tokens read, 300 written), data centre included | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 2.23 Wh | Llama 3.1 405B, short question (100 tokens read, 300 written), data centre included | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 2.33 Wh to 17.2 Wh | GPT-5, a question of medium length, from minimal reasoning to high reasoning | other scope | estimated |
| de Vries, Joule ↗The growing energy footprint of artificial intelligence | 2023 | 2.9 Wh | "10 times a Google search", based on a sentence about the cost | other scope | outdated |
| Database checks and calculations | |||||
| Projet Celsius, with Watershed factors ↗Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement | Oct. 2026 | 0.46 Wh | 100 tokens read and 400 written, default factors, data centre included | check | computation |
CO2e
| Source | Date | CO2e | What the figure measures | Scope | Status |
|---|---|---|---|---|---|
| Google, Gemini Apps ↗Measuring the environmental impact of delivering AI at Google scale | August 2025 | 0.03 gCO2e | Median question to Gemini, May 2025, electricity counted according to Google's contracts, chip manufacturing included | other scope | computation |
| Mistral AI, LCA of Large 2 ↗Our contribution to a global environmental standard for AI | Jul. 2025 | 1.14 gCO2e | An answer of 400 tokens, life cycle, manufacturing included | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 17 gCO2e | DeepSeek-R1, a long query of 10,000 tokens read and 1,500 written, electricity from the grid of the data centre's region, excluding manufacturing | other scope | estimated |
Water
| Source | Date | Water | What the figure measures | Scope | Status |
|---|---|---|---|---|---|
| Google, Gemini Apps ↗Measuring the environmental impact of delivering AI at Google scale | August 2025 | 0.26 mL | On-site water consumption, median of a text prompt | same scope | measured |
| OpenAI, Sam Altman ↗The Gentle Singularity | June 2025 | 0.32 mL | "An average query", with no method | other scope | published |
| Li, Yang, Islam and Ren, UC Riverside ↗Making AI Less « Thirsty » (Communications of the ACM) | 2025 | 17 mL | GPT-3 in Microsoft data centres in the United States, a medium-length response: 2.2 mL on site and 14.7 mL to produce the electricity | other scope | estimated |
| Mistral AI, LCA of Large 2 ↗Our contribution to a global environmental standard for AI | Jul. 2025 | 45 mL | A response of 400 tokens, water over the whole life cycle | other scope | estimated |
| Jegham et al. ↗How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference | 2025 | 200 mL | DeepSeek-R1, a reasoning model, on DeepSeek servers in China, a long query of 10,000 tokens read and 1,500 written: on-site cooling and water for electricity | other scope | estimated |
| Li et al., UC Riverside ↗Making AI Less Thirsty | 2023 | 500 mL | For 10 to 50 GPT-3 responses, including power plant water | other scope | discarded |
The published figures do not all count the same scope; the scope of each figure is written next to it. Why do the estimates diverge?
Calculation assumptions
Quantities, nature and source of each assumption
Read 1,000 tokens : × 0.1
The tokens of the question, 100 by default (convention C1)
Write 1,000 tokens : × 0.4
The tokens of the response, 400 by default (convention C1)
Methodology conventions
With their alternatives
| The standard query | A query counts 100 tokens read and 400 written |
|---|---|
| The data centre | Server electricity is multiplied by 1.10 for the data centre |
| The default location | Without other information, the servers are in the world, at the world average electricity |
| Electricity, location-based, on a life cycle basis | The kilowatt-hour is that of the servers' country, on a life cycle basis, from a single source for all countries |
| Two significant figures | A calculated value is rounded to two significant figures, a published value is reproduced with the precision of its source; the calculation keeps full precision |
| The formula | the measured value, adjusted to the case settings in the proportion in which they change the database calculation, then the CO2e: electricity (Wh) multiplied by the carbon intensity of the server location (gCO2e/kWh), divided by 1,000 |
| Calculate a case without an entry | The formula, the default values and a step-by-step example |
Uses in the hub
Charts, figures, analyses and articles that cite this entry
- Analysis"An email written by an AI is a bottle of water" →
- ChartElectricity use of a question, an image, a video →
- ChartOne question, 4 ways of counting CO2e →
- ChartThe water of a question, from 0.26 to 50 mL depending on scope →
- Blog articleHow much electricity does a ChatGPT query use? ↗
- Blog articleWhich AI to choose to pollute the least? ↗
- Blog articleHow can you reduce the impact of your AI use? ↗
History
Entry in version 0.3 of 2 November 2026, in the database in version 0.4
- 12 November 2026Press source (Washington Post, 519 mL of water for an email, 2024) removed.A press article repeating a calculation by UC Riverside; the scientific source (Li et al.) remains in the published values. No figure in the database changes.
- 11 November 2026Recomposition of the database, nature : Factors calibrated on non-production services becomes Watershed factors derived from Google's measurement, gap mainly due to the token split.The recomposition (0.46 Wh) does not change; only its justification is corrected.
- 3 November 2026Status : measured becomes estimated.The selected value, 0.3 Wh, is the median of a production measurement (Google, 0.24 Wh), two estimates (Epoch AI, Microsoft) and a statement (OpenAI): it is a database calculation from measured values; the value does not change.
- 2 November 2026Designation : A question to an AI assistant becomes Text query, conversational assistant.More precise designation: the object first, the specification after the comma, the quantity or period at the end; the former title remains searchable.
- 2 November 2026Rows : model size: medium, small and large (no value) becomes removed.No value exists for a small or large model; an option with no value does not exist in public.
- 2 November 2026Designation : Text query, conversational assistant becomes Ask a quick question.The name in people's own words.
- 2 November 2026Short title : Text query becomes Quick question.In people's own words.
- 2 November 2026Address : requete-texte becomes poser-une-question-rapide.The old address redirected.
- 2 November 2026Functional unit : 1 query becomes 1 question.In people's own words.
- 31 October 2026Selected value : 0.46 Wh becomes 0.3 Wh.Production measurements and modelling take precedence over recomposition.
- 31 October 2026Sentence : one sentence becomes two sentences.The citable sentence with no missing value.
- 31 October 2026Creation of the database, in its first version (0.1).
