Skip to content
by Projet Celsius
Projet Celsius
AI Impact Database
A technician checks servers in an aisle of a data centre
GuideAI Impact Database · version 0.4

Methodology and conventions

How the database counts, selects and dates each value.

Scope

An operation counts the computation on the server and the data centre; the rest is documented separately.

  1. Close-up of a chip on its gilded mountDocumented separatelyManufactureof servers and chips1.31tCO2eThe motherboard of 8 H100 GPUs, according to NvidiaManufacturing a server
  2. A training server roomDocumented separatelyTrainingof the model552tCO2eper training run, GPT-3, 2020Training a model
  3. Servers in their rackCounted in an operationComputation on the serverchips, host, spare capacity900Whper GPU hourH100 hour
  4. Aerial view of a data centreCounted in an operationData centrecooling and losses× 1.10on the electricity of the serversPUE of the data centre
  5. Network cables on a switchDocumented separatelyNetworkinternet, up to the userNo entry in the database
  6. A phone open on a conversation with an AIDocumented separatelyDevicethe user's deviceNo entry in the database
Counted and not counted, in detail
Counted, for operations and what derives from them

The computation of the answer on the servers (the chips, the host, the spare machines) and the data centre that houses them (convention C3). CO2e comes from the electricity and the location of the servers.

Not counted, and documented separately

The manufacture of servers and chips, model training, the network and the user's device. Each page says so in full; where a source measures them, they have their own entries, which can be compared and are not silently added together: Manufacturing a server, Chip manufacturing, Training a model.

Statuses

From the most reliable to the most fragile

Each value carries a status that says what it rests on; an entry with no selected value is not published.

Database entries by status

150 entries published, from the most reliable status to the most fragile

  1. measureda production measurement
    7 entries
  2. estimateda calculation from measurements
    114 entries

    ↑ 3 entries out of 4 are models or calculations by the database from measured values

  3. publishedtaken as is, without calculation
    26 entries
  4. conventionset by the database, written down, dated
    3 entries
  5. working valuenot yet verified by the hub
    0 entry
Source: AI Impact Database, version 0.4.
Definition of each status
measureda production measurement, with the same scope as the database, published by an operator or a research team; 7 entries
estimateda model or calculation by the database from measured values; 114 entries
publisheda figure taken as is from its source, without calculation (the sector, the grid factors); 26 entries
conventiona value set by the database, written down, dated, with its alternatives (the data centre, the cache, the typical query); 3 entries
working valuea value taken from a source that has not yet gone through the hub's verification chain; 0 entry

For figures published by others, 6 words: measured, estimated, calculation, published, discarded, outdated; and 3 scopes: same scope, other scope, check (the check is the recomposition of the database from its factors).

Selected value, range, location

Example: quick question

Among the published figures, measurements with the same scope take priority; the range runs from the lowest to the highest of these readings, and the servers are in the default world.

Published values and selected value

Quick question, electricity in Wh; 4 readings with the same scope

  1. Google, Gemini AppsAugust 2025 · measured
    0.24 Wh
  2. Epoch AIFebruary 2025 · estimated
    0.3 Wh
  3. Microsoft Research2026 · estimated
    0.31 Wh 0.16 to 0.6
  4. OpenAI, Sam AltmanJune 2025 · published
    0.34 Wh
  1. Selected valueby the database, for a quick question
    0.3 Wh
  2. Projet Celsius, with Watershed's factorsOctober 2026 · check
    0.46 Wh
10 other readings, with a different scope, are on the entry page.

Emissions by server location

Quick question, 0.3 Wh, in gCO2e

  1. Poland589 gCO2e per kWh
    0.18 gCO2e
  2. World458 gCO2e per kWh · default location
    0.14 gCO2e
  3. United States384 gCO2e per kWh
    0.12 gCO2e
  4. European Union210 gCO2e per kWh
    0.063 gCO2e
  5. France41 gCO2e per kWh
    0.012 gCO2e
Grid electricity on a life cycle basis, Ember 2025; 11 countries and zones in the database.
The 5 selection rules
The ruleproduction measurements and models with the same scope take precedence; the database's own calculation from its factors (tokens, data centre) serves as a check, and the gap is displayed
The rangethe minimum and maximum of the readings with the same scope; failing that, the range published by the source; failing that, none. Never an invented range
The locationthe world by default (458 gCO2e/kWh); 11 countries and zones; a single source for all (Ember 2025, life cycle, according to the grid of the location, known as location-based), because a ratio between two countries only makes sense within a single source
Displaya calculated value is rounded to two significant figures, a published value is reproduced with the precision of its source; the calculation keeps full precision (convention C8)
The sectora published total (a country, a company, the world) is reproduced as it stands, with its year; it is not converted into CO2e with an average factor and is not added to calculated values

Water

Cooling, electricity, life cycle · convention C6

Water is counted within three scopes, each with its own readings, which are never added together without saying so.

Water consumed per kWh, on site and at the power plant

Published readings, in litres per kWh; the on-site reference value is 1.15 L per kWh, not yet applied to uses

On site, for coolingwater consumed by the data centre · On-site water

  1. Microsoft's global fleetMicrosoft, 2024
    0.3 L
  2. Average of US data centres up to 2023Lawrence Berkeley National Laboratory, 2024
    0.36 L
  3. Google data centres that serve language modelsGoogle, 2025
    1.15 L

At the power plant, to generate the electricitywater consumed by generation type · Water for electricity generation

  1. Ground-mounted photovoltaicMacknick et al., NREL, 2012
    0.004 L
  2. Combined-cycle gas plant with cooling towerMacknick et al., NREL, 2012
    0.78 L
  3. Once-through cooled nuclear plantMacknick et al., NREL, 2012
    1.02 L
  4. Nuclear plant with cooling towerMacknick et al., NREL, 2012
    2.54 L
  5. Coal plant with cooling towerMacknick et al., NREL, 2012
    2.6 L
  6. Water consumed to generate the electricity of US data centres in 2023Lawrence Berkeley National Laboratory, 2024
    4.52 L

    ↑ 13 times the water consumed on site per kWh, according to the same source

  7. HydropowerMacknick et al., NREL, 2012
    17 L

Over the whole life cycle, manufacturing includedanother unit, published by a provider · Le Chat, LCA

45 mLper response, Mistral AI, LCA of Large 2

The database shows each reading in its own column, with no water value per operation until the rule (convention C6) is settled.

Conventions

What the database sets itself, when no source settles the matter, is written with its date, its reason and its alternatives.

  1. C1

    The standard queryA query counts 100 tokens read and 400 written

    100 read, 400 written
  2. C2

    The working day and yearA working year counts 220 days and 44 weeks, a personal year 365 days; the task lists of occupations give, as a median, about 2 requests to AI per working day at low adoption, 6 at standard adoption, 21 at intensive adoption, excluding code suggestions

    220 days; about 2, 6 and 21 requests per day
  3. C3

    The data centreServer electricity is multiplied by 1.10 for the data centre

    × 1.10
  4. C4

    The cacheA token re-read from the cache costs 10% of a token read

    10 %
  5. C5

    The default locationWithout other information, the servers are in the world, at the world average electricity

    world
  6. C6

    The three watersOn-site water, power plant water and life cycle water are never added together without saying so; the factor remains to be settled

    to be settled
  7. C7

    Electricity, location-based, on a life cycle basisThe kilowatt-hour is that of the servers' country, on a life cycle basis, from a single source for all countries

    Ember and Our World in Data, 2025
  8. C8

    Two significant figuresA calculated value is rounded to two significant figures, a published value is reproduced with the precision of its source; the calculation keeps full precision

    2 figures
  9. C9

    Calibration against production measurementsThe tokens of the tasks are multiplied by a coefficient: the quick question recomposed from the machine's factors is brought back to its selected value, that of the consensus of published figures (one production measurement by Google, two estimates by Epoch AI and Microsoft, one statement by OpenAI). The code agent hour is not calibrated: its tokens, mostly context re-read from the cache, do not resemble those of a question

    × 0.656 (0.3 Wh selected for 0.458 Wh recomposed)
Reason and alternatives for the 9 conventions

  1. C1 The standard query : 100 read, 400 written

    A query counts 100 tokens read and 400 written.

    Why: Providers do not all say the same thing about a "query"; we set a single definition, with input and output kept separate. Alternatives: 410 and 90 (Watershed's calibration); 1,000 and 400. Since 31 October 2026.

  2. C2 The working day and year : 220 days; about 2, 6 and 21 requests per day

    A working year counts 220 days and 44 weeks, a personal year 365 days; the task lists of occupations give, as a median, about 2 requests to AI per working day at low adoption, 6 at standard adoption, 21 at intensive adoption, excluding code suggestions.

    Why: Readable benchmarks, drawn from the baskets and not set alongside them. Alternatives: 365 days for personal use; the profiles of the previous version (5, 20 and 80 queries per day), replaced by the task lists. Since 2 November 2026.

  3. C3 The data centre : × 1.10

    Server electricity is multiplied by 1.10 for the data centre.

    Why: Watershed's default factor for a large operator, consistent with recent data centres; Watershed uses 1.56 for an unknown operator; as Watershed's factors are derived from a Google measurement that already contains its data centre, this × 1.10 may count about 8% twice, a gap below the precision of the factors. Alternatives: × 1.72 from the chips to the full service at Google, never to be combined. Since 31 October 2026.

  4. C4 The cache : 10 %

    A token re-read from the cache costs 10% of a token read.

    Why: Agents re-read their context at every step; no provider publishes this cost. Alternatives: bounds of 1% and 25%. Since 31 October 2026.

  5. C5 The default location : world

    Without other information, the servers are in the world, at the world average electricity.

    Why: We almost never know where a query runs. Alternatives: the United States, where most large assistants run. Since 31 October 2026.

  6. C6 The three waters : to be settled

    On-site water, power plant water and life cycle water are never added together without saying so; the factor remains to be settled.

    Why: Published figures mix three different waters. Alternatives: 1.3 or 3.67 L/kWh for France depending on the source. Since 31 October 2026.

  7. C7 Electricity, location-based, on a life cycle basis : Ember and Our World in Data, 2025

    The kilowatt-hour is that of the servers' country, on a life cycle basis, from a single source for all countries.

    Why: A ratio between two countries only makes sense within a single source. Alternatives: market-based, with purchase contracts, displayed alongside when published. Since 31 October 2026.

  8. C8 Two significant figures : 2 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.

    Why: The precision displayed must not exceed that of the sources. Since 31 October 2026.

  9. C9 Calibration against production measurements : × 0.656 (0.3 Wh selected for 0.458 Wh recomposed)

    The tokens of the tasks are multiplied by a coefficient: the quick question recomposed from the machine's factors is brought back to its selected value, that of the consensus of published figures (one production measurement by Google, two estimates by Epoch AI and Microsoft, one statement by OpenAI). The code agent hour is not calibrated: its tokens, mostly context re-read from the cache, do not resemble those of a question.

    Why: Watershed's per-token factors are themselves derived from Google's measurement (0.24 Wh for the median text query of the Gemini app, related to a query of about 500 tokens, a written token costing 3 times a read token). Recomposed with the database's split (100 tokens read, 400 written), the quick question gives 0.458 Wh; the gap with 0.3 Wh is mainly due to this split. The selected value, 0.3 Wh, also draws on Epoch AI, Microsoft and OpenAI. A single coefficient for now, to be refined by factor family. Alternatives: the raw factors, without calibration, under the "raw factors" setting. Since 2 November 2026.

Calculating a case without an entry

Formula and calculator

A use without an entry is calculated from the nearest entry, with the formula below; the calculator does this for the quick question.

  • Electricity of the case = selected value × (database calculation for this case ÷ database calculation for the default case)
  • CO2e = electricity × location factor ÷ 1,000
50
Model size

Emissions of the case

0.62 gCO2e

15 Wh of electricity, servers in France

50 × 0.3 Wh = 15 Wh0.015 kWh × 41 gCO2e per kWh = 0.62 gCO2e

The default values: the standard query (C1), 220 days; about 2, 6 and 21 requests per day (C2), the data centre (C3), the cache (C4), the world (C5). The grid factors: Grid electricity.

Reading by an AI

An AI finds the entry, reads its page, calculates its case with the formula above, and cites the database version.

  1. 1

    Find the entryThe table and its synonyms

  2. 2

    Read the entryThe value, the unit, the scope (counted, not counted), the location, the status, the sources

  3. 3

    CalculateWith the parameters (location, quantity, duration) and the formula above

  4. The filesThe JSON and Markdown of each entry, the database CSV, llms.txt

The 3 steps and the files
Find the entrythe table and its synonyms: "prompt", "query", "a question to ChatGPT" lead to IA-0042; each entry has a stable identifier and a short link, projetcelsius.com/en/ai/database/IA-0042/
Read the entrythe value, the unit, the scope (counted, not counted), the location, the status, the sources: the entry page of each page, or its JSON
Calculatewith the parameters (location, quantity, duration) and the formula above; a case is cited with its parameters, the database version (0.4) and the engine version (0.3.1)
The filesthe JSON and Markdown of each entry, the database CSV, llms.txt

poser-une-question-rapide.json

{
  "id": "IA-0042",
  "version": "0.3",
  "date": "2026-11-02",
  "designation": "Poser une question rapide",
  "unite": "1 question",
  "statut": "estimé",
  "valeur": {
    "electricite": {
      "valeur": 0.3,
      "basse": 0.24,
      "haute": 0.34,
      "unite": "Wh",
      "fourchette": "releves",
      "publie": false
    }
  }
  …
}
The database output formats

Versions

Version 0.4 · 13 November 2026

The changelog

A value never changes without its changelog line, and its page flags it for 30 days.

Changelog lines per day

221 changes and 24 address changes recorded since the database was created

  1. 31 October 2026version 0.1, creation of the database
    3 lines
  2. 1st November 2026
    21 lines
  3. 2 November 2026
    152 lines
  4. 3 November 2026
    4 lines
  5. 11 November 2026
    13 lines
  6. 12 November 2026
    43 lines
  7. 13 November 2026version 0.4
    9 lines
Engine 0.3.1. Each line gives the before, the after, the reason and the source.
Contents of a version and citation
Current version0.4 of 13 November 2026, engine 0.3.1
What a version freezeseach entry carries its version and date; a value never changes without its changelog line (before, after, why, source), and the page flags it for 30 days. A calculated value also changes with the entries it depends on: the citation of each entry therefore carries the database version
Citing a versionProjet Celsius (2026). AI Impact Database, version 0.4 of 13 November 2026. https://projetcelsius.com/en/ai/database/. CC BY 4.0 licence.
Published versions0.1, 0.3, then 0.4

Photos: header, Google, press photo, cited document ; manufacturing and life cycle, Mister rf, Wikimedia Commons, CC BY-SA 4.0 ; training, Amazon (AWS), aboutamazon.com, cited document ; server, NOIRLab/NSF/AURA/T. Slovinský, Wikimedia Commons, CC BY 4.0 ; data centre, Chad Davis, Wikimedia Commons, CC BY 2.0 ; network, J. Barande, École polytechnique, Wikimedia Commons, CC BY-SA 2.0 ; device, Jernej Furman, Wikimedia Commons, CC BY 2.0 ; power plant, formulanone, Wikimedia Commons, CC BY-SA 2.0.

Projet Celsius, environmental strategy consultancy

We support companies that account for AI in their Bilan Carbone®

Visiter Projet Celsius ↗
  • Institut Pasteur
  • SNCF Gares & Connexions
  • Radio France
  • Institut Curie
  • Institut Polytechnique de Paris
  • ESSEC Business School
  • Qualiopi, certified process