The calculation model
Each entry is calculated from smaller entries, from machine operations to territories, and each figure says what it rests on: published readings, our conventions or our assumptions.
Database structure
Each level is calculated from those below it. Infrastructure, published by others, stands apart and overlaps with the territories without adding to them.
The map of the database
- Calculation factorsthe foundation of the calculation7 entries
- Machine operationswhat the server calculates11 entries, from 0.032 Wh to 17 Wh
- Usesone request to the AI39 entries, from 0.002 Wh to 670 Wh
- Occupations and activitiesone person, one year38 entries, from 340 Wh to 470 kWh
- Organisationsone organisation, one year18 entries, from 1.1 kWh to 10 MWh
- Territoriescities, territories, worldwide services19 entries, from 450 kWh to 1.4 TWh
What the 6 levels and the infrastructure are made of
- Calculation factorsthe grid kilowatt-hour by country, water, the data centre, the GPU hour, manufacturing: values taken from readings
- Machine operationstokens read or written by model class, an image, a second of video, a minute of audio: values taken from readings, in server electricity
- Useswhat a person asks once, in their own words (an email, a page, an image): its size in human units, converted into tokens
- Occupations and activitiesan occupation, a personal use, a project entrusted to AI: a task basket, each task with its frequency and the share done with AI
- Organisationsits occupations (the structure of its jobs), the share that uses AI, and the AI services it provides for itself
- Territoriescities, territories, populations, global services: headcounts and shares of users
- Infrastructuredata centres, companies, training, chip manufacturing, as published: they overlap with the world calculated from the bottom up, they never add to it
Calculation rules
A figure is the sum of what composes it, multiplied by the frequency and the share done with AI, then converted into CO2e according to the servers' location.
The 6 rules
- A recipethe electricity of a block is the sum, over its components, of their quantity multiplied by their electricity, then multiplied by the data centre and the reasoning
- A basketan activity is the sum, over its tasks, of their frequency, multiplied by the number of periods in the year, by the share done with AI (depending on adoption) and by the electricity of the task
- The settingsthe location of the servers, the model class, reasoning, adoption, tooling and sector pass down the whole lineage; the duration (one year, one month, one day) is read at the top; a setting never creates a block
- A measurement from abovewhen a block has a published measurement and a recipe, the measurement is selected; the recipe makes it vary with the settings and the gap is displayed. For "Quick question", the recomposition gives 0.46 Wh and the selected value, 0.3 Wh, comes from production measurements
- What a figure rests oneach contribution takes the weakest nature along its path: readings alone, conventions of the database, or at least one assumption of Projet Celsius; the three shares are shown on each calculated block
- CarbonCO2e is the electricity multiplied by the kilowatt-hour factor of the location, divided by 1,000; a published carbon figure is used as it stands
Units of use
Uses are counted in words, pages or e-mails; their equivalent in tokens is a convention of the database.
The 22 units
- A word1 word, 1.6 token.
- An instruction25 words, 40 tokens.
- A question60 words, 100 tokens.
- An answer250 words, 400 tokens.
- An e-mail150 words, 240 tokens.
- A page400 words, 640 tokens.
- An article of 1,000 words1,000 words, 1,600 tokens.
- A document of 30 pages12,000 words, 19,200 tokens.
- A novel80,000 words, 128,000 tokens.
- The Lord of the Rings480,000 words, 768,000 tokens.
- A minute of speech150 words, 240 tokens.
- An hour of meeting11,000 words, 17,600 tokens.
- A web page1,000 words, 1,600 tokens.
- A web page excerpt200 words, 320 tokens.
- A photo sent to an AI1,100 tokens.
- A minute of voice listened to by an AI600 tokens.
- A minute of voice spoken by an AI1,200 tokens.
- A line of code12 tokens.
- A code function480 tokens.
- A code file3,600 tokens.
- A spreadsheet of 200 rows9,300 tokens.
- The instructions of an automated workflow500 tokens.
End-to-end calculation
SME of 50 people, 1 year · servers in France
Occupations and activities
- Developer, 1 year20 %
- Videographer, 1 year14 %
- Consultant, 1 year12 %
- Office worker, 1 year9 %
- Manager or executive, 1 year7 %
- 9 others29 %
- Outside "Occupations and activities"9 %
Uses
- Web search with an AI12 %
- Code suggestion11 %
- Summarise a meeting11 %
- Quick question6 %
- Write an email5 %
- 24 others55 %
Machine operations and data centre
- Read 1,000 tokens36 %
- Write, with reasoning18 %
- PUE of the data centre9 %
- Generate an image9 %
- Second of video8 %
- 6 others20 %
1.1 kgCO2e in total: 27 kWh of electricity × 41 gCO2e/kWh (servers in France). Each bar splits the emissions between the entries of a category.
Inputs of the calculation
PUE of the data centre, Read 1,000 tokens, Write, with reasoning, Write 1,000 tokens, Generate an image, Second of video, Transcribe one minute, Write an email, Quick question, Presentation of 10 slides, Clip of 5 seconds, Web search with an AI, Code suggestion, Summarise a meeting, Manager or executive, 1 year, Consultant, 1 year, Developer, 1 year, Graphic designer, 1 year, Videographer, 1 year, Office worker, 1 year, Field-based occupation, 1 year
Correspondence with the life cycle assessment
A block is a unit process, a task is the functional unit, and the location of the servers is the context that sets the kilowatt-hour factor.
The 7 correspondences
- A blocka unit process: incoming flows, a transformation, an outgoing flow, a written scope (counted, not counted)
- The first two levelsthe background data: electricity and water per kilowatt-hour, the data centre, tokens, an image at the server
- One taskthe functional unit: an e-mail, a page, an image, a video
- An activity, an organisation, the worldthe product system: tasks multiplied by frequencies, shares of use and headcounts
- The location of the serversthe context: it changes the kilowatt-hour factor, and all the CO2e with it
- The conventionsthe parameters of the model: written, dated, with their alternatives
- The infrastructurethe inventories published by others: used as they stand, compared, never added to the calculations
