UsesLevel 4 of 6 occupation or activityOccupation or activity, level 4 of 6Used to calculate 9 entries, levels 5 and 6Calculated with 11 entries, levels 1 and 3Database structure 
6Territories
5Organisations
3Uses
2Machine operations
1Calculation factorsDeveloper, 1 year
One working year of a developer who uses AI: about 35,000 code suggestions, 460 generated functions, 140 web searches with AI, and 7 other tasks. Projet Celsius estimate.
- Code and agents account for 71% of emissions.
- About 2,100 requests to AI over the year, or 9.4 per working day, not counting 35,000 code suggestions.
- 3.4 times the AI footprint of an office employee, for the same use.
- An estimate: the frequencies and the share done with AI are our assumptions.
Sentence to citeDeveloper, 1 year, Projet Celsius estimate: 2 kgCO2e per role per year, for 4.3 kWh of electricity, servers worldwide (458 gCO2e/kWh).
Breakdown of emissions by item
2 kgCO2e in total: 4.3 kWh of electricity × 458 gCO2e/kWh (servers worldwide).
How the calculation is built
One working year of a developer who uses AI, task by task: code suggestions, functions, reviews, questions, debugging; with agents, hours of coding agent. The frequencies and the share done with AI are assumptions based on the Stack Overflow survey and usage studies, and can be adjusted.
Comparison with other occupations and activities
Characteristics
| Unit | 1 person, 1 year : 220 days |
|---|---|
| Default location | world, 458 gCO2e/kWh (Ember 2025, life cycle); 11 other locations to choose from |
| Counted | the tasks done with AI over a working year, including data centre cooling |
| Not counted | server manufacturing, model training, the person's device and the tasks done without AI |
| Status | estimated, 1 other published value |
All the characteristics
| Identifier | IA-0071, projetcelsius.com/en/ai/database/IA-0071/ |
|---|---|
| Version | 0.3 of 2 November 2026 |
| Category | Occupations and activities, one person, one year |
| Choice of the selected value | database calculation from the entries this one is made of, at the quantities of the unit |
| Sources | calculation of the database |
| Licence | CC BY 4.0 |
Usage parameters
Advanced settings
Tasks and frequencies
Frequency and share done with AI, task by task · typical use
Adjust each task
| Task | Frequency | Share with AI | Per year | Electricity | Share |
|---|---|---|---|---|---|
| Code suggestion | 35,200 | 2.4 kWh | 56 % | ||
| Generate a function | 462 | 540 Wh | 12 % | ||
| Have code reviewed | 110 | 140 Wh | 3 % | ||
| Quick question | 990 | 300 Wh | 7 % | ||
| Conversation of 10 exchanges | 88 | 310 Wh | 7 % | ||
| Query a PDF | 34.3 | 160 Wh | 4 % | ||
| Write a page | 18.5 | 8.4 Wh | 0 % | ||
| Write an email | 194 | 45 Wh | 1 % | ||
| Write meeting minutes | 28.2 | 40 Wh | 1 % | ||
| Web search with an AI | 139 | 380 Wh | 9 % | ||
| Coding agent hourdoes not apply with this tooling |
Result for the chosen case
220 days, servers worldwide: 2 kgCO2e, 4.3 kWh.
For a carbon footprint
| Item | purchases of services (GHG Protocol scope 3, category 1); regulatory BEGES, category 4 |
|---|---|
| Quantity | 1 role over one year |
| Factor | 2 kgCO2e per role over one year, IA-0071 version 0.3, according to the local grid (location-based), servers worldwide |
| Total | 2 kgCO2e · not counted: server manufacturing, model training, the person's device and the tasks done without AI |
Typical usage profile
Projet Celsius assumption
The typical developer with an assistant accepts about 800 suggestions, has 10 functions written, has 2 to 3 pieces of code reviewed, asks 22 technical questions and opens 2 debugging sessions. About 47 requests per week, not counting suggestions. Over the year, about 4.3 kWh, of which 56% for code suggestions, which re-read the file each time. Light use, 1.5 kWh; intensive, 8.5 kWh. The same developer with agents hands them about 13 hours of work per week: the year rises to about 400 kWh, 93 times more, almost all of it in agent hours.
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 |
|---|---|---|---|---|---|
| Code suggestion | 35,000 times | 0.069 Wh | 2.4 kWh | 56 % | assumption |
| Generate a function | 460 times | 1.2 Wh | 540 Wh | 12 % | assumption |
| Web search with an AI | 140 times | 2.7 Wh | 380 Wh | 9 % | assumption |
| Conversation of 10 exchanges | 88 times | 3.5 Wh | 310 Wh | 7 % | assumption |
| Quick question | 990 times | 0.3 Wh | 300 Wh | 7 % | assumption |
| Query a PDF | 34 times | 4.7 Wh | 160 Wh | 4 % | assumption |
| Have code reviewed | 110 times | 1.3 Wh | 140 Wh | 3 % | assumption |
| Write an email | 190 times | 0.23 Wh | 45 Wh | 1 % | assumption |
| Write meeting minutes | 28 times | 1.4 Wh | 40 Wh | 1 % | assumption |
| Write a page | 18 times | 0.45 Wh | 8.4 Wh | 0 % | assumption |
Sources of the calculation
22 publications, for this entry and those used to calculate it, with the chosen parameters.
The result is calculated by Projet Celsius. The sources below support its uses, its factors and its conversion; they do not all measure this total.
Uses and assumptions
These publications shed light on AI tasks and uses. They do not measure the carbon footprint of the case calculated.
Anthropic · February 2025
Role and limits : Anthropic
Author. Kunal Handa et al. (Anthropic)
What the publication covers. Claude.ai conversations matched to O*NET tasks and occupations
What it does not cover. API use and enterprise plans
Reading by the database. A single assistant, more oriented towards code than average
Published observation on uses
Scope of the figure used. Share of Claude.ai conversations linked to computing and mathematics occupations
Limit. Support for the place of code among users of a code-oriented assistant
Anthropic · March 2026
Anthropic Economic Index report: Learning curves ↗ (new tab)
Role and limits : Anthropic
What the publication covers. Claude.ai conversations from 5 to 12 February 2026 and API use
What it does not cover. the other assistants
Reading by the database. Tracks how the uses of a single assistant evolve
Published observation on uses
Scope of the figure used. Share of Claude.ai conversations linked to computing and mathematics, February 2026
Limit. Code remains the first use, a little less concentrated than in 2025
Apec · June 2025
Role and limits : Apec
What the publication covers. frequency of generative AI use at work, 2,000 private-sector managers surveyed in March 2025
What it does not cover. non-managers and the number of requests
Reading by the database. Recent French survey on frequency of use at work; copy read on a specialist press site
Support for the "Typical use" setting
Scope of the figure used. The median user among those who use AI, end of 2026: they turn to it every working day for part of their questions, writing and reading, about 5 to 15 requests per day in an office job. In an organisation or a territory, it also sets the share of employees who use it (25%).
Limit. This setting remains an assumption of the database, not a measurement of the occupation.
EnronSent
Role and limits : EnronSent
Author. Will Styler
What the publication covers. 96,106 messages sent by Enron employees, 13,810,266 words
What it does not cover. the messages received
Reading by the database. Professional emails from 2000 and 2001: an order of magnitude, not a norm
Write an email → : Write, with reasoning →
The email written, 150 words
Scope of the figure used. server, excluding the data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Write an email → : Read 1,000 tokens →
Length convention: an email
Scope of the figure used. A typical email: the sent messages of the Enron corpus average about 144 words; figure read in an excerpt, notice not verified
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Write an email → : Write, with reasoning →
Length convention: an email
Scope of the figure used. A typical email: the sent messages of the Enron corpus average about 144 words; figure read in an excerpt, notice not verified
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
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
Conversation of 10 exchanges → : Re-read from cache →
The context re-read from cache at each turn: the previous exchanges, and the document already in the conversation if there is one
Scope of the figure used. server, convention C4
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
The 10 other publications in this section
Eurostat · June 2026
Individuals - use of generative AI tools (isoc_ai_iaiu) ↗ (new tab)
Role and limits
Author. Eurostat, survey on ICT use in households
What the publication covers. share of people aged 16 to 74 who used generative AI in the last 3 months, at work, privately, for studies, by country and category, 2025 data
What it does not cover. the frequency of use
Reading by the database. Harmonised European survey, first year
Published observation on uses
Scope of the figure used. France, IT professionals, used generative AI at work in the last 3 months, 2025
Limit. Share of users, which depends on the organisation and the population; here, support for the gap between occupations
Role and limits
What the publication covers. about 6,000 characters of code read for a suggestion
What it does not cover. electricity
Reading by the database. A post of 2023, updated in 2024: the current size is not published
Code suggestion → : Read 1,000 tokens →
The code around the cursor, about 6,000 characters or 125 lines, read by a small model
Scope of the figure used. server, excluding the data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Microsoft Research · July 2025
Working with AI: Measuring the Applicability of Generative AI to Occupations ↗ (new tab)
Role and limits
Author. Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, Siddharth Suri (Microsoft)
What the publication covers. 200,000 conversations with Bing Copilot matched to O*NET activities; applicability score by occupation, in open data (version 7 of October 2026)
What it does not cover. actual adoption by occupation
Reading by the database. Measures what AI is used for and succeeds at, not who uses it
Published observation on uses
Scope of the figure used. Generative AI applicability score for software developers (code 15-1252), 200,000 conversations with Bing Copilot, from 0 to 1
Limit. Used for the occupation coefficient: square root of the score divided by 0.25, the score of office employees
Developer, 1 year → : Query a PDF →
About 3 long documentation sets queried per week, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Scope of the figure used. document read, response, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Write an email →
About 8 long emails or messages per day, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Scope of the figure used. tokens read and written, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Write meeting minutes →
About 2 minutes or incident reports per week, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Scope of the figure used. simple tasks, data centre included
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Web search with an AI →
About 3 web searches per day, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Scope of the figure used. searches, pages read, response, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Role and limits
Author. National Center for O*NET Development, US Department of Labor
What the publication covers. tasks of each occupation and frequency reported by job holders, in 7 classes from "once a year or less" to "every hour or more"
What it does not cover. occupations without a frequency survey, such as data scientists
Reading by the database. The reference classification of tasks; US job holders
Developer, 1 year → : Write a page →
About 2 pages of documentation per week, documentation being weekly or more often in O*NET; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Scope of the figure used. tokens read and written, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
OpenAI · September 2025
How People Use ChatGPT (NBER Working Paper 34255) ↗ (new tab)
Role and limits
Author. Aaron Chatterji, Tom Cunningham, David Deming, Zoë Hitzig, Christopher Ong, Carl Shan, Kevin Wadman (OpenAI, Duke, Harvard)
What the publication covers. exact message volumes of the consumer plans; classification of a sample of about 1.1 million conversations by topic, intent and link to work
What it does not cover. enterprise plans, under-18s, logged-out users
Reading by the database. The only published measurement of the volumes and uses of a major assistant; internal data, automatic classification
Published observation on uses
Scope of the figure used. Share of programming in consumer ChatGPT messages
Limit. Code mostly goes through development tools and agents, outside this measurement
Published observation on uses
Scope of the figure used. Share of technical help in the work messages of computing occupations
Limit. Support for the place of technical questions in the developer's task basket
OpenAI · December 2025
Role and limits
What the publication covers. actual use by ChatGPT enterprise customers and a survey of 9,000 employees
What it does not cover. the consumer plans and any absolute volume per employee
Reading by the database. Publisher report; the gaps between users are the useful data
Published observation on uses
Scope of the figure used. Gap between the 95th percentile and the median for code messages, the largest of all uses
Limit. Support for the high intensive use for code
Projet Celsius · November 2026
Tokens per word, per line of code and per cell, measured with 7 public tokenisers
Role and limits
What the publication covers. 479,888 words from 196 blog articles and their English translation, code from the site and the Python library, a table of measurements; tokenisers of GPT-4o, GPT-4, Llama 3, Gemini (via Gemma 3), Mistral, DeepSeek and Qwen
What it does not cover. the Claude tokeniser, not public
Reading by the database. A measurement on technical texts, figures included; to be repeated if publishers change their tokeniser
Projet Celsius measurement, repeated on request with the public tokenisers
Code suggestion → : Read 1,000 tokens →
Length convention: a line of code
Scope of the figure used. From 10 tokens per line (Python standard library) to 19 (the TypeScript code of this site), measured with GPT-4o tokenisation
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Code suggestion → : Write 1,000 tokens →
Length convention: a line of code
Scope of the figure used. From 10 tokens per line (Python standard library) to 19 (the TypeScript code of this site), measured with GPT-4o tokenisation
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Get code reviewed → : Write, with reasoning →
Length convention: one page
Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Conversation of 10 exchanges → : Re-read from cache →
Length convention: one page
Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Query a PDF → : Read 1,000 tokens →
Length convention: one page
Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Write a page → : Write, with reasoning →
Length convention: one page
Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Proofread a page → : Read 1,000 tokens →
Length convention: one page
Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Proofread a page → : Write, with reasoning →
Length convention: one page
Scope of the figure used. A report page, headings and paragraphs included: between the 1,500-character sheet (250 words) and a full A4 page in single spacing (about 550 words, at 6.1 characters per word measured on our articles)
Limit. This publication supports the quantity of text or content processed, not its carbon footprint.
Stack Overflow · 2025
Role and limits
What the publication covers. frequency of use of AI and agents, tasks handed over, about 33,000 respondents
What it does not cover. the number of requests
Reading by the database. Online survey of volunteer developers, more engaged than average
Published observation on uses
Scope of the figure used. Use AI tools every day in their development work (17.4% every week), 26,004 respondents
Limit. Support for the developer's typical use
Published observation on uses
Scope of the figure used. Write code partly with AI (16.9% mostly with AI, on a different base of respondents: the two do not add up)
Limit. Support for the use of function generation
Published observation on uses
Scope of the figure used. Use AI agents at work every day (9.2% every week)
Limit. Support for the use of agents in 2025
Developer, 1 year → : Code suggestion →
About 200 suggestions offered per day when the assistant is active, assumption; share done with AI based on the half of developers who use AI every day (Stack Overflow, 2025).
Scope of the figure used. code read, suggestion written, small model, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Generate a function →
About 6 functions written per day, assumption; share done with AI based on the 59% of respondents who write code partly with AI (Stack Overflow, 2025); replaced by agent hours when there are agents.
Scope of the figure used. tokens read and written, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Get code reviewed →
About 2 code reviews per day, assumption; share done with AI based on the 23% of respondents who have their code partly reviewed by AI (Stack Overflow, 2025).
Scope of the figure used. tokens read and written, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Quick question →
About 10 technical questions per day, assumption; share done with AI set high, looking for an answer being the main stated use (Stack Overflow, 2025).
Scope of the figure used. computation of the response, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Developer, 1 year → : Conversation of 10 exchanges →
About 2 debugging sessions per day, fixing code being weekly or more often in O*NET; share done with AI based on the 47% of respondents who debug partly with AI (Stack Overflow, 2025).
Scope of the figure used. messages, context re-read from cache, responses, data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Stack Overflow · 2026
Role and limits
What the publication covers. daily duration of AI use, code assistants and agents, tasks handed over
What it does not cover. the share of professional developers
Reading by the database. Same survey a year later; some questions changed
Published observation on uses
Scope of the figure used. Use AI more than 4 hours per day (24% for 2 to 4 hours, 25% for 1 to 2 hours), 10,926 respondents
Limit. Support for agent hours: the median falls between 2 and 4 hours per day
Published observation on uses
Scope of the figure used. Use code assistants and agents every day
Limit. Support for agent tooling in 2026
Zhang and Nam · June 2026
HANSEL: Extracting Breadcrumbs from Web Agent Trajectories for Interactive Verification ↗ (new tab)
Role and limits
Author. Yujin Zhang and Daye Nam
What the publication covers. 45 web agent trajectories: 13.16 steps and 6.02 pages visited per task on average
What it does not cover. electricity
Reading by the database. A small sample, which gives the order of magnitude of an agent's steps
Web search with an AI → : Read 1,000 tokens →
The 5 pages found, read by the model, and the question: web agents visit 6 pages per task on average
Scope of the figure used. server, excluding the data centre
Limit. The quantity and the share entrusted to AI are database assumptions, not a measured footprint.
Energy and factors
The references of the selected values and of the entries used in this calculation.
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
Google · January 2009
Role and limits : Google
What the publication covers. a search, data centre
Reading by the database. Never remeasured
Reference of the selected value
Scope of the figure used. A search, data centre included
Limit. Never publicly remeasured; Google's AI answers are not published
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
ML.ENERGY · February 2026
Role and limits : ML.ENERGY
Author. Chung et al., University of Michigan
What the publication covers. GPU energy per token, per response, per image, in steady state, according to the number of requests served at once
What it does not cover. processors, host memory, data centre cooling, manufacturing
Reading by the database. The best open benchmark: it serves to compare model sizes with each other, not as a production value
Support for the "Small model" setting
Scope of the figure used. A model of about 8 billion active parameters, like Llama 3.1 8B: about 6 times less electricity per token than the medium model, according to the ML.ENERGY benchmark.
Limit. The setting is a convention of the database supported by this publication.
Reference of the variant "Small model, about 8 billion parameters"
Scope of the figure used. Llama 3.1 8B, one H100 GPU, 64 requests at a time, GPU only, per 1,000 tokens written, reading of the question included
Limit. GPU only, on a benchmark: the database keeps only the ratio to the 70 billion model, 1 to 6.7
The measurement covers a different scope; it is not a direct measurement of this entry.
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
The other publication in this section
Watershed · Aug. 2026
Estimating GHG Emissions from AI Use: Framework for Corporate-Level Measurement ↗ (new tab)
Role and limits
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 factor: 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. This measurement concerns the reference variant, not a measurement of "Small model, about 8 billion parameters". 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. The 0.24 Wh already include the overhead of Google's data centre (0.02 Wh, 8%): with 375 tokens read and 125 written, on 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, to be finalised in the next version of the framework
Reference factor: 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. This measurement concerns the reference variant, not a measurement of "Small model, about 8 billion parameters". 3 times the read factor, based on the write-to-read ratio adopted by Watershed (Patel et al., Splitwise, 2024). Same derivation and same caveat as for reading: Google's 0.24 Wh already include about 8% for the data centre. Values proposed for review, to be finalised in the next version of the framework
Code suggestion → : 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
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
Generate a function → : 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.
Get code reviewed → : 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.
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.
Conversation of 10 exchanges → : 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.
Query a PDF → : 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.
Write a page → : 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.
Write an email → : 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.
Proofread a page → : 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.
Web search with an AI → : 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.
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
1 figure for the same operation, each within its own scope.
| Source | Date | Electricity | What the figure measures | Scope | Status |
|---|---|---|---|---|---|
| Developer, 1 yearselected by the database, v0.3 | 2026 | 4.35 kWh | the value selected by the database | same scope | estimated |
| Zeke Hausfather, The Climate Brink ↗The real energy use of agentic AI | August 2026 | 3 kWh | An average day of intensive Claude Code use, one user, 8 weeks | other scope | estimated |
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?
Nature of the data
Published readings, conventions, assumptions
- Projet Celsius assumptions 100 %
Usage assumptions
Frequency, share done with AI according to adoption, source of each assumption
Code suggestion : 200 per day, 80% with AI in typical use
About 200 suggestions offered per day when the assistant is active, assumption; share done with AI based on the half of developers who use AI every day (Stack Overflow, 2025).
Generate a function : 6 per day, 35% with AI in typical use
About 6 functions written per day, assumption; share done with AI based on the 59% of respondents who write code partly with AI (Stack Overflow, 2025); replaced by agent hours when there are agents.
Have code reviewed : 2 per day, 25% with AI in typical use
About 2 code reviews per day, assumption; share done with AI based on the 23% of respondents who have their code partly reviewed by AI (Stack Overflow, 2025).
Quick question : 10 per day, 45% with AI in typical use
About 10 technical questions per day, assumption; share done with AI set high, looking for an answer being the main stated use (Stack Overflow, 2025).
Conversation of 10 exchanges : 2 per day, 20% with AI in typical use
About 2 debugging sessions per day, fixing code being weekly or more often in O*NET; share done with AI based on the 47% of respondents who debug partly with AI (Stack Overflow, 2025).
Query a PDF : 3 per week, 26% with AI in typical use
About 3 long documentation sets queried per week, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Write a page : 2 per week, 21% with AI in typical use
About 2 pages of documentation per week, documentation being weekly or more often in O*NET; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Write an email : 8 per day, 11% with AI in typical use
About 8 long emails or messages per day, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Write meeting minutes : 2 per week, 32% with AI in typical use
About 2 minutes or incident reports per week, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Web search with an AI : 3 per day, 21% with AI in typical use
About 3 web searches per day, assumption; share done with AI taken from the value common to all occupations for this task, then adjusted to the occupation using the Microsoft study of 200,000 conversations.
Coding agent hour : 6 per day, 45% with AI in typical use
About 6 hours of development per day in which an agent can work, assumption; share done with AI based on the 31% who use AI more than 4 hours a day (Stack Overflow, 2026).
Published figures in support
OpenAI, Chatterji et al. (2025): 4.2% of messages
Share of programming in consumer ChatGPT messages
OpenAI, Chatterji et al. (2025): 37% of work messages
Share of technical help in the work messages of computing occupations
OpenAI (2025): 17 times the median
Gap between the 95th percentile and the median for code messages, the largest of all uses
Anthropic, Handa et al. (2025): 37.2% of conversations
Share of Claude.ai conversations linked to computing and mathematics occupations
Anthropic (2026): 35% of conversations
Share of Claude.ai conversations linked to computing and mathematics, February 2026
Microsoft Research, Tomlinson et al. (2026): 0.278 applicability score
Generative AI applicability score for software developers (code 15-1252), 200,000 conversations with Bing Copilot, from 0 to 1
Eurostat (2025): 65.4% of people
France, IT professionals, used generative AI at work in the last 3 months, 2025
Stack Overflow (2025): 50.6% of professional developers
Use AI tools every day in their development work (17.4% every week), 26,004 respondents
Stack Overflow (2025): 59% of respondents to the question
Write code partly with AI (16.9% mostly with AI, on a different base of respondents: the two do not add up)
Stack Overflow (2025): 14.9% of professional developers
Use AI agents at work every day (9.2% every week)
Stack Overflow (2026): 31% of respondents
Use AI more than 4 hours per day (24% for 2 to 4 hours, 25% for 1 to 2 hours), 10,926 respondents
Stack Overflow (2026): 73% of respondents
Use code assistants and agents every day
Methodology conventions
With their alternatives
| The working day and year | 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 |
|---|---|
| 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 sum of what makes up the entry, then 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 |
History
Entry in version 0.3 of 2 November 2026, in the database in version 0.4
- 2 November 2026Designation : A day of a developer with an agent becomes Developer with agent, 1 day.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 2026Designation : Developer with agent, 1 day becomes Developer, 1 year.Database overhaul into tiers: the block becomes the developer's task basket, over one year; the day becomes a duration.
- 2 November 2026Short title : Developer with agent, 1 day becomes Developer, 1 year.Database overhaul into tiers: the block becomes the developer's task basket, over one year; the day becomes a duration.
- 2 November 2026Address : developpeur-avec-agent-1-jour becomes developpeur-1-an.Database overhaul into tiers: the block becomes the developer's task basket, over one year; the day becomes a duration; the old address now redirects.
- 2 November 2026Functional unit : 1 user · day becomes 1 role · year.Database overhaul into tiers: the block becomes the developer's task basket, over one year; the day becomes a duration.
- 2 November 2026Status : unpublished becomes estimated.With the default tooling (assistant), the basket is calculated without the agent hour, which has no value until it is set.
- 31 October 2026Creation of the database, in its first version (0.1).
