Uses14 November 2026Which emission factors should be used for AI in business?
What Microsoft, OpenAI, Google, Mistral AI and Anthropic publish, and what the AI Impact Database estimates, tool by tool
A hackathon at ETH Zurich. Photo Arthur Gamsa, CC BY 4.0, cropped, via Wikimedia Commons.
No provider publishes an emission factor for its enterprise offerings: neither Microsoft for Copilot, nor OpenAI for ChatGPT Enterprise, nor Anthropic for Claude. Google and Mistral AI publish a value per question, with its method, for their consumer assistant; OpenAI gives a figure without a method. For a carbon footprint, an estimate is therefore needed: according to the AI Impact Database, an office assistant licence represents about 1 to 1.3 kgCO2e per year for server electricity on the global grid, and a coding assistant licence with an agent about 29 kgCO2e.
Contents
To count AI in a Bilan Carbone®, an organisation multiplies what it buys, licences or compute hours, by an emission factor. For most digital services, this factor exists; for AI assistants in enterprise offerings, no provider publishes one. This is the clearest gap we have found among organisations, in November 2026.
A licence for 1 year per employee, in kgCO2e: AI Impact Database estimates, server electricity at the world grid. No provider publishes these values.
These 3 values are estimates built task by task: meetings summarised, emails, pages drafted, questions, and for the coding assistant suggestions, functions and agent hours. They count server electricity, cooling included, at the average factor of world electricity, 458 gCO2e/kWh. They count neither server manufacturing nor model training. Usage frequencies are assumptions of the database, which anyone can change on the entry's page.
What providers publish
No provider publishes a value for an enterprise offering. What exists relates to the consumer offering, and each provider gives it in its own way:
- Google : 0.24 Wh and 0.03 gCO2e for the median text question to the Gemini apps, measured in production in May 2025; the CO2e is calculated according to its electricity contracts (Elsworth et al.) ;
- Mistral AI : 1.14 gCO2e and 45 mL of water for a response of about one page from its assistant, over the life cycle, server manufacturing included, in a analysis carried out with Carbone 4 and ADEME ;
- OpenAI : 0.34 Wh for the average ChatGPT query, in a Sam Altman blog post from June 2025, with no method or scope;
- Microsoft : from 0.16 to 0.60 Wh for a typical question to large models in production, in a blog post from June 2026 which cites Copilot as an example of a model without saying whether this range applies to it; no CO2e figure;
- Anthropic and GitHub : nothing.

These values cannot be compared with one another: Google's counts the data centre but not manufacturing, Mistral AI's counts manufacturing, OpenAI's does not say what it counts. Our analysis of ChatGPT sets out what each is worth.
The table, tool by tool
The table places side by side, for each tool common in business, the value published by the provider and that of the AI Impact Database, with its range. All the database values are also in its downloadable files, to reuse in a spreadsheet.
| Tool and unit | Published by the provider | Estimated by the AI Impact Database | Status |
|---|---|---|---|
| Microsoft 365 Copilot, licence for 1 year | nothing for Copilot | 2.9 kWh; 1.3 kgCO2e (1.0 to 1.7) | estimated |
| ChatGPT Business or Enterprise, licence for 1 year | nothing for the offering | 2.2 kWh; 1.0 kgCO2e | estimated |
| GitHub Copilot, licence for 1 year | nothing | 63 kWh; 29 kgCO2e (14 to 54) | estimated |
| Coding agent, 1 hour | nothing | 0.67 kWh; 310 gCO2e (140 to 590) | estimated |
| Question to Gemini | 0.24 Wh; 0.03 gCO2e based on contracts | 0.11 gCO2e at the world grid | measured by the provider |
| Question to ChatGPT | 0.34 Wh, no method | 0.16 gCO2e at the world grid | declared by the provider |
| One-page response from Le Chat (Mistral AI) | 1.14 gCO2e over the life cycle | published value reproduced | published life cycle assessment |
| Question to Claude | nothing | 0.3 Wh; 0.14 gCO2e (Haiku 0.05 Wh, Opus 1.2 Wh) | estimated |
| Generated image | nothing for closed generators | 3.2 Wh; 1.5 gCO2e (0.7 to 5.8) | measured on open models |
Server electricity, data centre cooling included, excluding server manufacturing and model training, except for Mistral AI. CO2e at the average factor of world electricity, 458 gCO2e/kWh. AI Impact Database, version of 14 November 2026.
Why the coding assistant weighs 20 to 30 times more
Office assistant licences weigh little because their actions weigh little: a question costs about 0.3 Wh, an email or a page less than 1 Wh. The heavy actions, a summarised meeting (about 30 Wh) and a deep research (about 100 Wh), are rare and make up three quarters of a Microsoft 365 Copilot licence. One hour of coding agent consumes about 0.67 kWh, because the agent chains about 120 steps and rereads its whole context each time (our analysis of agents). The database assumes about 90 agent hours per year per GitHub Copilot licence; they account for almost the entire value. GitHub does not publish the equivalence between its credits and agent hours: this is the setting to change first for a team that uses it more or less.
Using it in a carbon footprint
For an organisation, this means multiplying the licences and agent hours actually used by the value in the table, with its range, using the electricity factor of the servers' country when it is known; in a Bilan Carbone® in France, the ADEME factor, 51.9 gCO2e/kWh. Including AI in your carbon footprint, on our blog, gives the full method.
For 200 Microsoft 365 Copilot licences, this makes about 260 kgCO2e per year at the world grid, and about 120 kg with servers in the European Union: a very small part of a services company's carbon footprint. Our analysis of the Bilan Carbone® sets out this calculation and the item under which to file it; for a local authority, see the responsible digital footprint.
Our reading
For us at Projet Celsius, the absence of a value published by providers is no reason to leave AI out of the carbon footprint: a documented estimate, with its range, is better than a zero. For office assistants, the order of magnitude is low and stable; for agents, it depends mainly on the number of hours, which the organisation can measure. The questions to ask your AI provider are in the same blog article.
The AI Impact Database values are estimates: usage frequencies for a licence are assumptions, adjustable on each entry's page. The values published by Google, Mistral AI and OpenAI relate to their consumer offerings and do not count the same things.
Sources
5 documents
- Projet Celsius, AI Impact Database, environmental database of artificial intelligence, engine 0.3.12026entry 144
- Elsworth et al. (Google), Measuring the environmental impact of delivering AI at Google scale2025entry 14
- Mistral AI, with Carbone 4 and ADEME, Our contribution to a global environmental standard for AI2025entry 17
- Sam Altman (OpenAI), The Gentle Singularity2025entry 21
- Microsoft, Scaling AI with 8 to 20x energy efficiency2026entry 284
Every figure was read in its source; the hub's methodology is described in Methodology and sources.
AI Impact, Projet Celsius, "Which emission factors should be used for AI in business?", 14 November 2026. https://projetcelsius.com/en/ai/analyses/ai-emission-factors-for-businesses/
















Climate






