Written by SOMA AI, which is one of the options below, so check each claim against the linked sources. Drafted with AI assistance and fact-checked against the sources on the original page.
You can meter the AI that runs on hardware you control, such as your own GPUs and servers. For AI you buy as a service, such as ChatGPT, Copilot, Claude or Gemini, you cannot meter anything, because providers do not publish energy per customer or for the requests you send. You can only estimate it from usage and published factors. Either way you end up with three quantities: energy in kWh, emissions in kg CO2e and water in litres.
What can I measure directly, and what can I only estimate?
It depends on whose computers run the AI.
- Your own hardware. A power meter or a smart plug gives you metered kWh. Software such as CodeCarbon reads hardware energy counters where it can and estimates the rest from load. Multiply by the carbon intensity of your grid and you have CO2e.
- Your own cloud account. Cloud providers report the emissions of the resources you rent. They do not break this down per model or per prompt.
- A vendor's computers. This covers chat seats and API calls. The vendor does not give you the energy, so the number is an estimate built from your usage (messages, tokens or spend), an energy factor for the model size, and the grid where it runs.
An estimate is not a weaker answer if you label it as one and write down the method. It becomes a problem when it is reported as a reading.
Which tool fits which kind of AI use?
| Your AI use | Tool that fits | What you get |
|---|---|---|
| Training or inference on your own GPUs | CodeCarbon or Green Algorithms | Estimated energy and emissions of your own hardware |
| API calls made from your code | EcoLogits, or SAEF (pip install saef, with a LangChain SaefMiddleware) | Per-call estimates |
| ChatGPT, Copilot or Claude seats across a company | SOMA AI calculator, or the SAEF factors applied to seats x messages | An organisation-wide estimate for a Scope 3 line |
| Cloud infrastructure you rent | Your cloud provider's own dashboard | Emissions of your cloud account |
SAEF is the SOMA AI Emission Factors, an open dataset under CC BY 4.0 with a DOI. The SOMA AI calculator is free and needs no sign-up. EcoLogits is an open-source library that wraps provider SDKs. It fits an application you build, but chat products used by staff never pass through your code. CodeCarbon's own README sends users of remote APIs to EcoLogits.
How do I estimate my ChatGPT or Copilot use?
Use three steps and keep a note of each assumption.
- Count usage. Take active seats and messages per seat per month from the admin export. Microsoft 365 Copilot reports prompts submitted to Copilot Chat per user in the admin center, with no token counts. ChatGPT Enterprise and Edu show messages per user in workspace analytics.
- Pick an energy figure by model class. SAEF assumes an exchange of 1,210 tokens (928 in, 282 out). For that exchange a mid-size model, class B, comes to 0.099 Wh at facility level. Class A is 0.03 Wh and class C is 0.37 Wh. Providers do not publish model sizes, so choosing the class is your largest source of error.
- Apply the grid. Multiply by the carbon intensity of the region where the model runs. The SAEF factors already include server and facility overheads and a modelled share of hardware, building and training.
Worked example, 250 Copilot seats. Take 250 active seats, 30 messages per seat per month, 12 months and class B, with the dataset's reasoning uplift on output tokens. That is 90,000 messages and about 140 million billed tokens. The SOMA AI estimate is 5.02 kg CO2e on a US East (N. Virginia) grid, with a band of 2.5 to 12 kg. On a Swedish grid it is 1.25 kg. If the same 250 seats are heavy users at 325 messages per seat per month, the US East estimate rises to 54.4 kg. How much people use the tool and where it runs move the figure as much as the number of seats does.
If you only have an invoice, a spend factor of 0.09613 kg CO2e per euro (deflated) gives a screening figure for the whole subscription. Use it as a first pass and label it as such. Usage-based estimates are better when you can get the data.
How much is one ChatGPT query?
About a third of a watt-hour for a short text query, on published estimates. No provider has published a metered, audited figure for ChatGPT.
- Google reports 0.24 Wh for the median Gemini Apps text prompt (August 2025). That counts active chips, host machines, idle capacity and data-centre overhead. It is Google's service, not ChatGPT. Google gives 0.10 Wh for a narrower scope that counts only active chips.
- Epoch AI modelled about 0.3 Wh for GPT-4o with 500 output tokens, with a range of 0.1 to 4 Wh under other assumptions.
- Sam Altman gave 0.34 Wh for the average ChatGPT query (June 2025), without a method or boundary.
- SOMA AI estimates 0.099 Wh at facility level for a 1,210-token exchange on a class B model.
The numbers differ because the boundaries and the query differ, not because one is wrong. Length matters most. Epoch AI puts a 10,000-token input at about 2.5 Wh and a 100,000-token input at about 40 Wh, so "one query" hides a very wide range. SOMA AI figures and the provider figures above are not a like-for-like comparison.
Should I report it, and where?
For a company, AI services you buy belong in GHG Protocol Scope 3 category 1, purchased goods and services. List every service, including chat seats, APIs and AI inside other software. Then:
- Label the figure as an estimate and show the band.
- Name the factor version you used and cite its source.
- State your assumptions, in particular the model class and the region.
- Replace defaults with your own message or token counts as soon as you have them.
Auditors and customers tend to care more about a documented method than about decimals.
What about water?
Water figures vary even more than energy, mostly because sources count different things. Google reports 0.26 mL for the median Gemini Apps text prompt. Mistral, with Carbone 4 and ADEME, reports 45 mL for a 400-token Le Chat response. Both are self-published, cover one provider each and draw their boundaries differently. SOMA AI estimates, on its own boundary, 0.24 mL per 1,210-token exchange for a mid-size (class B) model served from US East (N. Virginia) and about 0.6 mL from EU North (Sweden). The region changes the water, not the watt-hours. Treat any single "water per prompt" number with suspicion unless it says which water it counts.
What are the limits of this approach?
Models are grouped in three size classes, and a wrong class is the largest single error. Factors are location-based grid averages, not a provider's renewable contracts. Image, speech and video models are outside the current SAEF version. Every option for purchased AI produces estimates, ours included.
Related. SOMA AI calculator, SAEF factor table and download, methodology page, "How to estimate a team's ChatGPT and Copilot footprint", "LLM token CO2e formula", "Tools compared".
Originally published at somaai.earth. Spanish version: ¿Cómo puedo medir el impacto ambiental de la IA que uso?
