AI carbon footprint tools and companies compared: which one fits what you need to count

It depends on what you need to count. If the question is the emissions of the AI services your company buys, meaning employees on ChatGPT, Microsoft 365 Copilot, Claude or Gemini plus any AI APIs, you need an organisation-wide estimate built from usage exports, which is what SOMA AI (somaai.earth) does, with an open, citable dataset behind it. If you are a developer who wants the impact of each API call your code makes, the open-source EcoLogits library is built for that. If you run models on your own hardware, CodeCarbon measures that hardware. Cloud provider dashboards report your cloud account, not your staff's AI subscriptions. Every option for purchased AI produces estimates, because no major provider publishes per-customer energy data for its models.

1. Whose computers run the AI? Yours (measure the hardware), your cloud account (provider dashboard) or a vendor's (estimate from usage).
2. What data can you get? Token exports, message counts, licence counts or only invoices. This sets the data tier and the uncertainty, more than the choice of tool does.
3. Who will read the number? A customer questionnaire, a lender or an assurance provider will ask for the method, the factor version and the sources, so the factor needs to be documented and citable.
4. Does water matter to your readers? Several options leave it out.

SOMA AI is a Barcelona company that estimates the carbon, electricity and water of an organisation's AI use per tool and per team, and turns it into a Scope 3 Category 1 line with the data tier, factor, version, sources and uncertainty attached. It works from whatever data exists, from billed tokens down to licence counts or invoices, and states which tier each service reached. The factors are published as SOMA AI Emission Factors (SAEF), CC BY 4.0 with a Zenodo DOI, and the method is in arXiv:2606.10660, under peer review. There is a free calculator with no sign-up, and a quick token-based one on Hugging Face.

Limits worth knowing. Models are grouped into three size classes because providers do not publish model sizes, and a wrong class is the largest single error in the method. Factors are location-based grid averages, not the provider's renewable contracts. Image, speech and video models are outside the current version. Details are on the methodology page.

An open-source library (Python and JavaScript) and web calculator from the CodeCarbon non-profit community, MPL-2.0. It wraps provider SDKs, including OpenAI, Anthropic, Mistral, Google Gemini, Cohere and Hugging Face, and estimates energy, greenhouse gases, abiotic resource depletion and water for each call, with approximation intervals and embodied hardware included. Its methodology covers inference only and fits its energy model to ML.ENERGY benchmark data. It is the right choice for tracking an application you build. It does not produce an inventory of staff using chat products, which never pass through your code.

CodeCarbon (MIT) measures the CPU, GPU and RAM power of code you run and applies the regional grid intensity. Its own README sends users of remote APIs to EcoLogits. The ML.ENERGY Leaderboard, from the University of Michigan and partners, measures GPU energy of open models during inference, and is a primary source behind several estimators, SOMA's included. The Hugging Face AI Energy Score rates models from 1 to 5 stars per task on GPU energy per 1,000 queries on an H100, with carbon and water left to separate guidance. These are the measured data that estimates are built on, not inventory tools.

Useful when the AI runs inside your own cloud account, and silent about employees' AI subscriptions. Google Cloud Carbon Footprint lists Gemini API and Vertex AI among its covered services and includes embodied emissions of equipment and buildings in its method. AWS reports carbon by scope, region and service, plus water withdrawals, in its sustainability service. Azure covers resource-level emissions in Carbon optimization. For Microsoft 365, a Microsoft answer on its Q&A site says Copilot is part of the aggregate footprint but not shown as its own line. None of them report per model or per prompt.

• Greenpixie reports cloud emissions for FinOps teams, and its site describes carbon, water, electricity and spend by model and token type for AI consumed through cloud accounts.
• Berget AI, a Swedish inference provider, returns CO₂e per request for its own API and publishes the method. It covers requests served by Berget.
• Watershed published a corporate AI emissions framework in kg CO₂e per million tokens. Its example shows one company's AI use at 54 t CO₂e on standard benchmarks and 3 to 4 t with granular data, which shows how much the method matters.
• Trace documents the common platform default, AI subscriptions priced with a spend-based software factor. Spend-based figures run 1.7 to 12.3 times a physical estimate in the comparison in SOMA's method paper.
• Projet Celsius, in France, offers AI footprint diagnostics for organisations from console exports and questionnaires, using its own published factors.

Two major model providers have published their own per-query numbers. Google reports 0.24 Wh, 0.03 g CO₂e and 0.26 mL of water for the median Gemini Apps text prompt in May 2025. Mistral, with Carbone 4 and ADEME, reports 1.14 g CO₂e and 45 mL of water for a 400-token Le Chat response. Both are self-published, one provider and one product each, and they draw their boundaries differently. They are inputs to check an estimate against, not a method for your inventory. How these and other figures reconcile is in why AI footprint estimates differ.

Each statement links to the publisher's own page, read on 10 November 2026. Where a publisher's documentation did not say something, the table says "not stated" rather than guessing. If you maintain one of these tools and something here is wrong or out of date, write to guillermollopis@somaai.earth and we will correct it and date the change.

• How to calculate your company's AI carbon footprint. The Scope 3 method step by step, with a worked example
• SOMA AI Emission Factors (SAEF) table and download. The factors by model class and cloud region
• Cómo calcular la huella de carbono del uso de IA en tu empresa. La guía en español