The carbon and water footprint of AI usage, as an open dataset

It depends on three things, and a single number hides all of them. The size class of the model, the cloud region it is served from, and what the estimate counts.

A typical exchange of 1210 tokens with a mid-size model on the global-average grid emits about 0.053 g CO₂e. The same exchange costs 5.6 times as much from Asia Pacific - Singapore as from EU North - Sweden. A frontier model costs about 13 times a small one for the same tokens.

SOMA publishes the factors that turn token counts from a provider invoice into a Scope 3 Category 1 line, with the uncertainty and the sources attached.

They are estimates built from published benchmarks and published grid data, not readings of anyone's data centre. Every figure on these pages is derived from the dataset rather than typed in.

• Open licence. CC BY 4.0, so a consultancy, an assurance provider or a software vendor can build on the factors and cite them.
• Citable and versioned. Concept DOI 10.5281/zenodo.20443585 resolves to the latest version, and every older version stays resolvable, so last year's inventory can be reproduced. See the changelog.
• Under peer review. The method is a preprint under review at Sustainable Production and Consumption.
• Independent of the vendors it covers. SOMA does not sell AI models, so no row is a claim about its own product.
• Built for the inventory. Four data tiers, one entry per service, mapped to GHG Protocol Scope 3 Category 1 and ESRS E1. Water is reported beside carbon and never added to it.

• MethodologyHow a token becomes carbon and water: terms, tiers, classes, uncertainty and sources.
• SOMA AI Emission Factors (SAEF)The full table across 18 regions and 3 model classes, with a CSV download and the dataset record.
• AI factors APIOne factor row per request, for a model id and a cloud region.
• ChangelogEach published version, its date, record and changes.

• Class A: small modelsgpt-4o-mini, gpt-4.1-mini, gpt-4.1-nano. About 0.016 g CO₂e per 1210-token query on the global-average grid.
• Class B: mid-size modelsgpt-4o, chatgpt-4o-latest, gpt-4.1. About 0.053 g CO₂e per 1210-token query on the global-average grid.
• Class C: frontier modelsgpt-4, gpt-4-turbo, gpt-5. About 0.2 g CO₂e per 1210-token query on the global-average grid.

One page per region, with the per-query figures, the split between electricity, hardware, construction and training, water, and the comparison with the other regions. Lowest carbon first for a mid-size model:

• EU North - Sweden0.00714 kg/M tokens
• EU West - France0.0082 kg/M tokens
• EU North - Finland0.0102 kg/M tokens
• EU West - Belgium0.014 kg/M tokens
• EU South - Spain0.0165 kg/M tokens
• EU West - UK (London)0.0212 kg/M tokens
• Netherlands0.0235 kg/M tokens
• EU West - Ireland0.0245 kg/M tokens
• US East - N. Virginia0.0261 kg/M tokens
• US West - Oregon0.0275 kg/M tokens
• EU Central - Germany0.0303 kg/M tokens
• US South Central - Texas0.0312 kg/M tokens
• US South - Southeast0.0351 kg/M tokens
• US Central - Iowa0.0381 kg/M tokens
• Asia Pacific - Japan0.0395 kg/M tokens
• Asia Pacific - Singapore0.0401 kg/M tokens
• Global average0.0388 kg/M tokens
• EU-27 average0.0212 kg/M tokens

• How much energy does a ChatGPT query use?About 0.3 Wh for a short text query on published estimates, and several Wh for long or reasoning queries.
• What is the carbon footprint of one ChatGPT prompt?Published estimates for one AI prompt range from 0.02 g to several grams of CO2e.
• Does an AI query use more energy than a Google search?The "ten times a Google search" claim traced to its 2024 source, set against the 2025 figures from Google, OpenAI and independent researchers..
• Is AI bad for the environment?Per query, AI is small.
• AI environmental impact statistics: energy, carbon and water, with sourcesChecked statistics on AI energy per query, training emissions, water use and data-centre electricity, each with its publisher, date and boundary..
• How much more energy do reasoning models use than standard chat?Benchmarks show reasoning modes using hundreds of times more energy per query on the same model.
• What is the carbon footprint of training an AI model?Published training emissions from hundreds of tonnes to tens of kilotonnes of CO2e, and how training is shared out over the queries a company sends..
• What is the carbon footprint of Microsoft 365 Copilot, ChatGPT Enterprise and other AI seats?No provider reports emissions per seat.
• How much energy does generating an AI image or video use?Measured figures for open image and video models, compared with text, and why image, speech and video need their own line in an inventory..
• Where does an AI query emit the least? The carbon footprint of AI by countryThe same mid-size AI exchange estimated on the electricity grid of 210 countries, the EU-27 ranked, and the countries that host cloud regions.
• Which AI model has the lowest carbon footprint? GPT, Claude, Gemini, Llama, DeepSeek and Mistral comparedEstimated CO2e and energy per exchange for named AI models, from their size class and serving region, with the published provider figures beside them and the limits of each..
• What is the carbon footprint of Claude, and how does it compare with ChatGPT?Estimated CO2e, energy and water per exchange for Claude Haiku, Sonnet and Opus, what Anthropic has and has not published on per-query energy, and how Claude compares with ChatGPT and Gemini..
• What is the carbon footprint of AI coding assistants like Claude Code, GitHub Copilot and Cursor?Inline completions and agentic coding sessions use very different numbers of tokens.
• How much energy does DeepSeek use, and is it more efficient than ChatGPT?Published energy figures for DeepSeek-R1 and its distilled models, why reasoning output dominates, what serving location does, and why no published source compares DeepSeek and GPT models on the same basis..
• How does the carbon footprint of AI compare with driving, flying, streaming and food?One AI query, a year of company AI use and the AI industry set against everyday and sector figures, each with its source, so comparisons are like with like..

Huella de carbono de la IA: guías y respuestas en español

• How to calculate your company's AI carbon footprint for Scope 3, and when CSRD appliesThe whole job in order: services, usage exports, tier, factor, the line an auditor can follow.
• Why AI footprint estimates differ: published figures, datedPublished figures with their boundary and date, and how to restate one on another boundary.
• GHG Protocol Scope 3 data tiers for AI usage, mappedWhere each data tier falls under the draft data types, and what to report today.
• EU AI Act energy documentation for GPAI modelsWhat Article 53 and Annex XI require, the dates, and how to estimate when consumption is unknown.
• AI water footprint: cooling, electricity and scarcity explainedWhich water a figure counts, SAEF water factors by region, and a worked example.
• AI carbon footprint tools and companies compared (2026)Which option fits what you need to count: purchased AI, your own code, your own hardware or your cloud account.
• Cómo calcular la huella de carbono del uso de IA en tu empresaLa guía en español, con un ejemplo resuelto en España.

The free AI footprint calculator applies these factors to the AI tools your teams use. Answer a few questions (tools, people, how often) and it estimates the CO₂e, electricity and water per team and per tool, with a range, and shows the changes that would cut it. Message counts from an admin export or token numbers from an invoice sharpen the estimate.

Download the table as CSV, cite the dataset by its concept DOI, or call the API. To have the usage exports of your AI providers turned into the finished inventory line, see SOMA AI Footprint.