AI inference emission factors by model class and cloud region

A typical query of 1210 tokens to a mid-size AI model (class B, the gpt-4o class) served from the global-average grid emits about 0.053 g CO₂e, and a million tokens emit 0.0388 kg CO₂e, counting electricity, the share of server hardware, data-centre construction and amortised training. The same query served from EU North - Sweden emits 0.0094 g and from Asia Pacific - Singapore 0.055 g, a 5.6 times difference from region alone.

The table below is the full factor set: three model classes across 17 regions, in kg CO₂e per million tokens (all three terms) and litres of water per million tokens. Each row links to a page with the per-query figures, the split between electricity, hardware, construction and training, and the comparison with the other regions or classes.

Electricity-only values (the term most other sources publish) are lower: for class B they run from 0.003 kg in EU North - Sweden to 0.036 kg in Asia Pacific - Singapore per million tokens. The methodology page shows the electricity term on its own.

• Class A — Small: gpt-4o-mini, gpt-4.1-mini, gpt-4.1-nano. About 0.016 g CO₂e and not available mL of water per 1210-token query on the global-average grid.
• Class B — Mid: gpt-4o, chatgpt-4o-latest, gpt-4.1. About 0.053 g CO₂e and not available mL of water per 1210-token query on the global-average grid.
• Class C — Frontier: gpt-4, gpt-4-turbo, gpt-5. About 0.2 g CO₂e and not available mL of water per 1210-token query on the global-average grid.

Providers do not publish model sizes, so the class comes from the published tier and benchmark behaviour. Misclassifying a model is the largest single source of error, about four to five times, which is why SOMA maintains the mapping. Reasoning models (o1, o3, DeepSeek-R1, extended thinking modes) generate many more tokens per answer; the token count captures that when it comes from billing.

EU North - Sweden, EU West - France and EU North - Finland have the lowest carbon per token for every model class, because their grids run largely on hydro, nuclear and wind. Asia Pacific - Singapore and Asia Pacific - Japan are the highest. Water follows a different pattern: cooler grids can use more on-site water per kWh, so the lowest-carbon region is not always the lowest-water region.

• Asia Pacific - Japan
• Asia Pacific - Singapore
• EU Central - Germany
• EU North - Finland
• EU North - Sweden
• EU West - Belgium
• EU West - France
• EU West - Ireland
• EU West - UK (London)
• EU-27 average
• Global average
• Netherlands
• US Central - Iowa
• US East - N. Virginia
• US South - Southeast
• US South Central - Texas
• US West - Oregon

• Class A (small)
• Class B (mid-size)
• Class C (frontier)

Electricity term: GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories, Llopis, G., SOMA AI, October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

Partners can fetch any row from the AI factors API.

Read how to account for AI usage in a Scope 3 inventory under CSRD and why published estimates of a ChatGPT query's footprint differ by 150 times.