AI environmental impact statistics: energy, carbon and water, with sources
For journalists, analysts and sustainability teams who need a number they can quote with its boundary. Sources checked on 31 October 2026. Factor version soma-ai-ef-2026.10-dataset-v3.0.0. Last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.
Part of the SOMA AI Emission Factors cluster.
AI environmental impact statistics: energy, carbon and water, with sources
For one short text query, the published energy figures cluster between 0.2 and 0.4 Wh. Google reports 0.24 Wh for its median Gemini Apps prompt including idle machines and overhead, Epoch AI models 0.3 Wh for GPT-4o, and Sam Altman states 0.34 Wh with no method. Long inputs and reasoning push single queries to several Wh or more. Training a single large model has been put at tens of tonnes to tens of thousands of tonnes of CO₂e, depending largely on the boundary, and the electricity use of all data centres at 415 TWh in 2024, according to the IEA. Each table below gives the publisher, date, boundary and basis of every figure, so that none is quoted without what it counts. They are estimates, and SOMA AI (somaai.earth) computes its own from the same sources.
How much energy does one AI query use?
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 0.047 Wh per inference | Luccioni, Jernite and Strubell, November 2023 | Text-generation models, mean across a benchmark, measured on an 8-GPU A100 node including idle power of the unused GPUs | Measured, benchmark datasets |
| 0.3 Wh per query | Epoch AI, 7 February 2025 | GPT-4o, 500 output tokens, inference only. Epoch calls it relatively pessimistic and says other assumptions give 0.1 to 4 Wh | Modelled |
| 2.5 and almost 40 Wh per query | Epoch AI, 7 February 2025 | Same GPT-4o method with 10,000 and 100,000 input tokens | Modelled |
| 0.42 Wh per query | Jegham et al., 14 May 2025 | GPT-4o, 100 input and 300 output tokens, including data-centre overhead. 0.37 Wh without it | Modelled |
| 0.443 to 29.078 Wh per query | Jegham et al., 14 May 2025 | Prompts of 10,000 input and 1,500 output tokens. Low end Llama 3.1 8B, high end DeepSeek-R1 on DeepSeek's own servers | Modelled |
| 0.032 Wh per response (114 J) | MIT Technology Review, 20 May 2025 | Llama 3.1 8B. 57 J of GPU energy, doubled to allow for the rest of the system | GPU measured, doubling assumed |
| 1.86 Wh per response (6,706 J) | MIT Technology Review, 20 May 2025 | Llama 3.1 405B. 3,353 J of GPU energy, doubled in the same way | GPU measured, doubling assumed |
| 0.21 Wh per request | ML.ENERGY Leaderboard, v3.0, data as of 16 February 2026 | Llama 3.1 405B FP8, mean energy per request, average output 367 tokens, at one heavily batched 8 x H100 setting. GPU energy only. Across 15 configurations of the same model the range is 435 to 5,110 J, or 0.12 to 1.42 Wh | Measured |
| 0.31 Wh per query (IQR 0.16 to 0.60) | Oviedo et al. (Microsoft), Joule 2026, first posted 24 September 2025 | Models above 200 billion parameters on H100, 300 median output tokens, with a modelled PUE and production batching | Modelled |
| 0.24 Wh per prompt | Google (Elsworth et al.), 21 August 2025 | Median Gemini Apps text prompt, May 2025. Active chips, host CPU and memory, idle machines and overhead. Training excluded | Reported by the provider |
| 0.10 Wh per prompt | Google, 21 August 2025 | Same prompt, active TPU and GPU consumption only. Google calls it an optimistic scenario | Reported by the provider |
| 0.34 Wh per query | Sam Altman, June 2025 | Average ChatGPT query. No method, token count or boundary stated | Stated, no method |
| 2.9 Wh per ChatGPT request, against 0.3 Wh per Google search | EPRI white paper, May 2024 (reported by Utility Dive, 30 May 2024) | No method or boundary given in the report | Estimate, no method |
How to quote this group. Give the model, the token counts and whether the figure covers only the GPU or the whole facility. Google's own two methods differ by a factor of 2.4 on the same prompt, from 0.10 to 0.24 Wh. Never write "a ChatGPT query uses X Wh" as a measured fact, since the ChatGPT figures above are modelled or stated without method. The figures guide sets the same sources side by side.
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 0.03 Wh per exchange | SOMA AI, SAEF dataset | Class A, small models, 1210 tokens, facility energy, training excluded from the energy figure | Estimated |
| 0.099 Wh per exchange | SOMA AI, SAEF dataset | Class B, mid-size models, 1210 tokens, facility energy, training excluded from the energy figure | Estimated |
| 0.37 Wh per exchange | SOMA AI, SAEF dataset | Class C, frontier models, 1210 tokens, facility energy, training excluded from the energy figure | Estimated |
How much carbon does one AI query emit?
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 0.03 g CO₂e per prompt | Google, 21 August 2025 | Median Gemini Apps text prompt. Market-based factor of 94 g per kWh, scope 1 and embodied emissions included, training excluded | Reported by the provider |
| 0.02 g CO₂e per prompt | Google, 21 August 2025 | Same prompt on the narrower active-chip method | Reported by the provider |
| 1.14 g CO₂e per response | Mistral AI with Carbone 4 and ADEME, 22 July 2025 | Le Chat, 400-token response. Marginal inference impacts including upstream manufacturing, excluding users' devices | Reported, lifecycle study |
| 2.2 g CO₂e per query | Tomlinson, Black, Patterson and Torrance, 8 March 2023 | ChatGPT in 2023. 1.84 g of GPT-3 training plus 0.382 g of operation, resting on an informal estimate and an assumed 10 million queries a day | Modelled |
| Under 0.3 to about 17 g CO₂e per query | Jegham et al., 14 May 2025 | Prompts of 10,000 input and 1,500 output tokens. Energy times regional carbon intensity. The 17 g is DeepSeek-R1, and 2.5 g on Azure | Modelled |
| 0.4 and 3.2 g CO₂ per request | Berget AI, 21 October 2026 | Kimi K3, 5,000 input and 387 output tokens, B300 node in Sweden and Texas. Location-based electricity plus hardware manufacturing, training excluded | Company press release |
| 0.4 g CO₂ per query | GreenPT, 22 April 2026 | Typical chat reply, no model named. The page says both measured and derived from US hyperscaler workloads with EU grid intensity | Company figure, basis unclear |
| 1.25 to 1,325.12 g CO₂e per 500 questions | Dauner and Socher, 19 June 2025 | Open models on one A100 GPU at 480 g per kWh. The range joins two question phases. Cogito 70B with reasoning emits 1,341.1 g per 1,000 questions | GPU measured, fixed grid factor |
How to quote this group. State the grid and whether hardware manufacturing is in. Google's 0.03 g uses a market-based factor of 94 g per kWh, while the Dauner and Socher figures use a fixed 480 g per kWh, which alone is a fivefold difference before any model effect. The Berget AI row is a company press release, and the GreenPT page words its method inconsistently, so do not call that figure measured.
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 0.01 g and 0.0025 g | SOMA AI, SAEF dataset | Class A, 1210 tokens, in US East - N. Virginia and EU North - Sweden respectively | Estimated |
| 0.036 g and 0.0094 g | SOMA AI, SAEF dataset | Class B, 1210 tokens, in US East - N. Virginia and EU North - Sweden respectively | Estimated |
| 0.13 g and 0.037 g | SOMA AI, SAEF dataset | Class C, 1210 tokens, in US East - N. Virginia and EU North - Sweden respectively | Estimated |
How much water does AI use?
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 0.26 mL per prompt | Google (Elsworth et al.), 21 August 2025 | Median Gemini Apps text prompt at a fleet WUE of 1.15 L per kWh | Reported by the provider |
| 0.12 mL per prompt | Google, 21 August 2025 | Same prompt, active chips only | Reported by the provider |
| 0.32 mL per query (0.000085 US gallons) | Sam Altman, June 2025 | Average ChatGPT query. No method stated | Stated, no method |
| 16.904 mL per request | Li, Yang, Islam and Ren, 6 April 2023 | GPT-3, medium-length response, US average. 2.2 mL on-site plus 14.704 mL for electricity. Embodied water excluded | Modelled |
| Under 4 to over 200 mL per query | Jegham et al., 14 May 2025 | Prompts of 10,000 input and 1,500 output tokens. The top end is DeepSeek-R1 on DeepSeek's servers, and 34 mL on Azure | Modelled |
| 45 mL per response | Mistral AI with Carbone 4 and ADEME, 22 July 2025 | Le Chat, 400-token response. Includes upstream impacts, and the water scope is not stated | Reported, lifecycle study |
How to quote this group. Say which water is in: on-site cooling alone, or cooling plus the power plants behind the grid. Li and co-authors find most of their 16.9 mL is off-site, so Google's figure, which is computed from data-centre WUE and does not state that power-plant water is included, cannot be set against it. Withdrawal and consumption are different quantities. The water guide explains the three kinds of water. For comparison, SAEF gives 0.24 mL for a class B exchange served from US East - N. Virginia and 0.59 mL from EU North - Sweden, on-site plus generation water.
How much carbon and energy does training an AI model take?
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 24.7 t CO₂e, or 50.5 t CO₂e | Luccioni, Viguier and Ligozat, 3 November 2022 | BLOOM 176B, final training only. 24.7 t is dynamic power alone, 50.5 t adds equipment manufacturing and idle power | Modelled |
| 552 t CO₂e and 1,287 MWh | Patterson et al., 21 April 2021 | GPT-3 175B, from OpenAI-supplied inputs on 10,000 V100 GPUs for 14.8 days. The authors' calculation, not an OpenAI figure | Estimated |
| 626,155 lb CO₂e, or about 284 t | Strubell, Ganesh and McCallum, 5 June 2019 | One neural architecture search run for a big Transformer. Patterson et al. give a much lower alternative in their appendix. Conversion at 0.45359237 kg per lb | Estimated, contested upper case |
| 420, 2,040 and 8,930 tons CO₂e (as written on the model card) | Meta model card, 2024 | Training of Llama 3.1 8B, 70B and 405B, location-based. Market-based is 0 because Meta matches its electricity with renewables | Reported by the developer |
| 20.4 kt CO₂e | Mistral AI with Carbone 4 and ADEME, 22 July 2025 | Mistral Large 2, as of January 2025 and after 18 months of use. The boundary is not itemised | Reported, lifecycle study |
| 281,000 m³ of water consumed | Mistral AI with Carbone 4 and ADEME, 22 July 2025 | Same sentence and caveat as the 20.4 kt | Reported, lifecycle study |
| 700,000 L directly evaporated, 5.4 million L in total | Li, Yang, Islam and Ren, 6 April 2023 | Training GPT-3 in Microsoft's US data centres. The 700,000 L is on-site water only. The paper says it can directly evaporate | Modelled |
How to quote this group. Say whether the figure is final training only or a lifecycle total, and whether it is location-based. The 24.7 t and 50.5 t for BLOOM are the same run on two boundaries. The Llama 3.1 market-based zero reflects renewable matching, not an absence of electricity use. Per-query shares of training are covered on the page about training footprints.
How much electricity and water do AI and data centres use in total?
| Figure | Publisher and date | What it counts | Basis |
|---|---|---|---|
| 176 TWh in 2023, 4.4% of US electricity | Lawrence Berkeley National Laboratory via US DOE, 20 December 2024 | All US data centres, all uses, not only AI | Estimated |
| 325 to 580 TWh in 2028, 6.7% to 12% | Lawrence Berkeley National Laboratory via US DOE, 20 December 2024 | Scenario range for total US data-centre electricity, all uses | Projection |
| 415 TWh in 2024 and around 945 TWh by 2030 | International Energy Agency, April 2025 | All data centres worldwide, 1.5% of world electricity in 2024. The 2030 value is a projection | Estimate and projection |
| 85 to 134 TWh a year by 2027 | de Vries, Joule, 18 October 2023 | About 1.5 million NVIDIA AI servers running at full power. A scenario | Scenario |
| 29.2 TWh a year | de Vries, Joule, 18 October 2023 | Hypothetical case where every Google search uses AI, from a SemiAnalysis server count | Scenario |
| 4.2 to 6.6 billion m³ in 2027 | Li, Yang, Islam and Ren, 6 April 2023 | Global AI water withdrawal, not consumption | Projection |
| 10.9 billion US gallons in 2025 | Google environmental report, 2026 | Water consumed across all Google data centres and offices, not AI alone | Reported by the company |
How to quote this group. Data-centre totals are not AI totals. The IEA and Lawrence Berkeley figures count every workload, and the de Vries figures are scenarios built on server shipments. Label each as an estimate, a projection or a scenario. The page on whether AI is bad for the environment discusses how per-query and total figures fit together.
What about images and video?
MIT Technology Review reports about 1,141 J of GPU energy for one standard 1024 by 1024 image from Stable Diffusion 3 Medium, and an estimated 2,282 J, or 0.63 Wh, once doubled for the rest of the system. For a five-second CogVideoX video it reports about 3.4 million J, which is 0.94 kWh at 3.6 million J per kWh. Both are from 20 May 2025, and the doubling is an assumption. Details are on the image and video page.
What should a company report?
A company does not report these statistics. It reports its own use, converted with a stated method. The free calculator turns a token export into an estimate with a range, and the Scope 3 guide shows where the line sits in a CSRD inventory and which data tier to use. Any public statement should name the factor version, the region and the model class, and say that it is an estimate. The equations are on the methodology page.
Where do the factors come from, and how should they be cited?
Figures per query are one plain exchange, before the Tier 2b reasoning uplift. 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 (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). 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, and the whole table is a CSV download.
Keep reading
- How much energy does a ChatGPT query use? Every published figure for one query, with the SAEF estimate
- Why AI footprint estimates differ: published figures, dated. The same sources, set against each other
- AI water footprint: the three kinds of water. On-site, electricity and scarcity