Sources and method

Every number in this tool comes from one of the 231 sources below. Numbers in square brackets elsewhere on the site link here. In Pipeline you can see every individual data point, with the table, page or quote it came from.

Watch and read more

Talks and articles on the same question, for anyone who wants more than the numbers. None of the tool's figures come from these.

  1. [219]
    AI & CO2 - Webinar 9 oktober - Alumniweek 2026in Dutchvideo
    Universiteit van Amsterdam, 2026. Checked 2026-10-16.
    A 53-minute webinar in Dutch on the energy and CO₂ of AI, recorded for the UvA alumni week on 9 October 2026 (uploaded the same day). Covers the same questions as this tool, with more room for the story.

How the calculation works, and what it leaves out

  • Tokens. Each scenario is a fixed, pre-written conversation. Nothing is sent to an AI. Token counts follow published statistics (WildChat, LMSYS-Chat-1M, SWE-bench agent runs, Claude Code usage logs). Every turn re-reads the conversation so far, which is why long chats and agents use so much input.
  • Energy. Output tokens times the energy per output token. Input counts at a fraction of that, and cached input at a fraction again. Almost no provider publishes per-token energy, so a model gets the published measurements of comparable models. They differ by more than a factor of ten, mostly because of hardware and how many requests a server handles at once. The tool starts from the measurement we trust most for each model, says why, and shows you the whole spread so you can pick another.
  • Defaults. Every number is a set of published data points. For each one we picked the point we trust most (the most recent official figure, a measurement over an estimate, the scope that matches the rest of the sum) and wrote down why; that reason sits next to every setting. The median, the mean, any single point or your own value are one click away.
  • Comparisons. "Compare with" turns the result into a trip (the CO₂e as a distance with the chosen vehicle), into phone charges, laptop charges, a video call or an e-bike ride, counted by electricity and, charged at the departure place, by CO₂e, or into a slice of a whole industry: how many seconds of a year of the global livestock, Dutch livestock or textile industry it equals, or what fraction of everything Bitcoin has ever used, by energy and by the industry's own reported CO₂e. The water total is shown as showers, baths or flushes. Each of those is a sourced data set too, under Pipeline → Comparison and Task.
  • Images. The image scenario counts a fixed number of images times the measured energy per image of the chosen image model (GPU energy, with the datacenter overhead added once), and the list price per image where there is one. No tokens are counted. Open models have been measured by outside researchers at a stated size and step count; the closed ones publish nothing, so their points are third-party estimates.
  • Datacenter overhead. Energy figures are stored without overhead and multiplied by the datacenter's PUE once. Where a source already included overhead, we divided it out (the conversion is shown with the data point).
  • Grid. Annual average carbon intensity of the local grid. Sources mix lifecycle figures (including building power plants) and direct ones (burning fuel only); both are shown. Company green-energy contracts are not counted, because they do not change what the local grid emits at that moment.
  • Training. A separate scenario, not a surcharge on conversations: one published training run, counted as accelerator-hours times power per chip (what Meta, DeepMind and Ai2 disclose, and what BLOOM metered), with the overhead and grid of the datacenter you pick. For GPT-4 and Grok 4 nothing is disclosed; those points are reconstructions by outside analysts from leaked hardware counts. Which chips and which site each run used is listed under Pipeline → Model → Training run, because a V100 in 2020 and an H100 in 2024 do very different amounts of work per watt.
  • Left out by default. Making the hardware (you can switch that on), the network and your own device, and for training runs the development and ablation runs before the final one, which add about half again where they have been counted. Expect the result to be off by a factor of two either way: treat it as an order of magnitude.
  • Travel. Car figures are per vehicle with one person driving, public transport and flights per passenger at average occupancy. The "whole plane" and "whole ship" cards count the entire aircraft or vessel per km, as if you were its only passenger: fuel burn per km times the fuel's emission factor, or the per-seat factor times the passengers the source assumes on board. Sources differ in scope (tailpipe only, including fuel production, or full lifecycle). Distances are straight-line distance times a detour factor. Cars, buses and trains can only go where land connects.
  • Prices are list prices for API use. Consumer subscriptions (ChatGPT, Claude.ai) are paid monthly, so their real cost per conversation is different.

AI energy, water and efficiency

  1. [1]
    The ML.ENERGY Leaderboard v3.0 (data: public/data/tasks/*.json, last_updated 2026-02-16)
    Jae-Won Chung, Jeff J. Ma, Ruofan Wu, Jiachen Liu, Oh Jun Kweon, Yuxuan Xia, Zhiyu Wu, Mosharaf Chowdhury. The ML.ENERGY Initiative (University of Michigan), 2025. Checked 2026-10-02.
    Used for: model energy · image: Stable Diffusion 3.5 Large · image: FLUX.1 [dev]
  2. [2]
    How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference (arXiv:2505.09598v6)
    Nidhal Jegham, Marwan Abdelatti, Chan Young Koh, Lassad Elmoubarki, Abdeltawab Hendawi. arXiv, 2025. Checked 2026-10-02.
    Used for: model energy · PUE Tencent Cloud ap-shanghai · PUE xAI colossus-1 · PUE xAI colossus-2 · Datacenter overhead (PUE)
  3. [3]
    Measuring the environmental impact of delivering AI at Google Scale (arXiv:2508.15734)
    Cooper Elsworth, Keguo Huang, David Patterson, Ian Schneider, Robert Sedivy, Savannah Goodman, Ben Townsend, Parthasarathy Ranganathan, Jeff Dean, Amin Vahdat, Ben Gomes, James Manyika. Google / arXiv, 2025. Checked 2026-10-02.
    Used for: model energy · water Google Cloud europe-west4 · water Google Cloud europe-north1 · water Google Cloud europe-west9 · water Google Cloud europe-central2 · water Google Cloud us-west1 · and 22 more
  4. [4]
    How much energy does ChatGPT use?
    Josh You. Epoch AI, 2025. Checked 2026-10-02.
    Used for: model energy · Input token vs output token
  5. [5]
    The Gentle Singularity
    Sam Altman. Sam Altman (blog), 2025. Checked 2026-10-02.
    Used for: model energy
  6. [6]
    Our contribution to a global environmental standard for AI
    Mistral AI. Mistral AI (with Carbone 4 and ADEME), 2025. Checked 2026-10-02.
  7. [7]
    Energy Use of AI Inference, Efficiency Pathways, and Test-Time Scaling (arXiv:2509.20241; Joule 2026)
    Felipe Oviedo, Fiodar Kazhamiaka, Esha Choukse, Allen Kim, Amy Luers, Melanie Nakagawa, Ricardo Bianchini, Juan M. Lavista Ferres. Joule / arXiv (Microsoft), 2026. Checked 2026-10-02.
    Used for: model energy
  8. [8]
    From Words to Watts: Benchmarking the Energy Costs of Large Language Model Inference (arXiv:2310.03003)
    Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, Vijay Gadepally. IEEE HPEC / arXiv, 2023. Checked 2026-10-02.
    Used for: model energy
  9. [9]
    Power Hungry Processing: Watts Driving the Cost of AI Deployment? (arXiv:2311.16863v3, ACM FAccT 2024)
    Alexandra Sasha Luccioni, Yacine Jernite, Emma Strubell. ACM FAccT / arXiv, 2024. Checked 2026-10-02.
    Used for: model energy · image: Stable Diffusion XL · scenario: Generate an image
  10. [10]
    Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model (arXiv:2211.02001)
    Alexandra Sasha Luccioni, Sylvain Viguier, Anne-Laure Ligozat. JMLR / arXiv, 2022. Checked 2026-10-02.
    Used for: training: OPT-175B · training: BLOOM · Hardware manufacturing (embodied) · scenario: Train a model
  11. [11]
    From Prompts to Power: Measuring the Energy Footprint of LLM Inference (arXiv:2511.05597)
    Francisco Caravaca, Angel Cuevas, Ruben Cuevas. arXiv, 2025. Checked 2026-10-02.
    Used for: Input token vs output token
  12. [12]
    A Systematic Characterization of LLM Inference on GPUs (arXiv:2512.01644)
    Haonan Wang, Xuxin Xiao, Mingyu Yan, et al.. arXiv, 2025. Checked 2026-10-02.
    Used for: Input token vs output token
  13. [13]
    Characterization of Request and Token Energy Costs for LLM Inference Workloads on GPU Platforms (arXiv:2608.28044v1)
    (see arXiv listing). arXiv, 2026. Checked 2026-10-02.
    Used for: model energy · Input token vs output token
  14. [14]
    Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models (arXiv:2304.03271v5)
    Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren. arXiv / Communications of the ACM, 2025. Checked 2026-10-02.
    Used for: power-plant water: Finland · power-plant water: Netherlands · power-plant water: Ireland · power-plant water: USA (PJM Interconnection: Virginia, Ohio, ...) · power-plant water: USA (Tennessee Valley Authority: Memphis, Nashville, ...) · power-plant water: USA (Bonneville Power Administration, Oregon/Washington) · and 28 more
  15. [15]
    Data center efficiency (Google Data Centers)
    Google. Google, 2026. Checked 2026-10-02.
    Used for: PUE Google Cloud europe-west4 · PUE Google Cloud europe-north1 · PUE Google Cloud europe-west9 · PUE Google Cloud europe-central2 · PUE Google Cloud us-west1 · PUE Google Cloud us-west2 · and 10 more
  16. [16]
    Measuring energy and water efficiency for Microsoft datacenters
    Microsoft. Microsoft, 2025. Checked 2026-10-02.
    Used for: PUE Azure westeurope · water Azure westeurope · PUE Azure northeurope · water Azure northeurope · PUE Azure norwayeast · water Azure norwayeast · and 34 more
  17. [17]
    Inside Microsoft's two-decade push to cut water intensity while scaling for growth
    Microsoft. Microsoft, 2026. Checked 2026-10-02.
    Used for: water Google Cloud europe-west4 · water Google Cloud europe-north1 · water Google Cloud europe-west9 · water Google Cloud europe-central2 · water Google Cloud us-west1 · water Google Cloud us-west2 · and 16 more
  18. [18]
    AWS Sustainability
    AWS. Amazon Web Services, 2026. Checked 2026-10-02.
    Used for: water Google Cloud europe-west4 · water Google Cloud europe-north1 · water Google Cloud europe-west9 · water Google Cloud europe-central2 · water Google Cloud us-west1 · water Google Cloud us-west2 · and 20 more
  19. [19]
    Meta 2025 Sustainability Report - Environmental Data Index
    Meta. Meta, 2025. Checked 2026-10-02.
    Used for: water Google Cloud europe-west4 · water Google Cloud europe-north1 · water Google Cloud europe-west9 · water Google Cloud europe-central2 · water Google Cloud us-west1 · water Google Cloud us-west2 · and 20 more
  20. [20]
    Claude API pricing
    Anthropic. Anthropic, 2026. Checked 2026-10-02.
    Used for: price of Claude Opus 5.5 · price of Claude Sonnet 5.5 · price of Claude Haiku 4.5 · Cached input vs fresh input · Price of one web search
  21. [21]
    OpenAI API pricing
    OpenAI. OpenAI, 2026. Checked 2026-10-02.
    Used for: price of GPT-6 Astra · price of GPT-6 Luna · price of GPT-image-1 · price of DALL·E 3 · Cached input vs fresh input · Price of one web search
  22. [22]
    Gemini Developer API pricing
    Google. Google, 2026. Checked 2026-10-02.
    Used for: price of Gemini 3.1 Pro (Preview) · price of Gemini 3.8 Flash · Cached input vs fresh input · Price of one web search
  23. [23]
    DeepSeek API: Models & Pricing
    DeepSeek. DeepSeek, 2026. Checked 2026-10-02.
    Used for: price of DeepSeek-V4-Pro (0813) · Cached input vs fresh input

Models and prices

  1. [24]
    GPT-6 Astra model page
    OpenAI, 2026. Checked 2026-10-02.
  2. [25]
    GPT-6 Luna model page
    OpenAI, 2026. Checked 2026-10-02.
  3. [26]
    OpenAI API Changelog
    OpenAI, 2026. Checked 2026-10-02.
  4. [27]
    gpt-oss-120b model page
    OpenAI, 2025. Checked 2026-10-02.
  5. [28]
    Claude Models overview
    Anthropic, 2026. Checked 2026-10-02.
  6. [29]
    Claude Opus 5.5 overview
    Anthropic, 2026. Checked 2026-10-02.
  7. [30]
    Claude Sonnet 5.5 overview
    Anthropic, 2026. Checked 2026-10-02.
  8. [31]
    Gemini 3.8 Flash model page
    Google, 2026. Checked 2026-10-02.
  9. [32]
    Gemini 3.1 Pro Preview model page
    Google, 2026. Checked 2026-10-02.
  10. [33]
    Gemini API release notes
    Google, 2026. Checked 2026-10-02.
  11. [34]
    xAI Models and Pricing
    xAI, 2026. Checked 2026-10-02.
    Used for: price of Grok 4.7
  12. [35]
    Grok 4.7 model page
    xAI, 2026. Checked 2026-10-02.
  13. [36]
    DeepSeek-V4-Pro model card
    DeepSeek / Hugging Face, 2026. Checked 2026-10-02.
  14. [37]
    DeepSeek-V4-Pro-0813 model card
    DeepSeek / Hugging Face, 2026. Checked 2026-10-02.
  15. [38]
    Alibaba Cloud Model Studio model pricing
    Alibaba Cloud, 2026. Checked 2026-10-02.
    Used for: price of Qwen3.8-Max (Qwen3.8-2.4T-A95B)
  16. [39]
    Qwen3.8-2.4T-A95B model card
    Qwen / Hugging Face, 2026. Checked 2026-10-02.
  17. [40]
    Meta Model API pricing and rate limits
    Meta, 2026. Checked 2026-10-02.
    Used for: price of Muse Spark 1.3 · Price of one web search
  18. [41]
    Meta Model API models
    Meta, 2026. Checked 2026-10-02.
  19. [42]
    Muse-Glimmer-30B model card
    Meta / Hugging Face, 2026. Checked 2026-10-02.
  20. [43]
    Together AI pricing
    Together AI, 2026. Checked 2026-10-02.
    Used for: price of Muse Glimmer 30B · price of gpt-oss-120b
  21. [44]
    Mistral API pricing
    Mistral AI, 2026. Checked 2026-10-02.
    Used for: price of Mistral Large 3 · price of Mistral Small 4 · Price of one web search
  22. [45]
    Mistral-Large-3-675B-Instruct-2512 model card
    Mistral AI / Hugging Face, 2025. Checked 2026-10-02.
  23. [46]
    Mistral-Small-4-119B-2603 model card
    Mistral AI / Hugging Face, 2026. Checked 2026-10-02.
  24. [47]
    Kimi API Model Inference Pricing
    Moonshot AI, 2026. Checked 2026-10-02.
    Used for: price of Kimi K3
  25. [48]
    Kimi-K3 model card
    Moonshot AI / Hugging Face, 2026. Checked 2026-10-02.
  26. [49]
    Z.AI Pricing
    Z.AI (Zhipu), 2026. Checked 2026-10-02.
    Used for: price of GLM-5.3
  27. [50]
    GLM-5.3 guide
    Z.AI (Zhipu), 2026. Checked 2026-10-02.
  28. [51]
    GLM-5.3 model card
    Z.AI / Hugging Face, 2026. Checked 2026-10-02.
  29. [52]
    Epoch AI: Data on AI Models (all_ai_models.csv)
    Epoch AI, 2026. Checked 2026-10-02.

Datacenters and electricity grids

  1. [53]
    Carbon free energy for Google Cloud regions
    Google Cloud, 2025. Checked 2026-10-02.
    Used for: grid: Sweden (SE3, Stockholm) · grid: Finland · grid: France · grid: Canada (Québec) · grid: Netherlands · grid: Great Britain · and 14 more
  2. [54]
    AWS Cloud – Amazon Sustainability (PUE/WUE by geography)
    Amazon, 2025. Checked 2026-10-02.
    Used for: PUE AWS us-east-1 · water AWS us-east-1 · PUE AWS us-east-2 · water AWS us-east-2 · PUE AWS us-west-2 · water AWS us-west-2 · and 8 more
  3. [55]
    PUE, WUE, REF values – Fiscal Year 2025
    OVHcloud, 2025. Checked 2026-10-02.
    Used for: PUE OVHcloud GRA · water OVHcloud GRA
  4. [56]
    Environmental Leadership (PUE/WUE per data center)
    Scaleway, 2024. Checked 2026-10-02.
    Used for: PUE Scaleway fr-par-2 · water Scaleway fr-par-2
  5. [57]
    Sustainability at Hetzner
    Hetzner Online, 2025. Checked 2026-10-02.
    Used for: PUE Hetzner hel1 · PUE Hetzner fsn1 · PUE Hetzner ash
  6. [58]
    Alibaba Group FY2025 ESG Report (Form 6-K exhibit 99.2)
    Alibaba Group, 2025. Checked 2026-10-02.
    Used for: PUE Alibaba Cloud cn-beijing · PUE Alibaba Cloud cn-hangzhou
  7. [59]
    Huawei Cloud: Improving PUE to Build Green Data Centers
    Huawei, 2026. Checked 2026-10-02.
    The page carries no date; the year is when we checked it.
    Used for: PUE Huawei Cloud cn-southwest-2
  8. [60]
    Electricity Maps yearly zone carbon intensity (2024, 2025), republished in CO2.js under ODbL
    Electricity Maps / Green Web Foundation, 2026. Checked 2026-10-02.
    Used for: grid: Sweden (SE3, Stockholm) · grid: Norway (NO1, Oslo) · grid: Finland · grid: Iceland · grid: France · grid: Canada (Québec) · and 25 more
  9. [61]
  10. [62]
    Lifecycle carbon intensity of electricity (Ember yearly electricity data, via Our World in Data)
    Ember / Our World in Data, 2026. Checked 2026-10-02.
    Used for: grid: Sweden (SE3, Stockholm) · grid: Norway (NO1, Oslo) · grid: Finland · grid: Iceland · grid: France · grid: Netherlands · and 18 more
  11. [63]
    eGRID2023 Summary Tables (rev2) – Subregion Output Emission Rates
    US EPA, 2025. Checked 2026-10-02.
    Used for: grid: USA (PJM Interconnection: Virginia, Ohio, ...) · grid: USA (Tennessee Valley Authority: Memphis, Nashville, ...) · grid: USA (Bonneville Power Administration, Oregon/Washington) · grid: USA (California ISO) · grid: USA (New York ISO) · grid: USA (ISO New England)
  12. [64]
    Low-Carbon Power country pages (lifecycle carbon intensity, IPCC factors)
    lowcarbonpower.org, 2026. Checked 2026-10-02.
    Used for: grid: Sweden (SE3, Stockholm) · grid: Norway (NO1, Oslo) · grid: Finland · grid: Iceland · grid: France · grid: Netherlands · and 16 more
  13. [65]
    Bilan électrique 2025 – Émissions
    RTE, 2026. Checked 2026-10-02.
    Used for: grid: France
  14. [66]
    CO2-Emissionen pro Kilowattstunde Strom 2025
    Umweltbundesamt, 2026. Checked 2026-10-02.
    Used for: grid: Germany
  15. [67]
    Wskaźniki emisyjności CO2 ... dla energii elektrycznej ... za 2024 rok
    KOBiZE, 2025. Checked 2026-10-02.
    Used for: grid: Poland
  16. [68]
    CO2 Baseline Database for the Indian Power Sector, User Guide Version 20.0
    Central Electricity Authority (India), 2024. Checked 2026-10-02.
    Used for: grid: India (Western region, Mumbai)
  17. [69]
    Greenhouse Gas Emissions Accounting and Reporting Guidelines (Table 16, NGA scope 2 factors)
    NSW DCCEEW, 2025. Checked 2026-10-02.
    Used for: grid: Australia (New South Wales)
  18. [70]
    2023年电力二氧化碳排放因子 (MEE press briefing)
    Ministry of Ecology and Environment (China), 2026. Checked 2026-10-02.
    Used for: grid: China (national)
  19. [71]
    关于发布2022年电力二氧化碳排放因子的公告
    Ministry of Ecology and Environment / NBS (China), 2024. Checked 2026-10-02.
    Used for: grid: China (national)
  20. [72]
    GHG Emission Rate Associated with Residual Electricity Supplies 1990–2025
    Hydro-Québec, 2026. Checked 2026-10-02.
    Used for: grid: Canada (Québec)
  21. [227]
    Yearly electricity data (release_generation_yearly_global.csv), share of generation by source
    Ember, 2026. Checked 2026-10-25.
    CC-BY-4.0. Generation only: imports are not in the mix.
    Used for: power-plant water: Sweden (SE3, Stockholm) · grid mix: Sweden (SE3, Stockholm) · power-plant water: Norway (NO1, Oslo) · grid mix: Norway (NO1, Oslo) · power-plant water: Finland · grid mix: Finland · and 42 more
  22. [228]
    Hourly Electric Grid Monitor (EIA-930), balancing authority net generation by energy source, 2025 six-month files
    US Energy Information Administration, 2026. Checked 2026-10-25.
    Used for: power-plant water: USA (PJM Interconnection: Virginia, Ohio, ...) · grid mix: USA (PJM Interconnection: Virginia, Ohio, ...) · power-plant water: USA (Tennessee Valley Authority: Memphis, Nashville, ...) · grid mix: USA (Tennessee Valley Authority: Memphis, Nashville, ...) · power-plant water: USA (Bonneville Power Administration, Oregon/Washington) · grid mix: USA (Bonneville Power Administration, Oregon/Washington) · and 6 more
  23. [229]
    Provincial and Territorial Energy Profiles: Quebec
    Canada Energy Regulator, 2024. Checked 2026-10-25.
    Used for: power-plant water: Canada (Québec) · grid mix: Canada (Québec)
  24. [230]
    Behind-the-meter datacenters: Colossus (xAI, Memphis and Southaven)
    Cleanview, 2026. Checked 2026-10-25.
    Satellite imagery from 13 May 2026.
    Used for: on-site power xAI colossus-1 · on-site power xAI colossus-2
  25. [231]
    Inside Memphis’ fight against xAI (timeline)
    Southern Environmental Law Center, 2026. Checked 2026-10-25.
    Used for: on-site power xAI colossus-1

Transport

  1. [73]
    Greenhouse gas reporting: conversion factors 2026 (flat file, revised)
    UK Department for Energy Security and Net Zero, 2026. Checked 2026-10-02.
    Used for: Small petrol car · Medium petrol car · Large SUV or pickup · Electric car · Motorbike · City bus · and 8 more
  2. [74]
    2026 Government greenhouse gas conversion factors for company reporting: Methodology paper
    UK Department for Energy Security and Net Zero, 2026. Checked 2026-10-02.
    Used for: Plane, short-haul (whole plane) · Plane, long-haul (whole plane) · Flight detour factor
  3. [75]
    GHG Emission Factors Hub 2025 (xlsx)
    US Environmental Protection Agency, 2025. Checked 2026-10-02.
    Used for: Medium petrol car · Large SUV or pickup · Motorbike · City bus · Train · Plane, short-haul · and 1 more
  4. [76]
    fueleconomy.gov vehicle data (vehicles.csv) and field descriptions
    US DOE / US EPA (fueleconomy.gov), 2026. Checked 2026-10-02.
    Used for: Small petrol car · Medium petrol car · Large SUV or pickup · Electric car
  5. [77]
    2025 Tesla Model 3 Long Range RWD - fueleconomy.gov
    US DOE / US EPA (fueleconomy.gov), 2025. Checked 2026-10-02.
  6. [78]
    Impact CO2 - Transport comparator (ADEME Base Empreinte data)
    ADEME, 2026. Checked 2026-10-02.
    Used for: Small petrol car · Medium petrol car · Motorbike · City bus · Long-distance coach · Train · and 3 more
  7. [79]
    Life-cycle greenhouse gas emissions from passenger cars in the European Union: A 2025 update and key factors to consider
    International Council on Clean Transportation (Negri & Bieker), 2025. Checked 2026-10-02.
    Used for: Medium petrol car · Electric car
  8. [80]
    What if I told you cruising is worse for the climate than flying?
    International Council on Clean Transportation, 2022. Checked 2026-10-02.
    Used for: Cruise ship · Plane, short-haul
  9. [81]
    Vergleich der durchschnittlichen Emissionen einzelner Verkehrsmittel im Personenverkehr in Deutschland 2024 (TREMOD 6.71B)
    Umweltbundesamt (German Environment Agency), 2025. Checked 2026-10-02.
    Used for: Medium petrol car · City bus · Long-distance coach · Train · Plane, short-haul
  10. [82]
    Methodology for GHG Efficiency of Transport Modes - Final Report (Fraunhofer ISI & CE Delft for the EEA)
    Fraunhofer ISI / CE Delft for the European Environment Agency, 2020. Checked 2026-10-02.
    Used for: Medium petrol car · City bus · Long-distance coach · Train · High-speed train · Ferry · and 3 more
  11. [83]
    Which form of transport has the smallest carbon footprint? (chart: carbon-footprint-travel-mode)
    Our World in Data, 2023. Checked 2026-10-02.
    Used for: Medium petrol car · High-speed train · Ferry · Plane, short-haul · Plane, long-haul
  12. [84]
  13. [85]
    JR Central Annual Report 2025 - Environment section (Tokaido Shinkansen vs aircraft)
    Central Japan Railway Company, 2025. Checked 2026-10-02.
    Used for: High-speed train
  14. [86]
    JR Central - Operating Area
    Central Japan Railway Company, 2026. Checked 2026-10-02.
  15. [87]
    Tesla Impact Report 2024 (extended version, Chinese-language edition)
    Tesla, 2025. Checked 2026-10-02.
    Used for: Large SUV or pickup
  16. [88]
    Stromverbrauch Elektroautos: ADAC Test
    ADAC, 2026. Checked 2026-10-02.
    Used for: Electric car
  17. [89]
    Energy consumption of full electric vehicles (cheatsheet)
    EV Database, 2026. Checked 2026-10-02.
    Used for: Electric car
  18. [90]
    CTL.SC1x Lesson: One to Many Distribution - Selected values of kCF (citing Ballou, Rahardja & Sakai 2002, Transportation Research Part A 36(9):843-848)
    MIT Center for Transportation & Logistics (edX), 2015. Checked 2026-10-02.
    Used for: Road detour factor
  19. [91]
    Selected country circuity factors for road travel distance estimation (Ballou, Rahardja & Sakai)
    Transportation Research Part A 36(9) (TRID record), 2002. Checked 2026-10-02.
  20. [92]
    A nationwide comparison of driving distance versus straight-line distance to hospitals (Boscoe, Henry & Zdeb)
    The Professional Geographer 64(2), 2012. Checked 2026-10-02.
    Used for: Road detour factor
  21. [93]
    Driving the Extra Mile: Comprehensive Analysis of Road and Rail Transportation Distances in Europe (Heinold & Makowski)
    Networks and Spatial Economics (Springer), 2026. Checked 2026-10-02.
    Used for: Road detour factor · Rail detour factor
  22. [94]
  23. [95]
    Performance Review Report (PRR) 2025
    EUROCONTROL Performance Review Commission, 2026. Checked 2026-10-02.
    Used for: Flight detour factor
  24. [96]
    ICAO Carbon Emissions Calculator Methodology, Version 8 (May 2015)
    ICAO, 2015. Checked 2026-10-02.
    Used for: Flight detour factor
  25. [97]
  26. [98]
    Weekly Oil Bulletin - Weekly prices with taxes (prices in force 2026-09-28)
    European Commission DG ENER, 2026. Checked 2026-10-02.
  27. [99]
    Weekly U.S. All Grades All Formulations Retail Gasoline Prices
    US Energy Information Administration, 2026. Checked 2026-10-02.
  28. [189]
    EMEP/EEA air pollutant emission inventory guidebook 2023, chapter 1.A.3.a Aviation, with Annex 1: Master emissions calculator (v1.5, September 2024)
    European Environment Agency / EUROCONTROL, 2023. Checked 2026-10-16.
    Fuel per landing-and-take-off cycle and per climb/cruise/descent stage (125 to 8,000 nautical miles) by aircraft type, from EUROCONTROL's Advanced Emission Model (baseline 2022). A320 at 750 nm: 4,931 kg; 737-800: 5,098 kg; 787-9 at 4,000 nm: 43,812 kg; 777-300ER: 70,988 kg. The calculator's own CO₂ column uses 3.15 kg per kg of jet fuel.
    Used for: Plane, short-haul (whole plane) · Plane, long-haul (whole plane)
  29. [190]
    ICAO Carbon Emissions Calculator Methodology, Version 13.1
    International Civil Aviation Organization, 2024. Checked 2026-10-16.
    3.16 tonnes of CO₂ per tonne of jet fuel burnt, the factor used for the whole-plane fuel-burn points. CO₂ only; says nothing about non-CO₂ effects.
  30. [191]
    Basic analysis of energy-saving domestic marine unit load transportation (NILIM Annual Report 2014, p. 81)
    Tomoko Sasaki, Tomihiro Watanabe. National Institute for Land and Infrastructure Management, Japan, 2014. Checked 2026-10-16.
    2012 operator survey of 22 domestic ferries: 67 kg of fuel per km at 37.1 km/h and a 76% load factor (RoRo ships 44 kg/km, container ships 15 kg/km). Full data in Technical Note of NILIM No. 716 (Japanese). The fuel type is not named.
    Used for: Ferry (whole ship)
  31. [192]
    Resolution MEPC.281(70): amendments to the 2014 guidelines on the method of calculation of the attained EEDI for new ships (CO₂ emission factors per fuel)
    International Maritime Organization, Marine Environment Protection Committee, 2016. Checked 2026-10-16.
    Carbon factors in tonnes of CO₂ per tonne of fuel: diesel/gas oil 3.206, light fuel oil 3.151, heavy fuel oil 3.114, LPG 3.000 to 3.030. Mirror of the IMO document.
  32. [193]
    Carnival Corporation & plc FY2024 Sustainability Report
    Carnival Corporation & plc, 2025. Checked 2026-10-16.
    94 ships, 13.5 million passengers, 95.6 million available lower berth-days; scope 1 emissions 9,680 kt CO₂e; intensity 100 kg CO₂e per berth-day and 203 g CO₂e per berth-km in FY2024 (211 in FY2023, 241 in FY2022).
    Used for: Cruise ship (whole ship)
  33. [194]
    Model for estimation of fuel consumption of cruise ships
    Morten Simonsen, Hans Jakob Walnum, Stefan Gössling. Energies 11(5):1059, MDPI, 2018. Checked 2026-10-16.
    Measured and modelled fuel use of cruise ships in Norwegian waters: Artania 239.8 kg per nautical mile measured over a month (1,200 berths); Norwegian Star 329.1 kg per nautical mile modelled (3,000 berths); emission factor 3.114 kg CO₂ per kg of marine fuel oil.
    Used for: Cruise ship (whole ship)

Usage and scenarios

  1. [100]
    WildChat: 1M ChatGPT Interaction Logs in the Wild (Zhao, Ren, Hessel, Cardie, Choi, Deng)
    ICLR 2024 / arXiv, 2024. Checked 2026-10-02.
    Used for: scenario: Short question
  2. [101]
    LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset (Zheng et al.)
    arXiv / ICLR 2024, 2023. Checked 2026-10-02.
    Used for: scenario: Short question
  3. [102]
    How People Use ChatGPT (Chatterji, Cunningham, Deming, Hitzig, Ong, Shan, Wadman), NBER Working Paper 34255
    National Bureau of Economic Research, 2025. Checked 2026-10-02.
    Used for: scenario: Short question · scale: One week of ChatGPT (18 billion messages), each like this one
  4. [103]
  5. [104]
  6. [105]
    Claude Opus 5.5 system prompts (release notes)
    Anthropic, 2026. Checked 2026-10-02.
    Used for: scenario: Short question
  7. [106]
    Claude's System Prompt: Chatbots Are More Than Just Models
    Drew Breunig (blog), 2025. Checked 2026-10-02.
    Used for: scenario: Short question
  8. [107]
    Highlights from the Claude 4 system prompt
    Simon Willison's Weblog, 2025. Checked 2026-10-02.
  9. [108]
    ChatGPT users send 2.5B prompts a day
    TechCrunch (citing OpenAI statement to Axios), 2025. Checked 2026-10-02.
    Used for: scale: One day of ChatGPT (2.5 billion prompts), each like this one
  10. [109]
    ChatGPT reaches 900M weekly active users
    TechCrunch, 2026. Checked 2026-10-02.
  11. [110]
  12. [111]
  13. [112]
    Werkbare dagen (payroll note, 8-12-2025)
    Mysolution, 2025. Checked 2026-10-02.
    Used for: scale: Every working day for a year (230×) · scale: A 1,000-person company, every working day for a year
  14. [113]
  15. [114]
    Manage costs effectively (Claude Code docs)
    Anthropic, 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug · scenario: Max out Claude Code
  16. [115]
    Claude Code Pricing 2026: Real Costs
    Verdent (Rui Dai), 2026. Checked 2026-10-02.
  17. [116]
    What a task costs on Opus 5.5
    Anthropic (claude.dev blog, Addy Osmani), 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  18. [117]
    What is the Max plan?
    Anthropic Help Center, 2026. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  19. [118]
    Copy of 'About Claude's Max Plan Usage' help article (900 messages / 5h)
    GitHub gist (eonist), quoting Anthropic Help Center, 2025. Checked 2026-10-02.
  20. [119]
    Anthropic adds weekly limits to Claude, cites abuses
    Stark Insider (Clinton Stark), 2025. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  21. [120]
    Anthropic will set new weekly usage limits for Claude subscribers starting August
    The Decoder (Matthias Bastian), 2025. Checked 2026-10-02.
  22. [121]
    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
    arXiv (Yang et al., Princeton), 2024. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  23. [122]
    SWE-Effi: Re-Evaluating Software AI Agent System Effectiveness Under Resource Constraints
    arXiv (Fan et al.), 2025. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  24. [123]
  25. [124]
  26. [125]
    Agentic AI Workload Characteristics
    arXiv (Yuan et al.), 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  27. [126]
    SWE-bench leaderboards (embedded leaderboard-data JSON: instance_calls, instance_cost)
    SWE-bench (Princeton / Stanford), 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  28. [127]
    Claude Code /context command: token usage breakdown
    Wmedia, 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  29. [128]
    Anatomy of a Claude Code session
    codewithmukesh (Mukesh Murugan), 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  30. [129]
    My Real Claude Code Usage, Measured
    Chris Alarcon (blog), 2026. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  31. [130]
    Bug: Claude Code current session limit reaches 100% despite low visible local session usage (#54750)
    GitHub anthropics/claude-code (Troskiev83), 2026. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  32. [131]
    Comment on #84750 with ccusage blocks table (Max 5x, Opus 5)
    GitHub anthropics/claude-code (wgolden117), 2026. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  33. [132]
    [BUG] Abnormal token consumption still happening (#84750)
    GitHub anthropics/claude-code, 2026. Checked 2026-10-02.
    Used for: scenario: Build a feature · scenario: Hunt a bug
  34. [133]
    Weekly limit depletes disproportionately to 5h-session limit on Max 20x (#57699)
    GitHub anthropics/claude-code (Frisch12), 2026. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  35. [134]
  36. [135]
    Claude Code Usage Monitor (README plan limits)
    GitHub (Maciek-roboblog), 2026. Checked 2026-10-02.
    Used for: scenario: Max out Claude Code
  37. [136]
    ccusage and claude are not agree on my limits (#247)
    GitHub ccusage/ccusage, 2025. Checked 2026-10-02.
  38. [137]
    ccusage Blocks report guide
    ccusage, 2026. Checked 2026-10-02.
  39. [138]
    What it cost to build an AI energy calculator with AI (session logs of this tool's build)
    AI Codex, 2026. Checked 2026-10-02.
    Token counts from the Claude Code transcripts of the session that built this tool.
    Used for: scenario: Build this tool
  40. [139]
    Session log of the long chat example (course research with web search)
    AI Codex, 2026. Checked 2026-10-03.
    Token counts per model call from the Claude Code transcript of a real five-question research conversation on 3 October 2026.
    Used for: scenario: Long question

Maps and places

  1. [140]
    Natural Earth 1:110m land
    Natural Earth, 2024. Checked 2026-10-02.
    Public domain. World map outline.
    Used for: world map
  2. [141]
    GeoNames geographical database
    GeoNames, 2026. Checked 2026-10-02.
    City coordinates (city centre, rounded).
    Used for: place coordinates
  3. [142]
    Your own input
    You, 2026. Checked 2026-10-02.
    Values you entered yourself in the tool.

Household water use

  1. [143]
    Watergebruik Thuis (WGT) 2021
    Jeldrik Bakker, Francis van der Mooren, Harm Jan Boonstra. CBS (Statistics Netherlands), commissioned by Vewin, 2022. Checked 2026-10-02.
    Dutch household water survey, December 2020 to November 2021. Chapters 5 (bath), 6 (shower), 7 (toilet), 8 (dishwasher) and 9 (washing machine).
    Used for: Water for one shower · Water for one bath · Water for one dishwasher cycle · Water for one washing machine load · Water for one full toilet flush · Water for one small toilet flush
  2. [144]
    Showers (Home Water Works), citing the 2016 Residential End Uses of Water Study
    Alliance for Water Efficiency, 2016. Checked 2026-10-02.
    Used for: Water for one shower
  3. [145]
    WaterSense for Kids
    US EPA, 2025. Checked 2026-10-02.
    Used for: Water for one bath
  4. [146]
    Residential Toilets (WaterSense)
    US EPA, 2026. Checked 2026-10-09.
    Used for: Water for one full toilet flush
  5. [147]
    Dishwashers Key Product Criteria
    ENERGY STAR, 2023. Checked 2026-10-02.
    Used for: Water for one dishwasher cycle
  6. [148]
    Clothes Washers
    ENERGY STAR, 2025. Checked 2026-10-02.
    Used for: Water for one washing machine load

The bigger picture

  1. [149]
    Energy and AI – Executive summary
    International Energy Agency (IEA), 2025. Checked 2026-10-09.
    Used for: bigger picture: World energy-related CO₂, 2024 · bigger picture: energy → electricity → data centres → AI · bigger picture: headline figures
  2. [150]
  3. [151]
    Global Energy Review 2025 – CO2 emissions
    International Energy Agency (IEA), 2025. Checked 2026-10-09.
    Used for: bigger picture: World energy-related CO₂, 2024
  4. [152]
    Energy Efficiency 2023 (PDF)
    International Energy Agency (IEA), 2023. Checked 2026-10-09.
    Latest IEA edition that prints the world final-energy split by end-use sector in one sentence (2022 data).
    Used for: bigger picture: energy → electricity → data centres → AI
  5. [153]
    Energy Efficiency 2025 – Transport
    International Energy Agency (IEA), 2025. Checked 2026-10-09.
  6. [154]
    Cars and Vans (Energy System tracking page, 2024 edition)
    International Energy Agency (IEA), 2024. Checked 2026-10-09.
    The 2026 version of the page shows charts only. The quoted sentence is from the 2024 edition, repeated verbatim by Nordea (22 Feb 2024).
    Used for: bigger picture: headline figures
  7. [155]
    Cars, planes, trains: where do CO₂ emissions from transport come from?
    Hannah Ritchie. Our World in Data, 2020. Checked 2026-10-09.
    Shares from IEA data: transport share for 2016, split by mode for 2018.
    Used for: bigger picture: World energy-related CO₂, 2024
  8. [156]
    How much electricity does AI consume? [2025 summary]
    Hannah Ritchie. Sustainability by numbers (Substack), 2026. Checked 2026-10-09.
    Summarises the IEA Key Questions on Energy and AI figures (155 TWh for AI-focused data centres in 2025).
    Used for: bigger picture: energy → electricity → data centres → AI
  9. [157]
    Welke sectoren stoten broeikasgassen uit? (Dossier broeikasgassen)
    Statistics Netherlands (CBS), 2025. Checked 2026-10-09.
    Used for: bigger picture: Netherlands greenhouse gases, 2024
  10. [158]
    Datacenters verbruiken 4,6 procent van de elektriciteit
    Statistics Netherlands (CBS), 2025. Checked 2026-10-09.
    Used for: bigger picture: Netherlands greenhouse gases, 2024 · bigger picture: headline figures

Everyday electricity use

  1. [159]
    iPhone battery capacities compared: all iPhones battery life in mAh and Wh
    Simon Jary. Macworld, 2026. Checked 2026-10-16.
    Apple publishes no battery figures; the capacities come from the EU energy label (mAh) and regulatory filings (Wh).
    Used for: Electricity for one phone charge
  2. [160]
    Samsung Galaxy S25 - Full phone specifications
    GSMArena, 2025. Checked 2026-10-16.
    Used for: Electricity for one phone charge
  3. [161]
    Wireless Charging Drastically Less Efficient Than Charging With a Cord: REPORT (reporting the OneZero / iFixit measurements by Eric Ravenscraft)
    iPhone in Canada, 2020. Checked 2026-10-16.
    Secondary report of a test by OneZero (Eric Ravenscraft) with iFixit: a Google Pixel 4 charged from empty to full, energy metered at the wall.
    Used for: Electricity for one phone charge
  4. [162]
    MacBook Air (13-inch, M4, 2025) - Tech Specs
    Apple, 2025. Checked 2026-10-16.
    Used for: Electricity for one laptop charge
  5. [163]
    MacBook Pro (14-inch, M4, 2024) - Tech Specs
    Apple, 2024. Checked 2026-10-16.
    Used for: Electricity for one laptop charge
  6. [164]
    The carbon footprint of streaming video: fact-checking the headlines (IEA commentary, 11 December 2020)
    George Kamiya. International Energy Agency (IEA), 2020. Checked 2026-10-16.
    Used for: Electricity per minute of a video call
  7. [165]
    Carbon impact of video streaming (white paper with DIMPACT, 2021)
    Carbon Trust, 2021. Checked 2026-10-16.
    Used for: Electricity per minute of a video call
  8. [166]
    Assessing the Carbon Footprint of Virtual Meetings: A Quantitative Analysis of Camera Usage (arXiv:2601.06045)
    Félix Mortas. arXiv, 2026. Checked 2026-10-16.
    Nine meetings on one phone over 4G; the paper's electricity-per-GB range (0.12 to 0.26 kWh/GB) is for the mobile network only.
    Used for: Electricity per minute of a video call
  9. [167]
    Evaluation of in-service smartphone battery drainage profile for video calling feature in major apps (Scientific Reports 13, 11699)
    Hassan Ali, Hassan Abbas Khan, Michael Pecht. Scientific Reports, 2023. Checked 2026-10-16.
    Used for: Electricity per minute of a video call
  10. [168]
    Ecodesign: power consumption of PlayStation consoles
    Sony Interactive Entertainment, 2026. Checked 2026-10-23.
    Published under the EU games console self-regulatory initiative: the average of five European consoles per model, playing three PS5 games.
    Used for: Electricity for one hour of PlayStation gaming
  11. [169]
    El ajuste de tu televisor que puede multiplicar por dos el gasto de luz sin que te des cuentaes
    Matias Perea. Selectra, 2026. Checked 2026-10-23.
    On-mode power of current 55-inch TVs in SDR with the normal picture, from their EU energy labels (EPREL).
  12. [170]
    E-bike Impacts: Quantifying the Energy Use and Lifecycle Emissions in Response to Real World Driving Conditions (MASc thesis, University of Victoria)
    Daniel Clancy. University of Victoria, 2020. Checked 2026-10-16.
    Corporate fleet trial: 17 riders, 92 trips, 607 km, power measured at the rear hub with a PowerTap G3; Table 12 gives charger (0.95), charging (0.85) and motor (0.90) efficiencies.
    Used for: Electricity per km on an e-bike
  13. [171]
    What makes the eBike into a sustainable means of transport? (Bosch eBike Systems help center)
    Bosch eBike Systems, 2026. Checked 2026-10-16.
    The help center was temporarily unavailable on 2026-10-16; the quote is from the search index of the page.
    Used for: Electricity per km on an e-bike
  14. [172]
    Estimation of Energy Efficiency in Pedal-Assist Electric Bicycles (SSRN working paper 5525822)
    Daniel Jaramillo-Ramirez, David Florez-Rubio, Martha Cano-Morales, Max Herrera, Jesus Gonzalez-Llorente. SSRN / Pontificia Universidad Javeriana, 2025. Checked 2026-10-16.
    Used for: Electricity per km on an e-bike

AI training

  1. [173]
    Carbon Emissions and Large Neural Network Training (arXiv:2104.10350)
    David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, Jeff Dean. arXiv (Google / UC Berkeley), 2021. Checked 2026-10-16.
    Table 4 gives processor, chip count, days, energy and CO₂e for T5, Meena, GShard, Switch Transformer and GPT-3; OpenAI measured the V100 power and PUE for GPT-3.
    Used for: training: GPT-3 · scenario: Train a model
  2. [174]
    Scaling Language Models: Methods, Analysis & Insights from Training Gopher (arXiv:2112.11446)
    Jack W. Rae, Sebastian Borgeaud, Trevor Cai, et al.. arXiv (DeepMind), 2021. Checked 2026-10-16.
    Appendix F reports 4,096 TPU v3 chips, 920 hours, 283 W per chip, PUE 1.08 and 380 net tCO₂e.
    Used for: training: Gopher
  3. [175]
    OPT: Open Pre-trained Transformer Language Models (arXiv:2205.01068)
    Susan Zhang, Stephen Roller, Naman Goyal, et al.. arXiv (Meta AI), 2022. Checked 2026-10-16.
    "We trained OPT-175B on 992 80GB A100 GPUs" (Section 2.4); "an estimated carbon emissions footprint (CO2eq) of 75 tons" (Section 6).
    Used for: training: OPT-175B
  4. [176]
    Llama 2 70B model card: Carbon Footprint (Hugging Face)
    Meta, 2023. Checked 2026-10-16.
    GPU-hours, power per GPU (A100-80GB, 400 W) and tCO₂eq per model; 3,311,616 GPU-hours and 539 t for the family.
    Used for: training: Llama 2 70B
  5. [177]
    Llama 3.1 405B model card: Training Energy Use and Greenhouse Gas Emissions (Hugging Face)
    Meta, 2024. Checked 2026-10-16.
    39.3M GPU-hours on H100-80GB (700 W) for the family; 11,390 t location-based, 0 market-based. Methodology: arXiv:2204.05149.
    Used for: training: Llama 3.1 8B · training: Llama 3.1 70B · training: Llama 3.1 405B · scenario: Train a model
  6. [178]
    The Llama 3 Herd of Models (arXiv:2407.21783)
    Llama Team, AI @ Meta. arXiv (Meta AI), 2024. Checked 2026-10-16.
    "trained on up to 16K H100 GPUs, each running at 700W TDP with 80GB HBM3" on Meta's Grand Teton servers (Section 3.3.1); 15.6T tokens; a 54-day pre-training snapshot with 466 job interruptions (Section 3.3.4).
    Used for: scenario: Train a model
  7. [179]
    Llama 4 Maverick model card: Training Energy Use and Greenhouse Gas Emissions (Hugging Face)
    Meta, 2025. Checked 2026-10-16.
    Scout 5.0M and Maverick 2.38M GPU-hours on H100-80GB (700 W); 1,354 and 645 t location-based.
    Used for: training: Llama 4 Maverick
  8. [180]
    2 OLMo 2 Furious (arXiv:2501.00656)
    Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groeneveld, Kyle Lo, et al.. arXiv (Ai2), 2025. Checked 2026-10-16.
    Section 6.1 describes the Jupiter (Cirrascale, Austin) and Augusta (Google Cloud, Iowa) H100 clusters; Section 6.5 reports logged GPU power, CO₂e and water per model.
    Used for: training: OLMo 2 7B · training: OLMo 2 13B · scenario: Train a model
  9. [181]
    Holistically Evaluating the Environmental Impact of Creating Language Models (arXiv:2503.05804)
    Jacob Morrison, Clara Na, Jared Fernandez, Tim Dettmers, Emma Strubell, Jesse Dodge. arXiv (Ai2 / CMU), 2025. Checked 2026-10-16.
    Final runs of the OLMo series: 913 MWh, 312 t, 1,921 kL; development added about 50%; power draw swung between 15% and 85% of the hardware maximum.
    Used for: training: OLMo 2 7B · training: OLMo 2 13B · scenario: Train a model
  10. [182]
    DeepSeek-V3 Technical Report (arXiv:2412.19437)
    DeepSeek-AI, Aixin Liu, Bei Feng, et al.. arXiv (DeepSeek-AI), 2024. Checked 2026-10-16.
    Table 1: 2.788M H800 GPU-hours for the final run on a 2,048-GPU cluster; no energy or CO₂e figure.
    Used for: training: DeepSeek-V3 · scenario: Train a model
  11. [183]
    Position: Sustainable Open-Source AI Requires Tracking the Cumulative Footprint of Derivatives (arXiv:2601.21632)
    Shaina Raza, Iuliia Zarubiieva, Ahmed Y. Radwan, Nathaniel Lesperance, Deval Pandya, Sedef Akinli Kocak, Graham W. Taylor. arXiv, 2026. Checked 2026-10-16.
    Estimates DeepSeek-V3 training at about 1.07 GWh and 548 tCO₂eq assuming 350 W per H800, PUE 1.1 and 0.51 kg/kWh.
    Used for: training: DeepSeek-V3
  12. [184]
    The Environmental Impacts of Language Model Training Keep Rising: Now is the Time to Catch Impacts on the Rebound (arXiv:2510.09022)
    Clément Morand, Anne-Laure Ligozat, Aurélie Névéol, et al.. arXiv (LISN / STL), 2025. Checked 2026-10-16.
    Life-cycle model of GPT-4 training: 57M A100 GPU-hours, 27.4 GWh usage + 5.4 GWh infrastructure, 15 ktCO₂e including 3.3 kt embodied (Table 2).
    Used for: training: GPT-4 · scenario: Train a model
  13. [185]
    The carbon footprint of GPT-4 (Towards Data Science)
    Kasper Groes Albin Ludvigsen. Towards Data Science, 2023. Checked 2026-10-16.
    Rebuilds the run from leaked figures (25,000 A100s, 90 to 100 days) using the 6.5 kW rating of an 8-GPU server and Azure's PUE of 1.18.
    Used for: training: GPT-4
  14. [186]
    Grok 4 training resources: GPU-hours, electricity, water, CO₂ and cost (Epoch AI data insight)
    James Sanders, Luke Emberson, Yafah Edelman. Epoch AI, 2025. Checked 2026-10-16.
    246M H100-hours inferred from xAI statements; 310 GWh (700 W, 2.03× non-GPU hardware, PUE 1.2, 75% average draw), about 750 million litres of water and 140,000 to 154,000 t CO₂ at gas-turbine intensity.
    Used for: training: Grok 4 · scenario: Train a model
  15. [187]
    Our contribution to a global environmental standard for AI (Mistral Large 2 life-cycle assessment)
    Mistral AI, 2025. Checked 2026-10-16.
    20.4 ktCO₂e and 281,000 m³ of water for training plus 18 months of use (to January 2025); 1.14 gCO₂e and 45 mL per 400-token response. No energy figure.
  16. [188]
    What kind of environmental impacts are AI companies disclosing? (And can we compare them?)
    Sasha Luccioni, Theo Alves Da Costa. Hugging Face blog, 2025. Checked 2026-10-16.
    Compares the OpenAI, Mistral and Google disclosures of 2025: none publishes training energy; methods and query definitions differ, so per-query figures are not comparable.
    Used for: scenario: Train a model

AI image generation

  1. [195]
    AI Energy Score Leaderboard, February 2025 (image generation task)
    Sasha Luccioni, Boris Gamazaychikov, Emma Strubell, Yacine Jernite. Hugging Face, Salesforce, Cohere and Carnegie Mellon University, 2025. Checked 2026-10-16.
    GPU energy per 1,000 image prompts on one H100 80GB at batch size 1, FP32, 10 runs (raw file data/energy/image_generation.csv): SDXL base 1.0 1,639.85 Wh, SD 2.1 534.47 Wh, SDXL Turbo 385.57 Wh. The output size and step count are the models' defaults and are not documented; the FLUX.1 schnell and SD3 medium runs errored out.
    Used for: image: Stable Diffusion XL · image: Stable Diffusion 2.1
  2. [196]
    Energy Scaling Laws for Diffusion Models: Quantifying Compute in Image Generation
    Aniketh Iyengar, Jiaqi Han, Boris Ruf, Vincent Grari, Marcin Detyniecki, Stefano Ermon. arXiv (ACM FAccT 2026), 2025. Checked 2026-10-16.
    Dynamic GPU energy (idle subtracted) of FLUX.1 dev, SD 3.5 Large, SD 2 and Qwen-Image on an A100 at batch size 1 across output sizes, step counts and precisions; the appendix tables (6 to 9) give joules per 100 prompts.
    Used for: image: Stable Diffusion 3.5 Large · image: FLUX.1 [dev] · image: Stable Diffusion 2.1
  3. [197]
    Towards Sustainable Creativity Support: An Exploratory Study on Prompt Based Image Generation
    Daniel Hove Paludan, Julie Fredsgård, Kasper Patrick Bährentz, Ilhan Aslan, Niels van Berkel. arXiv, 2025. Checked 2026-10-16.
    A 24-person user study with a Stable Diffusion 1.5 fine-tune (Juggernaut Reborn) at 512×512 and 30 steps: about 0.3 Wh per finished image, metered with CodeCarbon on a local workstation whose GPU is not named.
    Used for: image: Stable Diffusion 1.5
  4. [198]
    We did the math on AI's energy footprint. Here's the story you haven't heard.
    James O'Donnell, Casey Crownhart. MIT Technology Review, 2025. Checked 2026-10-16.
    Stable Diffusion 3 Medium at 1024×1024: about 1,141 joules of GPU energy per image, measured for the article by the ML.ENERGY team; the article doubles it for the rest of the server and quotes 4,402 joules at 50 steps.
    Used for: image: Stable Diffusion 3.5 Large
  5. [199]
    Measure What Matters: Energy Economics of AI Image Generation
    Scope3, 2025. Checked 2026-10-16.
    Estimates in gCO₂e per image from a regression fitted to measured open models and the price charged per image: GPT-image-1 (4o) low 0.40, medium 1.50, high 5.63; DALL·E 3 standard 1.40, HD 2.74; SDXL 1.0 0.55; SD 3.5 Medium 1.37. No energy figures and no grid factor, so only the ratios are usable.
    Used for: image: GPT-image-1 · image: DALL·E 3
  6. [200]
    Stability AI Developer Platform: pricing
    Stability AI, 2026. Checked 2026-10-16.
    1 credit = $0.01. Stable Diffusion 3.5 Large 6.5 credits, Large Turbo 4, Medium 3.5; SDXL 1.0 from 0.9 credits per image.
    Used for: price of Stable Diffusion XL · price of Stable Diffusion 3.5 Large
  7. [201]
    Replicate pricing
    Replicate, 2026. Checked 2026-10-16.
    Per-image prices of hosted models: FLUX.1 dev $0.025, FLUX.1 schnell $3.00 per thousand, FLUX 1.1 pro $0.04, Ideogram v3 quality $0.09; hardware time A100 80GB $0.0014/s, L40S $0.000975/s, H100 $0.001525/s.
    Used for: price of FLUX.1 [dev]
  8. [202]
    Replicate: stability-ai/stable-diffusion (SD 1.5 and 2.1) and stability-ai/sdxl
    Replicate, 2026. Checked 2026-10-16.
    Stable Diffusion 1.5/2.1 costs about $0.0062 per run on an A100 80GB (about 5 seconds, 161 runs per dollar); SDXL about $0.0029 per run on an L40S (about 3 seconds). Hardware-time prices, not fixed list prices.
    Used for: price of Stable Diffusion 2.1 · price of Stable Diffusion 1.5

Industry

  1. [203]
    Pathways towards lower emissions: a global assessment of the greenhouse gas emissions and mitigation options from livestock agrifood systems
    FAO, Rome, 2023. Checked 2026-10-16.
    GLEAM 3, reference year 2015: livestock agrifood systems 6.2 Gt CO₂eq, about 12% of all human-caused emissions.
    Used for: Energy of the global livestock industry in a year · CO₂e of the global livestock industry in a year
  2. [204]
    Tackling climate change through livestock: a global assessment of emissions and mitigation opportunities
    Pierre Gerber, Henning Steinfeld, Benjamin Henderson, Anne Mottet, Carolyn Opio, Johannes Dijkman, Alessandra Falcucci, Giuseppe Tempio. FAO, Rome, 2013. Checked 2026-10-16.
    GLEAM 1, reference year 2005: 7.1 Gt CO₂-eq, 14.5% of human-caused emissions; fossil fuel along the supply chains about 20% of the sector's emissions.
    Used for: Energy of the global livestock industry in a year · CO₂e of the global livestock industry in a year
  3. [205]
    2006 IPCC Guidelines for National Greenhouse Gas Inventories, Volume 2 (Energy), chapter 1, Table 1.4: default CO₂ emission factors for combustion
    IPCC, 2006. Checked 2026-10-16.
    Natural gas 56,100, gas/diesel oil 74,100, motor gasoline 69,300, other bituminous coal 94,600 kg CO₂ per TJ; the 70 kg per GJ midpoint turns FAO's fossil-fuel CO₂ back into energy.
  4. [206]
    Greenhouse gas emissions in the Netherlands 1990–2023: National Inventory Report 2025 (RIVM report 2025-0005)
    RIVM / Emissieregistratie, 2025. Checked 2026-10-16.
    Table 5.1, agriculture in Tg CO₂ eq for 2023: enteric fermentation 9.2, manure management 4.4 (dairy cattle 6.1, other cattle 2.1, pigs 0.5).
    Used for: CO₂e of Dutch livestock farming in a year
  5. [207]
    Answer to parliamentary questions (ah-1266245) quoting the PBL Climate and Energy Outlook 2025: agriculture 24.8 Mt, process emissions of livestock and arable farming 17.8 Mt (2024)
    Tweede Kamer / Ministry of Agriculture, Fisheries, Food Security and Nature, 2025. Checked 2026-10-16.
    A secondary carrier of the PBL figure; the KEV 2025 itself was not fetched.
    Used for: CO₂e of Dutch livestock farming in a year
  6. [208]
    Emissies van de Nederlandse landbouw naar productgroepen (erratum), chapter 2: emissies naar lucht door de landbouw
    Olaf Koops, Patrick Bogaart, Kathleen Geertjes. Statistics Netherlands (CBS), 2024. Checked 2026-10-16.
    Table 2.1, 2021, Mt CO₂-eq: dairy 11.8, beef and veal 2.5, pork 2.3, poultry meat 0.06, eggs 0.07, other livestock 0.6; greenhouses 9.2; agriculture 27.9.
    Used for: CO₂e of Dutch livestock farming in a year
  7. [209]
    Emissies van broeikasgassen berekend volgens IPCC-voorschriften (StatLine 85669NED)
    Statistics Netherlands (CBS) / RIVM Emissieregistratie, 2026. Checked 2026-10-16.
    Updated 2026-09-03, 1990 to 2025 (2025 provisional), in billion kg CO₂-eq. Agriculture, stationary sources: 25.0 in 2023 and 2024, 24.6 in 2025.
    Used for: CO₂e of Dutch livestock farming in a year
  8. [210]
    Energiebalans; aanbod en verbruik, sector (StatLine 83989NED)
    Statistics Netherlands (CBS), 2026. Checked 2026-10-16.
    Updated 2026-06-12, 1990 to 2025. Agriculture 136.5 PJ in 2024, of which greenhouse horticulture 99.7 PJ of gas; slaughterhouses and meat processing (SBI 101) 5.1 PJ.
    Used for: Energy of Dutch livestock farming in a year
  9. [211]
    Energieverbruik land- en tuinbouw 1990–2013 (StatLine 80382ned)
    Statistics Netherlands (CBS), 2014. Checked 2026-10-16.
    Discontinued after 2013. Energy by farm type in PJ: greenhouses 109.2, grazing livestock 8.3, housed livestock 6.2, contractors 4.5.
    Used for: Energy of Dutch livestock farming in a year
  10. [212]
    Measuring Fashion: environmental impact of the global apparel and footwear industries study (full report and methodological considerations)
    Quantis / ClimateWorks Foundation, 2018. Checked 2026-10-16.
    Baseline year 2016: 3,990 Mt CO₂eq (apparel 3,290, footwear 700) and 48,640 PJ of non-renewable primary energy, use phase excluded. The live URL now redirects; the PDF was read from the Internet Archive.
    Used for: Energy of the global textile industry in a year · CO₂e of the global textile industry in a year
  11. [213]
    Fashion on Climate: how the fashion industry can urgently act to reduce its greenhouse gas emissions
    McKinsey & Company and Global Fashion Agenda, 2020. Checked 2026-10-16.
    About 2.1 billion tonnes of greenhouse gases in 2018, 4% of the global total, over the whole value chain.
    Used for: CO₂e of the global textile industry in a year
  12. [214]
    A new textiles economy: redesigning fashion's future (summary of findings)
    Ellen MacArthur Foundation, 2017. Checked 2026-10-16.
    Textiles production 1.2 billion tonnes CO₂e in 2015, based on McKinsey analysis; no energy total.
    Used for: CO₂e of the global textile industry in a year
  13. [215]
    Threads of Transformation: fashion, energy and women at the intersection of climate action
    Anna Strasser, Rachel Idem, Teniola Dorcas-Joseph. Sustainable Energy for All (SEforALL), 2025. Checked 2026-10-16.
    Repeats the '1 trillion kWh of electricity a year' claim (footnote 112 points to a 2019 blog that no longer carries it) and the 2 to 8% emissions range (WRI 2021, UNEP 2023). Not a primary data source.
    Used for: Energy of the global textile industry in a year
  14. [216]
    Bitcoin Energy Consumption Index: daily statistics API
    Digiconomist (Alex de Vries), 2026. Checked 2026-10-16.
    Daily electricity (24hr_kWh) and carbon (24hr_kgCO2) from 10 February 2017 via /wp-json/mo/v1/bitcoin/stats/{YYYYMMDD}; summed to 1,116 TWh and 601 Mt CO₂ up to 15 October 2026. The index page showed 204 TWh and 114 Mt annualised on the day.
    Used for: Electricity Bitcoin has used since 2009, all years combined · CO₂e of Bitcoin since 2009, all years combined
  15. [217]
    A deep dive into Bitcoin's environmental impact
    Alexander Neumueller. Cambridge Judge Business School, Cambridge Centre for Alternative Finance, 2022. Checked 2026-10-16.
    Cumulative 199.65 Mt CO₂e since 2009 up to 21 September 2022, about 92% of it after 2018; annualised 56.3 Mt (2021) and 48.4 Mt (2022). The index's data endpoints now sit behind a captcha.
    Used for: CO₂e of Bitcoin since 2009, all years combined

Food

  1. [218]
    Greenhouse gas emissions per kilogram of food product
    Joseph Poore and Thomas Nemecek (Science, 2018), processed by Our World in Data. Our World in Data, 2018. Checked 2026-10-23.
    Life-cycle emissions from farm to retail (land use, farm, feed, processing, transport, retail, packaging and losses) per kg of food product, mean of about 38,700 farms in 119 countries, reference year 2010. Downloaded as CSV.
    Used for: CO₂e of one kg of beef · CO₂e of one kg of chicken

Power plants: CO₂ and water

  1. [220]
    Annex III: Technology-specific cost and performance parameters (Table A.III.2, emissions of selected electricity supply technologies)
    Schlömer, S., Bruckner, T., Fulton, L., Hertwich, E., McKinnon, A., Perczyk, D., Roy, J., Schaeffer, R., Sims, R., Smith, P. and Wiser, R.. IPCC, Climate Change 2014: Mitigation of Climate Change (AR5 WG III), 2014. Checked 2026-10-25.
    Used for: CO₂ per kWh: Coal · CO₂ per kWh: Gas (combined cycle) · CO₂ per kWh: Gas turbines (simple cycle) · CO₂ per kWh: Gas engines · CO₂ per kWh: Nuclear · CO₂ per kWh: Geothermal · and 7 more
  2. [221]
    Operational water consumption and withdrawal factors for electricity generating technologies: a review of existing literature
    Jordan Macknick, Robin Newmark, Garvin Heath and K C Hallett. Environmental Research Letters 7, 045802 (NREL), 2012. Checked 2026-10-25.
    Used for: Water per kWh: Coal · Water per kWh: Gas (combined cycle) · Water per kWh: Gas turbines (simple cycle) · Water per kWh: Gas engines · Water per kWh: Oil · Water per kWh: Diesel generators · and 9 more
  3. [222]
    The consumptive water footprint of electricity and heat: a global assessment
    Mesfin M. Mekonnen, P. W. Gerbens-Leenes and Arjen Y. Hoekstra. Environmental Science: Water Research & Technology 1, 285–297, 2015. Checked 2026-10-25.
    Used for: Water per kWh: Coal · Water per kWh: Gas (combined cycle) · Water per kWh: Oil · Water per kWh: Nuclear · Water per kWh: Geothermal · Water per kWh: Hydro (reservoir) · and 2 more
  4. [223]
    The blue water footprint of electricity from hydropower
    Mesfin M. Mekonnen and Arjen Y. Hoekstra. Hydrology and Earth System Sciences 16, 179–187, 2012. Checked 2026-10-25.
    Used for: Water per kWh: Hydro (reservoir)
  5. [224]
    Electric Power Annual, Table 8.2: Average tested heat rates by prime mover and energy source
    US Energy Information Administration, 2026. Checked 2026-10-25.
    Used for: CO₂ per kWh: Gas turbines (simple cycle) · CO₂ per kWh: Gas engines · CO₂ per kWh: Oil · CO₂ per kWh: Diesel generators
  6. [225]
    Carbon dioxide emissions coefficients by fuel
    US Energy Information Administration, 2024. Checked 2026-10-25.
  7. [226]
    Bloom Energy Server datasheet (and water claim on the Honda customer page)
    Bloom Energy, 2026. Checked 2026-10-25.
    Vendor figures. The water claim is from https://www.bloomenergy.com/customer/honda/.
    Used for: CO₂ per kWh: Fuel cells (on gas) · Water per kWh: Fuel cells (on gas)

Data checked on 2 October 2026. Prices and models change fast; every value links to its source.