{ "licence": "CC BY 4.0", "licenceUrl": "https://creativecommons.org/licenses/by/4.0/", "notice": "The CO2Stats compilation of this ledger and our own writing in it are licensed CC BY 4.0: share them with a credit to CO2Stats and a link to https://co2stats.com/kettle/ledger.json. The figures and quotations come from the sources named in each claim and keep their original attribution and terms.", "claims": [ { "id": "altman-chatgpt-query-2025", "activity": "ai-text-prompt", "title": "Average ChatGPT query, per OpenAI's chief executive", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.34, "high": 0.34, "per": "ChatGPT query", "statistic": "mean" }, { "quantity": "water", "unit": "mL", "low": 0.32, "high": 0.32, "per": "ChatGPT query", "statistic": "mean", "original": "about 0.000085 gallons of water (converted at 1 US gallon = 3,785.41 mL)", "water": "unknown" } ], "counted": { "workload": "The average ChatGPT query", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "The Gentle Singularity", "publisher": "Sam Altman (blog.samaltman.com)", "kind": "blog post", "url": "https://blog.samaltman.com/the-gentle-singularity", "published": "2025-06-10", "quote": "People are often curious about how much energy a ChatGPT query uses; the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon." }, "method": "not stated", "claimant": { "type": "company", "name": "OpenAI (Sam Altman, chief executive)", "funder": "none" }, "status": "current", "dispute": "Google's August 2025 paper notes that the disclosure gives no explanation of the measurement boundary or method, so it cannot be compared with other estimates. https://arxiv.org/abs/2508.15734", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "berners-lee-email-attachment-2010", "activity": "email", "title": "An email with a long attachment, How Bad Are Bananas? (2010)", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 50, "high": 50, "per": "email with attachment", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "One email with a long attachment", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "What's the carbon footprint of ... email?", "publisher": "The Guardian (Mike Berners-Lee and Duncan Clark; draws on How Bad Are Bananas?)", "kind": "press article", "url": "https://www.theguardian.com/environment/green-living-blog/2010/oct/21/carbon-footprint-email", "published": "2010-10-21", "quote": "50g CO2e: An email with long and tiresome attachment" }, "method": "not stated", "claimant": { "type": "independent", "name": "Mike Berners-Lee", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "berners-lee-long-email-laptop-2024", "activity": "email", "title": "A long laptop-to-laptop email (10 minutes to write, 3 to read)", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 17, "high": 17, "per": "email", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "One email that takes 10 minutes to write and three minutes to read, sent from a laptop and received on one", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Concerned about your data use? Here is the carbon footprint of an average day of emails, WhatsApps and more", "publisher": "The Guardian (Chris Stokel-Walker)", "kind": "press article", "url": "https://www.theguardian.com/environment/2024/oct/31/concerned-about-your-data-use-here-is-the-carbon-footprint-of-an-average-day-of-emails-whatsapps-and-more", "published": "2024-10-31", "quote": "If the editor took 10 minutes to write the email ... and it took me three minutes to read, and if it was sent from a laptop and received on one, then we have generated 17g of carbon dioxide (CO2) emissions already, according to estimates by Mike Berners-Lee, a professor at Lancaster University’s Environment Centre" }, "method": "not stated", "claimant": { "type": "independent", "name": "Mike Berners-Lee (as reported by The Guardian)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "berners-lee-proper-email-2010", "activity": "email", "title": "A proper (non-spam) email, How Bad Are Bananas? (2010)", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 4, "high": 4, "per": "email", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "One proper (genuine, non-spam) email", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "What's the carbon footprint of ... email?", "publisher": "The Guardian (Mike Berners-Lee and Duncan Clark; draws on How Bad Are Bananas?)", "kind": "press article", "url": "https://www.theguardian.com/environment/green-living-blog/2010/oct/21/carbon-footprint-email", "published": "2010-10-21", "quote": "4g CO2e: A proper email" }, "method": "not stated", "claimant": { "type": "independent", "name": "Mike Berners-Lee", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "berners-lee-spam-email-2010", "activity": "email", "title": "A spam email, How Bad Are Bananas? (2010)", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 0.3, "high": 0.3, "per": "spam email", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "One spam email", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "What's the carbon footprint of ... email?", "publisher": "The Guardian (Mike Berners-Lee and Duncan Clark; draws on How Bad Are Bananas?)", "kind": "press article", "url": "https://www.theguardian.com/environment/green-living-blog/2010/oct/21/carbon-footprint-email", "published": "2010-10-21", "quote": "The carbon footprint of spam: 0.3g CO2e: A spam email" }, "method": "not stated", "claimant": { "type": "independent", "name": "Mike Berners-Lee", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "buckley-google-search-2009", "activity": "search", "title": "Google search including the user's PC, John Buckley's estimate in The Sunday Times", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 1, "high": 10, "per": "Google search", "statistic": "range", "accounting": "not stated" } ], "counted": { "workload": "A Google search, with or without starting the PC", "scope": "end to end", "included": [ "the user's PC (the range depends on whether the PC has to be started)" ], "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Revealed: the environmental impact of Google searches", "publisher": "The Sunday Times (Times Online; archived copy of 15 January 2009 on the Wayback Machine)", "kind": "press article", "url": "https://web.archive.org/web/20090115112144/http://technology.timesonline.co.uk/tol/news/tech_and_web/article5489134.ece", "published": "2009-01-11", "quote": "A separate estimate from John Buckley, managing director of carbonfootprint.com , a British environmental consultancy, puts the CO2 emissions of a Google search at between 1g and 10g, depending on whether you have to start your PC or not. Simply running a PC generates between 40g and 80g per hour, he says." }, "method": "not stated", "claimant": { "type": "independent", "name": "John Buckley (carbonfootprint.com), quoted by The Sunday Times", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "carbon-trust-video-streaming-europe-2020", "activity": "streaming-hour", "title": "One hour of video-on-demand streaming, European average in 2020", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 55, "high": 55, "per": "hour of video streaming", "statistic": "mean", "accounting": "location-based" }, { "quantity": "energy", "unit": "Wh", "low": 188, "high": 188, "per": "hour of video streaming", "statistic": "mean" } ], "counted": { "workload": "One hour of video-on-demand streaming in Europe in 2020, with a representative mix of viewing devices and a modelled average bitrate of 6.40 Mbps", "scope": "end to end", "included": [ "data centres and content delivery networks (encoding and storage)", "internet network transmission", "home routers", "end-user viewing devices", "TV peripherals such as set-top boxes" ], "excluded": [ "content creation", "embodied emissions of the equipment", "live streaming" ], "model": "Conventional allocation approach (average energy per data volume, kWh/GB)", "place": "Europe", "period": "2020" }, "source": { "title": "Carbon impact of video streaming", "publisher": "The Carbon Trust", "kind": "white paper", "url": "https://www.carbontrust.com/en-eu/node/1540", "published": "2021-06", "quote": "The European average footprint estimated in this white paper is approximately 55gCO2e per hour of video streaming for the conventional allocation approach. ... The equivalent energy consumption is a European average of 188Wh per hour of video streaming using the conventional approach" }, "method": "Modelled with the DIMPACT tool from operational electricity use of each component, using a European average grid emission factor of 0.295 gCO2e/Wh (national grid averages, so direct renewable purchases are not recognised); the body of the paper (section 4.1) gives the same estimate as 56 gCO2e per hour.", "claimant": { "type": "industry-funded", "name": "The Carbon Trust, in consultation with DIMPACT members", "funder": "Netflix" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "delavande-animatediff-video-2025", "activity": "ai-video", "title": "AnimateDiff video at default settings, Video Killed the Energy Budget", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.14, "high": 0.14, "per": "video (16 frames, 512 × 512, 10 fps, 4 denoising steps)", "statistic": "mean" } ], "counted": { "workload": "One video from AnimateDiff at its default Hugging Face settings (16 frames at 10 fps, 512 × 512, 4 steps), averaged over 50 prompts with five runs each", "scope": "servers", "included": [ "GPU energy", "CPU energy", "RAM energy (estimated)" ], "excluded": "not stated", "model": "AnimateDiff", "place": "not stated", "period": "not stated" }, "source": { "title": "Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models", "publisher": "arXiv (2509.19222)", "kind": "preprint", "url": "https://arxiv.org/abs/2509.19222", "published": "2025-09-23", "quote": "We observe orders-of-magnitude disparities: AnimateDiff requires only 0.14 Wh in total, while WAN2.1-T2V-14B consumes over 415 Wh, a factor of nearly 3000×." }, "method": "Run on one dedicated NVIDIA H100 SXM GPU with an 8-core AMD EPYC 7R13 CPU; GPU and CPU energy measured with CodeCarbon, RAM energy estimated with CodeCarbon's default heuristic; two warm-ups and five runs per prompt.", "claimant": { "type": "company", "name": "Julien Delavande, Régis Pierrard and Sasha Luccioni (Hugging Face; ENS Paris-Saclay)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "delavande-wan21-14b-video-2025", "activity": "ai-video", "title": "WAN2.1-T2V-14B video at default settings, Video Killed the Energy Budget", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 415, "high": 415, "per": "video (81 frames, 720 × 1280, 15 fps, 50 denoising steps)", "statistic": "lower bound", "original": "over 415 Wh" } ], "counted": { "workload": "One video from WAN2.1-T2V-14B at its default Hugging Face settings (81 frames at 15 fps, 720 × 1280, 50 steps), averaged over 50 prompts with five runs each", "scope": "servers", "included": [ "GPU energy", "CPU energy", "RAM energy (estimated)" ], "excluded": "not stated", "model": "WAN2.1-T2V-14B", "place": "not stated", "period": "not stated" }, "source": { "title": "Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models", "publisher": "arXiv (2509.19222)", "kind": "preprint", "url": "https://arxiv.org/abs/2509.19222", "published": "2025-09-23", "quote": "We observe orders-of-magnitude disparities: AnimateDiff requires only 0.14 Wh in total, while WAN2.1-T2V-14B consumes over 415 Wh, a factor of nearly 3000×." }, "method": "Run on one dedicated NVIDIA H100 SXM GPU with an 8-core AMD EPYC 7R13 CPU; GPU and CPU energy measured with CodeCarbon, RAM energy estimated with CodeCarbon's default heuristic; two warm-ups and five runs per prompt.", "claimant": { "type": "company", "name": "Julien Delavande, Régis Pierrard and Sasha Luccioni (Hugging Face; ENS Paris-Saclay)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "devries-chatgpt-request-2023", "activity": "ai-text-prompt", "title": "ChatGPT request in early 2023, from SemiAnalysis via de Vries", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 2.9, "high": 2.9, "per": "ChatGPT request", "statistic": "upper bound" } ], "counted": { "workload": "An average ChatGPT request in early 2023 (564 MWh per day over 195 million requests per day)", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "ChatGPT (early 2023)", "place": "not stated", "period": "early 2023" }, "source": { "title": "The growing energy footprint of artificial intelligence", "publisher": "Joule (Cell Press)", "kind": "journal commentary", "url": "https://doi.org/10.1016/j.joule.2023.09.004", "published": "2023-10-18", "quote": "This figure aligns with SemiAnalysis’ assessment of ChatGPT’s operating costs in early 2023, which estimated that ChatGPT responds to 195 million requests per day, requiring an estimated average electricity consumption of 564 MWh per day, or, at most, 2.9 Wh per request." }, "method": "Taken from SemiAnalysis' early-2023 estimate of ChatGPT's daily electricity use (3,617 Nvidia HGX A100 servers, 564 MWh per day) divided by 195 million requests per day; the paper states it as an upper bound (\"at most\").", "claimant": { "type": "independent", "name": "Alex de Vries (VU Amsterdam, Digiconomist)", "funder": "not stated" }, "status": "current", "dispute": "Epoch AI (February 2025) says this estimate, the origin of the widely cited 3 Wh figure, rests on outdated hardware and an unrepresentative assumption of about 4,000 input and 2,000 output tokens per query, and estimates roughly 0.3 Wh for a typical GPT-4o query. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "epoch-chatgpt-gpt4o-query-2025", "activity": "ai-text-prompt", "title": "Typical ChatGPT query on GPT-4o, Epoch AI estimate", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.3, "high": 0.3, "per": "ChatGPT query", "statistic": "single estimate" } ], "counted": { "workload": "A typical ChatGPT query with about 500 output tokens", "scope": "data centre", "included": [ "H100 GPU power", "server and data centre overhead (about 1,500 W per GPU at peak, at 70% average power)" ], "excluded": [ "model training and other upstream costs (discussed separately, not in the figure)", "extra cost of long inputs (discussed separately)" ], "model": "GPT-4o (assumed about 100 billion active parameters)", "place": "not stated", "period": "not stated" }, "source": { "title": "How much energy does ChatGPT use?", "publisher": "Epoch AI (Gradient Updates, by Josh You)", "kind": "blog post", "url": "https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use", "published": "2025-02-07", "quote": "We find that typical ChatGPT queries using GPT-4o likely consume roughly 0.3 watt-hours, which is ten times less than the older estimate." }, "method": "Model-based estimate from assumed FLOP per token, 500 output tokens, Nvidia H100 throughput at 10% utilisation, and about 1,500 W per GPU including server and data centre overhead at 70% average power.", "claimant": { "type": "independent", "name": "Epoch AI (Josh You)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "epri-chatgpt-request-2024", "activity": "ai-text-prompt", "title": "ChatGPT request, EPRI Powering Intelligence", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 2.9, "high": 2.9, "per": "ChatGPT request", "statistic": "single estimate" } ], "counted": { "workload": "One ChatGPT request", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "ChatGPT (model not stated)", "place": "not stated", "period": "not stated" }, "source": { "title": "Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption", "publisher": "Electric Power Research Institute (EPRI)", "kind": "white paper", "url": "https://www.epri.com/research/products/3002028905", "published": "2024-05", "quote": "At 2.9 watt-hours per ChatGPT request, AI queries are estimated to require 10x the electricity of traditional Google queries, which use about 0.3 watt-hours each" }, "method": "not stated", "claimant": { "type": "industry-funded", "name": "Electric Power Research Institute (EPRI)", "funder": "EPRI's members, chiefly electric utilities (no funder named for this report)" }, "status": "current", "dispute": "Epoch AI (February 2025) argues that the widely cited figure of about 3 Wh per ChatGPT query, which traces to de Vries (2023), is likely an overestimate, and puts a typical GPT-4o query at roughly 0.3 Wh. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "goodall-google-search-2009", "activity": "search", "title": "Google search with 15 minutes of computer use, Chris Goodall's estimate in The Sunday Times", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 7, "high": 10, "per": "Google search (assuming 15 minutes’ computer use)", "statistic": "range", "accounting": "not stated" } ], "counted": { "workload": "A Google search, assuming 15 minutes’ computer use", "scope": "end to end", "included": [ "15 minutes’ computer use" ], "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Revealed: the environmental impact of Google searches", "publisher": "The Sunday Times (Times Online; archived copy of 15 January 2009 on the Wayback Machine)", "kind": "press article", "url": "https://web.archive.org/web/20090115112144/http://technology.timesonline.co.uk/tol/news/tech_and_web/article5489134.ece", "published": "2009-01-11", "quote": "Chris Goodall, author of Ten Technologies to Save the Planet, estimates the carbon emissions of a Google search at 7g to 10g (assuming 15 minutes’ computer use)." }, "method": "not stated (the article says \"carbon emissions\"; recorded as gCO2 because the article reports the other estimates as CO2)", "claimant": { "type": "independent", "name": "Chris Goodall, quoted by The Sunday Times", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "google-gemini-prompt-2025-comprehensive", "activity": "ai-text-prompt", "title": "Median Gemini Apps text prompt, comprehensive boundary", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.24, "high": 0.24, "per": "text prompt", "statistic": "median" }, { "quantity": "co2", "unit": "gCO2e", "low": 0.03, "high": 0.03, "per": "text prompt", "statistic": "median", "accounting": "market-based" }, { "quantity": "water", "unit": "mL", "low": 0.26, "high": 0.26, "per": "text prompt", "statistic": "median", "water": "consumption" } ], "counted": { "workload": "The median Gemini Apps text prompt in May 2025", "scope": "data centre", "included": [ "active AI accelerators (prefill and decode, including networking between accelerators in the same AI computer)", "host CPU and DRAM", "idle machines provisioned for reliability and low latency", "data centre overhead (PUE)", "embodied (Scope 1 and Scope 3) emissions of the AI accelerators and host CPU and DRAM (CO2 figure)", "water consumed for cooling in data centres (water figure)" ], "excluded": [ "external networking", "end-user devices, including edge compute", "LLM training and data storage" ], "model": "Gemini Apps (all models serving the app, including supporting models for scoring, ranking, classification and routing)", "place": "Google data centre fleet (fleet-wide average emission and water factors)", "period": "May 2025" }, "source": { "title": "Measuring the environmental impact of delivering AI at Google Scale", "publisher": "Google (arXiv)", "kind": "preprint", "url": "https://arxiv.org/abs/2508.15734", "published": "2025-08-21", "quote": "Applying our comprehensive measurement methodology, we find that the median energy consumption for a Gemini Apps text prompt in May 2025 is 0.24 Wh. ... a median Gemini Apps text prompt generates 0.03 gCO2e and consumes 0.26 mL of water when measured comprehensively." }, "method": "Measured from internal telemetry across Google's AI serving fleet; emissions apply the previous calendar year's fleet market-based grid emission factor plus embodied Scope 1 and 3 emissions, and water applies the fleet consumptive water usage effectiveness (ISO WUE Category 2) to energy per prompt.", "claimant": { "type": "company", "name": "Google", "funder": "none" }, "status": "current", "dispute": "Researchers quoted in press coverage, including Shaolei Ren (UC Riverside) and Alex de Vries-Gao, said the water figure leaves out water consumed by the power plants that supply the data centres, and that market-based emissions do not show the carbon actually emitted. https://www.pcworld.com/article/2886730/experts-are-skeptical-about-googles-ai-water-consumption-claims.html", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "google-gemini-prompt-2025-narrow", "activity": "ai-text-prompt", "title": "Median Gemini Apps text prompt, active accelerators only", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.1, "high": 0.1, "per": "text prompt", "statistic": "median" }, { "quantity": "co2", "unit": "gCO2e", "low": 0.02, "high": 0.02, "per": "text prompt", "statistic": "median", "accounting": "market-based" }, { "quantity": "water", "unit": "mL", "low": 0.12, "high": 0.12, "per": "text prompt", "statistic": "median", "water": "consumption" } ], "counted": { "workload": "The median Gemini Apps text prompt in May 2025, measured with a narrower method Google calls the existing approach", "scope": "servers", "included": [ "active AI accelerators (TPU and GPU) only" ], "excluded": [ "host CPU and DRAM", "idle machines", "data centre overhead (PUE)", "external networking", "end-user devices", "LLM training and data storage" ], "model": "Gemini Apps (all models serving text prompts)", "place": "Google data centre fleet, subsampled to the 10% most energy-efficient data centres", "period": "May 2025" }, "source": { "title": "Measuring the environmental impact of AI inference", "publisher": "Google Cloud Blog", "kind": "company statement", "url": "https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference", "published": "2025-08-21", "quote": "When we apply this non-comprehensive methodology that only considers active TPU and GPU consumption, we estimate the median Gemini text prompt uses 0.10 Wh of energy, emits 0.02 gCO2e, and consumes 0.12 mL of water. This is an optimistic scenario at best and substantially underestimates the real operational footprint of AI." }, "method": "Google's own narrow comparison case, counting only active accelerator energy and sampling the most energy-efficient data centres; the companion paper (arXiv 2508.15734) states that all its emissions figures are market-based and its water figures are consumption.", "claimant": { "type": "company", "name": "Google", "funder": "none" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "google-search-2009", "activity": "search", "title": "Average Google search, Google's 2009 figure", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.3, "high": 0.3, "per": "Google search", "statistic": "mean", "original": "0.0003 kWh of energy per search, or 1 kJ (converted from kWh to Wh)" }, { "quantity": "co2", "unit": "gCO2", "low": 0.2, "high": 0.2, "per": "Google search", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "The average Google search query in January 2009", "scope": "data centre", "included": [ "the servers that work on the query", "other work performed before the search starts, such as building the search index" ], "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Powering a Google search", "publisher": "Google (Official Google Blog)", "kind": "blog post", "url": "https://googleblog.blogspot.com/2009/01/powering-google-search.html", "published": "2009-01-11", "quote": "Queries vary in degree of difficulty, but for the average query, the servers it touches each work on it for just a few thousandths of a second. Together with other work performed before your search even starts (such as building the search index) this amounts to 0.0003 kWh of energy per search, or 1 kJ. ... In terms of greenhouse gases, one Google search is equivalent to about 0.2 grams of CO2." }, "method": "not stated", "claimant": { "type": "company", "name": "Google (Urs Hölzle, Senior Vice President, Operations)", "funder": "none" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "iea-chatgpt-request-2024", "activity": "ai-text-prompt", "title": "ChatGPT request, IEA Electricity 2024", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 2.9, "high": 2.9, "per": "ChatGPT request", "statistic": "single estimate" } ], "counted": { "workload": "One request to OpenAI's ChatGPT", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "ChatGPT (model not stated)", "place": "not stated", "period": "not stated" }, "source": { "title": "Electricity 2024: Analysis and forecast to 2026", "publisher": "International Energy Agency", "kind": "report", "url": "https://iea.blob.core.windows.net/assets/6b2fd954-2017-408e-bf08-952fdd62118a/Electricity2024-Analysisandforecastto2026.pdf", "published": "2024-01", "quote": "When comparing the average electricity demand of a typical Google search (0.3 Wh of electricity) to OpenAI’s ChatGPT (2.9 Wh per request), and considering 9 billion searches daily, this would require almost 10 TWh of additional electricity in a year." }, "method": "not stated", "claimant": { "type": "independent", "name": "International Energy Agency", "funder": "none" }, "status": "current", "dispute": "Epoch AI (February 2025) argues that the widely cited figure of about 3 Wh per ChatGPT query, which traces to de Vries (2023), is likely an overestimate, and puts a typical GPT-4o query at roughly 0.3 Wh. https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "iea-google-search-2024", "activity": "search", "title": "Typical Google search, as stated by the IEA in Electricity 2024", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.3, "high": 0.3, "per": "Google search", "statistic": "mean" } ], "counted": { "workload": "A typical Google search, compared with a ChatGPT request (2.9 Wh)", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Electricity 2024 - Analysis and forecast to 2026", "publisher": "International Energy Agency", "kind": "report", "url": "https://iea.blob.core.windows.net/assets/6b2fd954-2017-408e-bf08-952fdd62118a/Electricity2024-Analysisandforecastto2026.pdf", "published": "2024-01", "quote": "When comparing the average electricity demand of a typical Google search (0.3 Wh of electricity) to OpenAI's ChatGPT (2.9 Wh per request), and considering 9 billion searches daily, this would require almost 10 TWh of additional electricity in a year." }, "method": "not stated (page 34 gives no source for the 0.3 Wh figure; the source notes for the nearby data centre charts cite de Vries, \"The growing energy footprint of artificial intelligence\", Joule, 2023)", "claimant": { "type": "independent", "name": "International Energy Agency", "funder": "none" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "iea-kamiya-streaming-hour-2019", "activity": "streaming-hour", "title": "One hour of streaming video, 2019 global average (IEA revised estimate)", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 36, "high": 36, "per": "hour of streaming video", "statistic": "single estimate", "accounting": "not stated" }, { "quantity": "energy", "unit": "kWh", "low": 0.077, "high": 0.077, "per": "hour of streaming video", "statistic": "single estimate" } ], "counted": { "workload": "One hour of streaming a Netflix video in 2019, based on average viewing habits", "scope": "end to end", "included": [ "viewing devices", "data transmission networks", "data centres" ], "excluded": "not stated", "model": "IEA analysis with energy intensity figures for data centres and networks updated in December 2020", "place": "global (global average electricity mix)", "period": "2019" }, "source": { "title": "The carbon footprint of streaming video: fact-checking the headlines", "publisher": "International Energy Agency (IEA)", "kind": "report", "url": "https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines", "published": "2020-12-10", "quote": "As a result, the central IEA estimate for one hour of streaming video in 2019 is now 36gCO2, down from 82gCO2 in the original analysis published in February 2020. ... my updated analysis suggests that streaming a Netflix video in 2019 typically consumed around 0.077 kWh of electricity per hour" }, "method": "Bottom-up estimate combining energy intensities of devices, networks and data centres with a global average electricity mix; the commentary also gives the same estimate as about 18 gCO2e per 30-minute show.", "claimant": { "type": "independent", "name": "George Kamiya, International Energy Agency", "funder": "not stated" }, "status": "current", "dispute": "The Shift Project (June 2020) accepted the bitrate error Kamiya found but disputed his network energy-intensity assumptions, which it says explain most of the remaining gap between the two estimates (this response addressed the February 2020 version of the analysis). https://theshiftproject.org/app/uploads/2025/04/2020-06_Did-TSP-overestimate-the-carbon-footprint-of-online-video_EN.pdf", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "jegham-deepseek-r1-long-prompt-2025", "activity": "ai-text-prompt", "title": "Long DeepSeek-R1 query on DeepSeek's own servers, How Hungry is AI? benchmark", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 29.078, "high": 29.078, "per": "long query (10,000 input, 1,500 output tokens)", "statistic": "mean" }, { "quantity": "co2", "unit": "gCO2e", "low": 17, "high": 17, "per": "long query (10,000 input, 1,500 output tokens)", "statistic": "not stated", "accounting": "not stated" } ], "counted": { "workload": "One long-form DeepSeek-R1 query (10,000 input tokens, 1,500 output tokens) served from DeepSeek's own data centres", "scope": "data centre", "included": [ "GPU power", "non-GPU server power (CPUs, SSDs, network, cooling control electronics)", "data centre overhead (PUE)" ], "excluded": [ "embodied emissions and water from hardware manufacturing and supply chains (Scope 3)", "training" ], "model": "DeepSeek-R1, hosted by DeepSeek", "place": "DeepSeek data centres in China (PUE and WUE taken as the average of the thirty most efficient data centres in China; carbon factors from the regions of its known or reported deployments)", "period": "not stated" }, "source": { "title": "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference", "publisher": "arXiv (Jegham, Abdelatti, Koh, Elmoubarki and Hendawi)", "kind": "preprint", "url": "https://arxiv.org/abs/2505.09598v6", "published": "2025-11-24", "quote": "DeepSeek-R1 and DeepSeek-V3 hosted on DeepSeek’s own servers consume 29.078 Wh and 13.162 Wh, while the same models on Azure use just 7.410 Wh and 3.696 Wh, over 70% less energy. ... DeepSeek-R1 (DS) consumes over 200 mL of water and emits approximately 17 gCO2e per long query" }, "method": "Estimated from public API throughput and latency data combined with inferred hardware configurations (Nvidia H800), rated GPU and non-GPU power, PUE and regional carbon intensity, reported as the mean of Monte Carlo draws.", "claimant": { "type": "independent", "name": "Nidhal Jegham et al. (University of Rhode Island and others)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "jegham-gpt4o-short-prompt-2025", "activity": "ai-text-prompt", "title": "Short GPT-4o query, How Hungry is AI? benchmark", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.42, "high": 0.42, "per": "short query (100 input, 300 output tokens)", "statistic": "mean" } ], "counted": { "workload": "One short GPT-4o query (100 input tokens, 300 output tokens)", "scope": "data centre", "included": [ "GPU power", "non-GPU server power (CPUs, SSDs, network, cooling control electronics)", "data centre overhead (PUE)" ], "excluded": [ "embodied emissions and water from hardware manufacturing and supply chains (Scope 3)", "training" ], "model": "GPT-4o (March 2025 version)", "place": "Microsoft Azure data centres (Azure-reported PUE; carbon and water factors from the locations of Microsoft data centres)", "period": "not stated" }, "source": { "title": "How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference", "publisher": "arXiv (Jegham, Abdelatti, Koh, Elmoubarki and Hendawi)", "kind": "preprint", "url": "https://arxiv.org/abs/2505.09598v6", "published": "2025-11-24", "quote": "A single short GPT-4o query consumes 0.42 Wh (±0.13 Wh), exceeding the footprint of a Google search (0.30 Wh) by approximately 40%." }, "method": "Estimated from public API throughput and latency data combined with inferred hardware configurations, rated GPU and non-GPU power, and provider-specific PUE, reported as the mean of Monte Carlo draws (the paper's energy table gives 0.423 ± 0.085 Wh for the same case).", "claimant": { "type": "independent", "name": "Nidhal Jegham et al. (University of Rhode Island and others)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "kamiya-carbon-brief-streaming-2020-02", "activity": "streaming-hour", "title": "Streaming Netflix in 2019, Kamiya's original February 2020 estimate", "figures": [ { "quantity": "energy", "unit": "kWh", "low": 0.12, "high": 0.24, "per": "hour of streaming video", "statistic": "range" }, { "quantity": "co2", "unit": "kgCO2e", "low": 0.028, "high": 0.057, "per": "30-minute show", "statistic": "range", "accounting": "not stated" } ], "counted": { "workload": "Streaming a Netflix video in 2019 (energy per hour; emissions per 30-minute show)", "scope": "end to end", "included": [ "viewing devices", "data transmission networks", "data centres" ], "excluded": "not stated", "model": "Kamiya's bottom-up analysis, original version", "place": "global (current global average electricity mix)", "period": "2019" }, "source": { "title": "Factcheck: What is the carbon footprint of streaming video on Netflix?", "publisher": "Carbon Brief (guest post by George Kamiya of the IEA; archived copy)", "kind": "blog post", "url": "https://web.archive.org/web/20200229140629/https://www.carbonbrief.org/factcheck-what-is-the-carbon-footprint-of-streaming-video-on-netflix", "published": "2020-02-25", "quote": "Taken together, my analysis suggests that streaming a Netflix video in 2019 typically consumed 0.12-0.24kWh of electricity per hour ... Powered by the current global average electricity mix, streaming a 30-minute show on Netflix would release 0.028-0.057kgCO2e (28-57 grammes" }, "method": "Bottom-up estimate combining energy intensities of devices, networks and data centres with the global average electricity mix; viewing devices were estimated at 72% of energy use, transmission 23% and data centres 5%.", "claimant": { "type": "independent", "name": "George Kamiya, International Energy Agency", "funder": "not stated" }, "status": "superseded", "supersededBy": "iea-kamiya-streaming-hour-2019", "dispute": "The Shift Project (June 2020) accepted the bitrate error Kamiya found but disputed his network energy-intensity assumptions. https://theshiftproject.org/app/uploads/2025/04/2020-06_Did-TSP-overestimate-the-carbon-footprint-of-online-video_EN.pdf", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "li-ren-gpt3-water-responses-2025", "activity": "ai-text-prompt", "title": "GPT-3 water for 10 to 50 medium-length responses, Making AI Less Thirsty", "figures": [ { "quantity": "water", "unit": "mL", "low": 500, "high": 500, "per": "10 to 50 medium-length responses", "statistic": "single estimate", "water": "consumption" } ], "counted": { "workload": "10 to 50 medium-length GPT-3 responses (each about 800 words or fewer of input and 150 to 300 words of output), depending on when and where the model is deployed", "scope": "data centre", "included": [ "on-site (Scope 1) water consumed for data centre cooling", "off-site (Scope 2) water consumed in generating the electricity" ], "excluded": "not stated", "model": "GPT-3 (175 billion parameters)", "place": "Microsoft's U.S. data centres (U.S. average and six state locations)", "period": "not stated" }, "source": { "title": "Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models", "publisher": "arXiv (Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren); accepted by Communications of the ACM", "kind": "preprint", "url": "https://arxiv.org/abs/2304.03271v5", "published": "2025-03-26", "quote": "Additionally, GPT-3 needs to “drink” (i.e., consume) a 500ml bottle of water for roughly 10 – 50 medium-length responses, depending on when and where it is deployed." }, "method": "Per-request server energy of 0.004 kWh (from OpenAI's estimate for GPT-3) multiplied by on-site water usage effectiveness plus PUE times the off-site electricity water intensity, using Microsoft-reported and location-specific values.", "claimant": { "type": "independent", "name": "Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren (UC Riverside, UT Arlington)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "luccioni-image-generation-least-efficient-2024", "activity": "ai-image", "title": "Least efficient open image generation model, Power Hungry Processing", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 11.49, "high": 11.49, "per": "image generated", "statistic": "single estimate", "original": "11.49 kWh for 1,000 inferences; converted from kWh per 1,000 to Wh per one" } ], "counted": { "workload": "1,000 image generations, run sequentially without batching, with the least efficient of the eight image generation models tested", "scope": "servers", "included": [ "GPU, CPU and RAM energy measured with Code Carbon" ], "excluded": "not stated", "model": "not stated (the sentence does not name the model)", "place": "AWS us-west-2 (Oregon)", "period": "not stated" }, "source": { "title": "Power Hungry Processing: Watts Driving the Cost of AI Deployment?", "publisher": "ACM FAccT '24 (arXiv 2311.16863)", "kind": "peer-reviewed paper", "url": "https://arxiv.org/abs/2311.16863", "published": "2024-06", "quote": "whereas the least efficient image generation model uses as much energy as 522 smartphone charges (11.49 kWh), or around half a charge per image generation" }, "method": "Each model ran 1,000 inferences on each of three datasets, ten times, on a node of 8 NVIDIA A100-SXM4-80GB GPUs, with energy measured by the Code Carbon package.", "claimant": { "type": "company", "name": "Alexandra Sasha Luccioni, Yacine Jernite (Hugging Face) and Emma Strubell (Carnegie Mellon University)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "luccioni-image-generation-mean-2024", "activity": "ai-image", "title": "Mean image generation across open text-to-image models, Power Hungry Processing", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 2.907, "high": 2.907, "per": "image generated", "statistic": "mean", "original": "2.907 kWh per 1,000 queries (Table 2; the text rounds it to 2.9kWh); converted from kWh per 1,000 to Wh per one" } ], "counted": { "workload": "1,000 image generations, run sequentially without batching, on prompts from each of three datasets (DiffusionDB, ImageReward, SD Prompts)", "scope": "servers", "included": [ "GPU, CPU and RAM energy measured with Code Carbon" ], "excluded": "not stated", "model": "Eight text-to-image models (stable-diffusion-v1-4, v1-5, 2-1 and xl-base-1.0, openjourney, dreamlike-photoreal-2.0, bk-sdm-tiny, tiny-sd), per the paper's Table 6", "place": "AWS us-west-2 (Oregon)", "period": "not stated" }, "source": { "title": "Power Hungry Processing: Watts Driving the Cost of AI Deployment?", "publisher": "ACM FAccT '24 (arXiv 2311.16863)", "kind": "peer-reviewed paper", "url": "https://arxiv.org/abs/2311.16863", "published": "2024-06", "quote": "Table 2. Mean and standard deviation of energy per 1,000 queries for the ten tasks examined in our analysis. ... to the most energy-intensive one, image generation, whose mean consumption is 2.9kWh." }, "method": "Each model ran 1,000 inferences on each of three datasets, ten times, on a node of 8 NVIDIA A100-SXM4-80GB GPUs, with energy measured by the Code Carbon package.", "claimant": { "type": "company", "name": "Alexandra Sasha Luccioni, Yacine Jernite (Hugging Face) and Emma Strubell (Carnegie Mellon University)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "luccioni-text-generation-inference-2024", "activity": "ai-text-prompt", "title": "Text generation inference, Power Hungry Processing", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.047, "high": 0.047, "per": "text generation inference", "statistic": "mean", "original": "0.047 kWh per 1,000 inferences (0.047 kWh / 1,000 = 0.047 Wh per inference)" } ], "counted": { "workload": "One text-generation inference generating 10 new tokens, unbatched, on benchmark datasets", "scope": "servers", "included": [ "GPU, CPU and RAM energy measured with the Code Carbon package" ], "excluded": "not stated", "model": "The 8 most popular text-generation models on the Hugging Face Hub at the time of the study", "place": "Amazon Web Services us-west-2 (Oregon), one node of 8 Nvidia A100-SXM4-80GB GPUs", "period": "not stated" }, "source": { "title": "Power Hungry Processing: Watts Driving the Cost of AI Deployment?", "publisher": "ACM FAccT 2024 (Luccioni, Jernite and Strubell; arXiv version)", "kind": "peer-reviewed paper", "url": "https://arxiv.org/abs/2311.16863", "published": "2024-06", "quote": "image generation uses, on average, over 60 times more energy than text generation (0.047 vs. 2.9 kWh)." }, "method": "Each model ran 1,000 inferences on each of 3 datasets, repeated 10 times without batching, with energy measured by Code Carbon; Table 2 gives the mean (0.047 kWh) and standard deviation (0.03) of energy per 1,000 queries for text generation.", "claimant": { "type": "company", "name": "Alexandra Sasha Luccioni, Yacine Jernite (Hugging Face) and Emma Strubell (Carnegie Mellon University)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "microsoft-frontier-query-2026", "activity": "ai-text-prompt", "title": "Frontier-scale model query on H100 nodes, Microsoft researchers' estimate", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.31, "high": 0.31, "per": "query", "statistic": "median" } ], "counted": { "workload": "One traditional (non-reasoning) query to a frontier-scale model of more than 200 billion parameters, median 300 output tokens and 500 input tokens", "scope": "data centre", "included": [ "GPU node power (8 or 10 Nvidia H100 GPUs, modelled around 70% of maximum power)", "data centre overhead (PUE, modelled between 1.05 and 1.40)" ], "excluded": [ "extra orchestration, tool calling and long-context management in agentic workflows" ], "model": "Frontier-scale models above 200 billion parameters (DeepSeek-R1 671B, Llama 3.1 405B, Llama 3.1 Nemotron Ultra 253B)", "place": "not stated", "period": "not stated" }, "source": { "title": "Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute", "publisher": "Joule (Oviedo et al., Microsoft; arXiv version 2 read)", "kind": "peer-reviewed paper", "url": "https://arxiv.org/abs/2509.20241v2", "published": "2026-06-09", "quote": "For frontier-scale models (>200B parameters) on H100 nodes, we estimate a median energy of 0.31 Wh/query (IQR 0.16–0.60), indicating widely cited estimates are overstated by 4–20×." }, "method": "Bottom-up Monte Carlo model from token throughput, node power and PUE under production serving assumptions (batching, FP8); interquartile range 0.16 to 0.60 Wh.", "claimant": { "type": "company", "name": "Microsoft (Felipe Oviedo et al.)", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mistral-le-chat-response-2025", "activity": "ai-text-prompt", "title": "Le Chat 400-token response, Mistral life-cycle analysis", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 1.14, "high": 1.14, "per": "400-token response", "statistic": "single estimate", "accounting": "location-based" }, { "quantity": "water", "unit": "mL", "low": 45, "high": 45, "per": "400-token response", "statistic": "single estimate", "water": "consumption" } ], "counted": { "workload": "The marginal impact of inference for one 400-token response from Le Chat", "scope": "not stated", "included": [ "marginal impacts of inference", "upstream emissions, such as from manufacturing servers" ], "excluded": [ "users' terminals" ], "model": "Le Chat (the study is a life-cycle analysis of Mistral Large 2)", "place": "not stated", "period": "not stated" }, "source": { "title": "Our contribution to a global environmental standard for AI", "publisher": "Mistral AI", "kind": "company statement", "url": "https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai", "published": "2025-07-22", "quote": "the marginal impacts of inference, more precisely the use of our AI assistant Le Chat for a 400-token response - excluding users’ terminals: 1.14 gCO₂e, 45 mL of water, and 0.16 mg of Sb eq." }, "method": "Life-cycle analysis with Carbone 4 and the French ecological transition agency (ADEME), peer-reviewed by Resilio and Hubblo, following the AFNOR Frugal AI methodology, the GHG Protocol Product Standard and ISO 14040/44; water is Water Consumption Potential and electricity emissions are location-based.", "claimant": { "type": "company", "name": "Mistral AI", "funder": "none" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-cogvideox-5s-video-2025", "activity": "ai-video", "title": "Five-second CogVideoX video at 16 frames per second (MIT Technology Review)", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 944.4, "high": 944.4, "per": "5-second video at 16 frames per second", "statistic": "single estimate", "original": "about 3.4 million joules; converted from J to Wh" } ], "counted": { "workload": "One 5-second video at 16 frames per second from the newer, larger CogVideoX model", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "CogVideoX (the larger model launched three months after the August version; exact version not stated)", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "But three months later the company launched a larger, higher-quality model that produces five-second videos at 16 frames per second ... The new model uses more than 30 times more energy on each 5-second video: about 3.4 million joules, more than 700 times the energy required to generate a high-quality image." }, "method": "Measured by Sasha Luccioni (Hugging Face) with Code Carbon; the article does not say whether the doubling for non-GPU energy was applied.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-cogvideox-older-video-2025", "activity": "ai-video", "title": "Older CogVideoX video at 8 frames per second (MIT Technology Review)", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 30.3, "high": 30.3, "per": "video at 8 frames per second (length not stated)", "statistic": "single estimate", "original": "about 109,000 joules; converted from J to Wh" } ], "counted": { "workload": "One video from the older CogVideoX model, at 8 frames per second and a grainy resolution", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "CogVideoX (the older version released in August; exact version not stated)", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "An older version of the model, released in August, made videos at just eight frames per second at a grainy resolution—more like a GIF than a video. Each one required about 109,000 joules to produce." }, "method": "Measured by Sasha Luccioni (Hugging Face) with Code Carbon; the article does not say whether the doubling for non-GPU energy was applied.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-llama-3-1-405b-response-2025", "activity": "ai-text-prompt", "title": "Llama 3.1 405B text response, MIT Technology Review analysis", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 1.86, "high": 1.86, "per": "text response", "statistic": "mean", "original": "an estimated 6,706 joules total (6,706 J / 3,600 = 1.863 Wh)" } ], "counted": { "workload": "One response from Llama 3.1 405B to a variety of text-generating prompts", "scope": "data centre", "included": [ "measured GPU energy (3,353 joules per response on average)", "estimate for cooling, other computations and other demands (GPU energy doubled)" ], "excluded": "not stated", "model": "Llama 3.1 405B", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "On average, this model needed 3,353 joules, or an estimated 6,706 joules total, for each response." }, "method": "GPU energy measured on Nvidia H100 GPUs by Jae-Won Chung (ML.Energy, University of Michigan), then doubled to estimate the whole operation, following a 2024 Microsoft paper.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-llama-3-1-8b-response-2025", "activity": "ai-text-prompt", "title": "Llama 3.1 8B text response, MIT Technology Review analysis", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.032, "high": 0.032, "per": "text response", "statistic": "not stated", "original": "an estimated 114 joules (114 J / 3,600 = 0.0317 Wh)" } ], "counted": { "workload": "One response from Llama 3.1 8B to a variety of text-generating prompts", "scope": "data centre", "included": [ "measured GPU energy (about 57 joules per response)", "estimate for cooling, other computations and other demands (GPU energy doubled)" ], "excluded": "not stated", "model": "Llama 3.1 8B", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "When tested on a variety of different text-generating prompts, like making a travel itinerary for Istanbul or explaining quantum computing, the model required about 57 joules per response, or an estimated 114 joules when accounting for cooling, other computations, and other demands." }, "method": "GPU energy measured on Nvidia H100 GPUs by Jae-Won Chung (ML.Energy, University of Michigan), then doubled to estimate the whole operation, following a 2024 Microsoft paper.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-sd3-medium-image-2025-gpu", "activity": "ai-image", "title": "Stable Diffusion 3 Medium 1024 × 1024 image, GPU energy only (MIT Technology Review)", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.317, "high": 0.317, "per": "image generated (1024 × 1024 pixels, standard quality)", "statistic": "single estimate", "original": "about 1,141 joules of GPU energy; converted from J to Wh" } ], "counted": { "workload": "One standard-quality 1024 × 1024 image from Stable Diffusion 3 Medium (2 billion parameters)", "scope": "servers", "included": [ "GPU energy" ], "excluded": [ "CPUs, fans and other equipment" ], "model": "Stable Diffusion 3 Medium", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "Generating a standard-quality image (1024 x 1024 pixels) with Stable Diffusion 3 Medium, the leading open-source image generator, with 2 billion parameters, requires about 1,141 joules of GPU energy." }, "method": "GPU energy measured by Jae-Won Chung (University of Michigan, ML.Energy) for the article.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-sd3-medium-image-2025-total", "activity": "ai-image", "title": "Stable Diffusion 3 Medium 1024 × 1024 image, estimated total energy (MIT Technology Review)", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 0.634, "high": 0.634, "per": "image generated (1024 × 1024 pixels, standard quality)", "statistic": "single estimate", "original": "an estimated 2,282 joules total; converted from J to Wh" } ], "counted": { "workload": "One standard-quality 1024 × 1024 image from Stable Diffusion 3 Medium (2 billion parameters)", "scope": "data centre", "included": [ "GPU energy", "an estimate for CPUs, fans and other equipment (GPU energy doubled)" ], "excluded": "not stated", "model": "Stable Diffusion 3 Medium", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "With diffusion models, unlike large language models, there are no estimates of how much GPUs are responsible for the total energy required, but experts suggested we stick with the “doubling” approach we’ve used thus far because the differences are likely subtle. That means an estimated 2,282 joules total." }, "method": "GPU energy measured by Jae-Won Chung (University of Michigan, ML.Energy), then doubled to estimate the whole operation, following a 2024 Microsoft paper on LLM inference.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mit-tr-sd3-medium-image-50-steps-2025", "activity": "ai-image", "title": "Stable Diffusion 3 Medium image at 50 diffusion steps (MIT Technology Review)", "figures": [ { "quantity": "energy", "unit": "Wh", "low": 1.223, "high": 1.223, "per": "image generated (1024 × 1024 pixels, 50 diffusion steps)", "statistic": "single estimate", "original": "about 4,402 joules; converted from J to Wh" } ], "counted": { "workload": "One higher-quality 1024 × 1024 image from Stable Diffusion 3 Medium, with the number of diffusion steps doubled to 50", "scope": "data centre", "included": "not stated", "excluded": "not stated", "model": "Stable Diffusion 3 Medium", "place": "not stated", "period": "not stated" }, "source": { "title": "We did the math on AI’s energy footprint. Here’s the story you haven’t heard.", "publisher": "MIT Technology Review", "kind": "press article", "url": "https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/", "published": "2025-05-20", "quote": "Improving the image quality by doubling the number diffusion steps to 50 just about doubles the energy required, to about 4,402 joules." }, "method": "Measured by Jae-Won Chung (University of Michigan, ML.Energy) for the article; the sentence follows the doubled 2,282 J total but does not say explicitly whether 4,402 J includes the doubling for non-GPU energy.", "claimant": { "type": "independent", "name": "MIT Technology Review (James O'Donnell and Casey Crownhart)", "funder": "Tarbell Center for AI Journalism (grant)" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mortas-video-call-camera-off-2026", "activity": "video-call-hour", "title": "One-hour Teams call with the camera off, over 4G (network data only)", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 4.8, "high": 10.8, "per": "hour of video call with the camera off", "statistic": "range", "accounting": "not stated" } ], "counted": { "workload": "A one-hour Microsoft Teams call with the camera off, measured at 8.76 MB per minute on a 4G connection", "scope": "network", "included": [ "electricity for mobile (4G) network data transfer" ], "excluded": [ "data centres", "end-user devices", "screen sharing" ], "model": "Measured data volume times mobile-network energy intensity (kWh/GB, minimum and maximum from the literature) times the French electricity carbon intensity (ADEME, 79.1 gCO2e/kWh)", "place": "France", "period": "not stated" }, "source": { "title": "Assessing the Carbon Footprint of Virtual Meetings: A Quantitative Analysis of Camera Usage", "publisher": "arXiv (Félix Mortas, independent researcher)", "kind": "preprint", "url": "https://arxiv.org/abs/2601.06045", "published": "2025-12-16", "quote": "Turning the camera off reduces these emissions to 4.8 to 10.8 g of CO2e per hour." }, "method": "Mobile data use was read from the operator app before and after each of nine 30 to 60 minute Teams meetings and converted to emissions with mobile-network energy intensities and the French grid factor; figures as in version 2 (18 January 2026).", "claimant": { "type": "independent", "name": "Félix Mortas", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "mortas-video-call-camera-on-2026", "activity": "video-call-hour", "title": "One-hour Teams call with the camera on, over 4G (network data only)", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 10.2, "high": 21.6, "per": "hour of video call with the camera on", "statistic": "range", "accounting": "not stated" } ], "counted": { "workload": "A one-hour Microsoft Teams call with the camera on, measured at 18.07 MB per minute on a 4G connection", "scope": "network", "included": [ "electricity for mobile (4G) network data transfer" ], "excluded": [ "data centres", "end-user devices", "screen sharing" ], "model": "Measured data volume times mobile-network energy intensity (kWh/GB, minimum and maximum from the literature) times the French electricity carbon intensity (ADEME, 79.1 gCO2e/kWh)", "place": "France", "period": "not stated" }, "source": { "title": "Assessing the Carbon Footprint of Virtual Meetings: A Quantitative Analysis of Camera Usage", "publisher": "arXiv (Félix Mortas, independent researcher)", "kind": "preprint", "url": "https://arxiv.org/abs/2601.06045", "published": "2025-12-16", "quote": "For video calls, having the camera on results in emissions ranging from 10.2 to 21.6 g of CO2e per hour." }, "method": "Mobile data use was read from the operator app before and after each of nine 30 to 60 minute Teams meetings and converted to emissions with mobile-network energy intensities and the French grid factor; figures as in version 2 (18 January 2026).", "claimant": { "type": "independent", "name": "Félix Mortas", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "netflix-streaming-hour-worldwide-2021", "activity": "streaming-hour", "title": "One hour of Netflix streaming worldwide, \"well below 100 gCO2e\"", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 100, "high": 100, "per": "hour of streaming", "statistic": "upper bound", "original": "well below 100gCO2e", "accounting": "not stated" } ], "counted": { "workload": "One hour of Netflix streaming worldwide, from Netflix's own data run through the DIMPACT tool", "scope": "not stated", "included": [ "electricity consumption during use" ], "excluded": [ "manufacturing of devices", "end-of-life of devices" ], "model": "DIMPACT carbon calculator (University of Bristol)", "place": "worldwide (includes grids in the U.S. and Canada, Latin America and Asia-Pacific)", "period": "not stated" }, "source": { "title": "The True Climate Impact of Streaming", "publisher": "Netflix", "kind": "company statement", "url": "https://about.netflix.com/en/news/the-true-climate-impact-of-streaming", "published": "2021-06-10", "quote": "Netflix recently ran its own data through the tool and found that the emissions from an hour of streaming the world over is well below 100gCO2e, or less than driving a gas-powered vehicle a quarter mile or 400 metres." }, "method": "Netflix ran its own data through the DIMPACT calculator; it published only an upper bound (\"well below 100gCO2e\"), recorded as that bound, not as an estimate.", "claimant": { "type": "company", "name": "Netflix (Emma Stewart, Netflix Sustainability Officer, with Daniel Schien, University of Bristol)", "funder": "none" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "ovo-thank-you-email-2019", "activity": "email", "title": "One email, 1 g CO2e (OVO Energy \"Think Before You Thank\")", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 1, "high": 1, "per": "email", "statistic": "single estimate", "original": "1g carbon = 0.000001 tonnes CO2e per email", "accounting": "not stated" } ], "counted": { "workload": "One email, used for unnecessary \"thank you\" emails in the UK", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "UK", "period": "not stated" }, "source": { "title": "'Think Before You Thank': If every Brit sent one less thank you email a day, we would save 16,433 tonnes of carbon a year - the same as 81,152 flights to Madrid", "publisher": "OVO Energy (archived copy; the original page has been removed)", "kind": "press release", "url": "https://web.archive.org/web/20220302172156/https://www.ovoenergy.com/ovo-newsroom/press-releases/2019/november/think-before-you-thank-if-every-brit-sent-one-less-thank-you-email-a-day-we-would-save-16433-tonnes-of-carbon-a-year-the-same-as-81152-flights-to-madrid", "published": "2019-11-26", "quote": "x1 email = 1g carbon = 0.000001 tonnes CO2e per email (Source: Mike Berners-Lee)" }, "method": "The press release takes 1 g per email from Mike Berners-Lee and multiplies it by survey-based counts of unnecessary emails (Censuswide, 3,000 British adults) to reach 16,433 tonnes a year.", "claimant": { "type": "company", "name": "OVO Energy (citing Mike Berners-Lee)", "funder": "none" }, "status": "current", "dispute": "BBC News (November 2020) reported that Mike Berners-Lee, whose research was used, told the Financial Times the work was based on \"back-of-the-envelope\" maths from 2010, and quoted Chris Preist (University of Bristol) that most of the system uses energy whether or not an email is sent, so \"the carbon saved will be far far less than 1g per email.\" https://www.bbc.co.uk/news/technology-55002423", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "purdue-obringer-video-call-hour-2021", "activity": "video-call-hour", "title": "One hour of videoconferencing or streaming, Purdue/Yale/MIT study", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 150, "high": 1000, "per": "hour of videoconferencing or streaming", "statistic": "range", "accounting": "not stated" }, { "quantity": "water", "unit": "L", "low": 2, "high": 12, "per": "hour of videoconferencing or streaming", "statistic": "range", "water": "unknown" } ], "counted": { "workload": "One hour of videoconferencing or streaming (the release adds that leaving the camera off cuts these footprints by 96%)", "scope": "not stated", "included": [ "processing and transmission of internet data" ], "excluded": "not stated", "model": "Footprints per gigabyte of data for 18 platforms, online gaming and web surfing, using models from Madani's research group and known energy use per gigabyte of fixed-line internet", "place": "not stated", "period": "not stated" }, "source": { "title": "Turn off that camera during virtual meetings, environmental study says", "publisher": "Purdue University (press release on Obringer et al., Resources, Conservation & Recycling, 2021)", "kind": "press release", "url": "https://www.purdue.edu/newsroom/releases/2021/Q1/turn-off-that-camera-during-virtual-meetings,-environmental-study-says.html", "published": "2021-01-14", "quote": "Just one hour of videoconferencing or streaming, for example, emits 150-1,000 grams of carbon dioxide (a gallon of gasoline burned from a car emits about 8,887 grams), requires 2-12 liters of water and demands a land area adding up to about the size of an iPad Mini. But leaving your camera off during a web call can reduce these footprints by 96%." }, "method": "Estimated from the data volume of each platform multiplied by per-gigabyte carbon, water and land footprints of electricity for internet data; the researchers call the estimates rough.", "claimant": { "type": "independent", "name": "Renee Obringer, Benjamin Rachunok, Debora Maia-Silva, Maryam Arbabzadeh, Roshanak Nateghi and Kaveh Madani (Purdue, Yale, MIT)", "funder": "Purdue Climate Change Research Center, Purdue Center for the Environment, MIT Energy Initiative, Yale MacMillan Center" }, "status": "current", "dispute": "Félix Mortas (arXiv preprint, 2026) measured data use on Microsoft Teams calls over 4G and found that turning the camera off roughly halves data use and the related emissions, which it presents as a challenge to this study's 96% figure, noting that the study is based on estimates and modelling rather than experimental data. https://arxiv.org/abs/2601.06045", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "shift-project-afp-half-hour-show-2019", "activity": "streaming-hour", "title": "Half-hour show, 1.6 kg CO2e (Shift Project figure as reported by AFP)", "figures": [ { "quantity": "co2", "unit": "kgCO2e", "low": 1.6, "high": 1.6, "per": "half-hour show", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "Watching a half-hour show online", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Chill your Netflix habit, climate experts say", "publisher": "AFP (France 24)", "kind": "press article", "url": "https://www.france24.com/en/20191028-chill-your-netflix-habit-climate-experts-say", "published": "2019-10-28", "quote": "Watching a half-hour show would lead to emissions of 1.6 kilograms of carbon dioxide equivalent, said Maxime Efoui-Hess of French think tank the Shift Project. That's equivalent to driving 3.9 miles (6.28 kilometres)." }, "method": "not stated", "claimant": { "type": "independent", "name": "Maxime Efoui-Hess, The Shift Project (in an interview with AFP)", "funder": "not stated" }, "status": "corrected", "dispute": "George Kamiya (Carbon Brief, February 2020; IEA, December 2020) traced the figure to a bitrate of 3 Mbps converted in error to 3 megabytes per second, and put the 2019 figure at about 18 gCO2e per half hour. The Shift Project accepted the error in June 2020, said the figure came from an unchecked oral interview and is not in its reports, and gave a corrected 0.400 kgCO2e per hour (see shift-project-online-video-hour-2020). https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "shift-project-online-video-hour-2020", "activity": "streaming-hour", "title": "One hour of online video, The Shift Project after its bitrate correction", "figures": [ { "quantity": "co2", "unit": "kgCO2e", "low": 0.4, "high": 0.4, "per": "hour of video viewed", "statistic": "single estimate", "original": "0.400 kgCO2e per hour", "accounting": "not stated" }, { "quantity": "energy", "unit": "kWh", "low": 0.77, "high": 0.77, "per": "hour of video viewed", "statistic": "single estimate" } ], "counted": { "workload": "One hour of online video viewed, from The Shift Project's \"1byte\" model after correcting the bit-to-byte bitrate error", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "The Shift Project \"1byte\" model", "place": "not stated", "period": "not stated" }, "source": { "title": "Did The Shift Project really overestimate the carbon footprint of online video? Our analysis of the IEA and CarbonBrief articles", "publisher": "The Shift Project", "kind": "report", "url": "https://theshiftproject.org/app/uploads/2025/04/2020-06_Did-TSP-overestimate-the-carbon-footprint-of-online-video_EN.pdf", "published": "2020-06-15", "quote": "3/ After correction of the bitrate, our model yields 0.400 kgCO2e per hour. ... After correction of the bitrate (bit-to-byte conversion error, i.e. a division by 8 of the bitrate), our model indicates 0.77 kWh per hour of video. However, we mention this figure merely on an indicative basis" }, "method": "The Shift Project's \"1byte\" model with the bitrate divided by 8; the Shift Project says the model is not designed to measure the impact of an individual action and gives the energy figure on an indicative basis only.", "claimant": { "type": "independent", "name": "The Shift Project (Maxime Efoui-Hess and Jean-Noël Geist)", "funder": "not stated" }, "status": "current", "dispute": "The IEA (George Kamiya, December 2020) puts 2019 streaming at 0.077 kWh per hour, 10 times less than the corrected Shift Project energy estimate (0.78 kWh), and says the Shift Project corrected the bitrate error but did not revise its other assumptions, which overestimate the energy intensity of data centres and networks. https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "sunday-times-google-search-2009", "activity": "search", "title": "Typical Google search, as reported by The Sunday Times (\"two searches boil a kettle\")", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 7, "high": 7, "per": "Google search", "statistic": "single estimate", "accounting": "not stated" } ], "counted": { "workload": "A typical Google search from a desktop computer", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Revealed: the environmental impact of Google searches", "publisher": "The Sunday Times (Times Online; archived copy of 15 January 2009 on the Wayback Machine)", "kind": "press article", "url": "https://web.archive.org/web/20090115112144/http://technology.timesonline.co.uk/tol/news/tech_and_web/article5489134.ece", "published": "2009-01-11", "quote": "Physicist Alex Wissner-Gross says that performing two Google searches uses up as much energy as boiling the kettle for a cup of tea ... Performing two Google searches from a desktop computer can generate about the same amount of carbon dioxide as boiling a kettle for a cup of tea, according to new research. While millions of people tap into Google without considering the environment, a typical search generates about 7g of CO2 Boiling a kettle generates about 15g." }, "method": "not stated (the body attributes the figure to \"new research\"; the standfirst attributes the kettle comparison to Alex Wissner-Gross)", "claimant": { "type": "independent", "name": "The Sunday Times (Jonathan Leake and Richard Woods)", "funder": "none" }, "status": "corrected", "dispute": "Google's official blog (11 January 2009) said the claim that a typical search produces 7 grams of CO2 is many times too high, and gave about 0.2 g CO2 per search (https://googleblog.blogspot.com/2009/01/powering-google-search.html). Alex Wissner-Gross told TechNewsWorld that his work had nothing to do with Google and that the kettle comparison was the newspaper's, not his (https://web.archive.org/web/20091217121803/http://www.technewsworld.com/story/Harvard-Prof-Sets-Record-Straight-on-Internet-Carbon-Study-65794.html). On 16 January 2009 The Sunday Times added a clarification: the 7 g does not refer to a one-hit search taking less than a second, for which it accepts Google's figure of about 0.2 g, but to a search that may involve several attempts and last several minutes, for which experts gave 1 g to 10 g (https://web.archive.org/web/20090729054637/http://technology.timesonline.co.uk/tol/news/tech_and_web/article5489134.ece).", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "website-carbon-page-view-median-2023", "activity": "website-visit", "title": "Median web page view as of July 2023, Website Carbon Calculator", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 0.8, "high": 0.8, "per": "page view", "statistic": "median", "accounting": "not stated" } ], "counted": { "workload": "One view of a typical web page (median of pages in HTTP Archive data)", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "global", "period": "July 2023" }, "source": { "title": "FAQ - Website Carbon Calculator", "publisher": "Wholegrain Digital (Website Carbon Calculator)", "kind": "company statement", "url": "https://www.websitecarbon.com/faq/", "published": "2023-07", "quote": "We used data from HTTP Archive. This is the median average and not the mean average. ... For reference, the median is approximately 0.8 grams per page view as of July 2023." }, "method": "Median of HTTP Archive page data; the FAQ says CO2 on the site always means CO2e.", "claimant": { "type": "company", "name": "Wholegrain Digital", "funder": "none" }, "status": "superseded", "supersededBy": "website-carbon-page-view-v4", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "website-carbon-page-view-v4", "activity": "website-visit", "title": "Average web page view, Website Carbon Calculator (Sustainable Web Design Model v4)", "figures": [ { "quantity": "co2", "unit": "gCO2e", "low": 0.36, "high": 0.36, "per": "page view", "statistic": "not stated", "accounting": "not stated" } ], "counted": { "workload": "One view of a typical web page, averaged globally", "scope": "end to end", "included": [ "data centres (hosting)", "transmission networks", "user devices", "operational and embodied emissions" ], "excluded": [ "the type of user device (not taken into account)" ], "model": "Sustainable Web Design Model v4", "place": "global (global average grid intensity)", "period": "not stated" }, "source": { "title": "Website Carbon Calculator v4", "publisher": "Wholegrain Digital (Website Carbon Calculator)", "kind": "company statement", "url": "https://www.websitecarbon.com/", "published": "2025-09-17", "quote": "Globally, the average web page produces approximately 0.36 grams CO2 equivalent per pageview." }, "method": "Page data transfer converted to emissions with the Sustainable Web Design Model v4 (adopted by the calculator on 14 July 2025); the page says \"average\" and does not say which average.", "claimant": { "type": "company", "name": "Wholegrain Digital", "funder": "none" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] }, { "id": "wissner-gross-website-second-2009", "activity": "website-visit", "title": "One second of visiting a typical website, Alex Wissner-Gross's figure", "figures": [ { "quantity": "co2", "unit": "gCO2", "low": 0.02, "high": 0.02, "per": "second of browsing a typical website", "statistic": "mean", "original": "about 20 milligrams of CO2 per second (converted from mg to g)", "accounting": "not stated" } ], "counted": { "workload": "Visiting a website, averaged over the web as a whole", "scope": "not stated", "included": "not stated", "excluded": "not stated", "model": "not stated", "place": "not stated", "period": "not stated" }, "source": { "title": "Harvard Physicist Sets Record Straight on Internet Carbon Study", "publisher": "TechNewsWorld (Renay San Miguel; archived copy on the Wayback Machine)", "kind": "press article", "url": "https://web.archive.org/web/20091217121803/http://www.technewsworld.com/story/Harvard-Prof-Sets-Record-Straight-on-Internet-Carbon-Study-65794.html", "published": "2009-01-12", "quote": "\"Our work has nothing to do with Google. Our focus was exclusively on the Web overall, and we found that it takes on average about 20 milligrams of CO2 per second to visit a Web site.\"" }, "method": "not stated", "claimant": { "type": "company", "name": "Alex Wissner-Gross (Harvard University), who ran CO2Stats.com at the time (this site's 2009 predecessor), quoted by TechNewsWorld", "funder": "not stated" }, "status": "current", "dispute": "none", "history": [ { "date": "2026-10-16", "change": "Added." } ] } ], "factors": [ { "activity": "ai-text-prompt", "label": "AI chat prompt (Google's median Gemini text prompt, May 2025)", "per": "text prompt", "scope": "data centre", "energy": { "low": 0.24, "high": 0.24, "from": "google-gemini-prompt-2025-comprehensive", "unit": "Wh" }, "co2": { "low": 0.03, "high": 0.03, "from": "google-gemini-prompt-2025-comprehensive", "unit": "gCO2e", "accounting": "market-based" }, "water": { "low": 0.26, "high": 0.26, "from": "google-gemini-prompt-2025-comprehensive", "unit": "mL", "water": "consumption" }, "approved": { "date": "2026-10-16", "by": "the build driver, as editor under the owner's delegation" } }, { "activity": "ai-image", "label": "AI image (mean of eight open models, one at a time on GPU servers, 2024 benchmark)", "per": "image generated", "scope": "servers", "energy": { "low": 2.907, "high": 2.907, "from": "luccioni-image-generation-mean-2024", "unit": "Wh" }, "co2": "unknown", "water": "unknown", "approved": { "date": "2026-10-16", "by": "the build driver, as editor under the owner's delegation" } }, { "activity": "search", "label": "Web search (Google's 2009 estimate)", "per": "Google search", "scope": "data centre", "energy": { "low": 0.3, "high": 0.3, "from": "google-search-2009", "unit": "Wh" }, "co2": { "low": 0.2, "high": 0.2, "from": "google-search-2009", "unit": "gCO2", "accounting": "not stated" }, "water": "unknown", "approved": { "date": "2026-10-16", "by": "the build driver, as editor under the owner's delegation" } }, { "activity": "streaming-hour", "label": "Video streaming (IEA, 2019 global average)", "per": "hour of streaming video", "scope": "end to end", "energy": { "low": 0.077, "high": 0.077, "from": "iea-kamiya-streaming-hour-2019", "unit": "kWh" }, "co2": { "low": 36, "high": 36, "from": "iea-kamiya-streaming-hour-2019", "unit": "gCO2", "accounting": "not stated" }, "water": "unknown", "approved": { "date": "2026-10-16", "by": "the build driver, as editor under the owner's delegation" } }, { "activity": "video-call-hour", "label": "Video call, camera on (4G mobile network only, France)", "per": "hour of video call with the camera on", "scope": "network", "energy": "unknown", "co2": { "low": 10.2, "high": 21.6, "from": "mortas-video-call-camera-on-2026", "unit": "gCO2e", "accounting": "not stated" }, "water": "unknown", "approved": { "date": "2026-10-16", "by": "the build driver, as editor under the owner's delegation" } }, { "activity": "website-visit", "label": "Website visit (Website Carbon's global average page)", "per": "page view", "scope": "end to end", "energy": "unknown", "co2": { "low": 0.36, "high": 0.36, "from": "website-carbon-page-view-v4", "unit": "gCO2e", "accounting": "not stated" }, "water": "unknown", "approved": { "date": "2026-10-16", "by": "the build driver, as editor under the owner's delegation" } } ] }