Training GPT-3 used an estimated 1,287 MWh of electricity and emitted about 552 tonnes of CO₂e, roughly three round trips of a passenger jet between San Francisco and New York.
Patterson et al. (Google), 2021
CorroboratedThe computation used in deep-learning research doubled every few months, an estimated 300,000× increase from 2012 to 2018.
Schwartz et al., 2019
Training GPT-3 in Microsoft's US data centres could directly evaporate 700,000 litres of clean freshwater, and that figure was not disclosed.
Li, Yang, Islam & Ren, 2023
Global data-centre electricity demand is set to more than double to about 945 TWh by 2030, growing nearly 12% a year from about 1.5% of world electricity today.
IEA, cited in You et al., 2026
Better practice, cleaner grids and leaner serving: how far the footprint per job can fall.
Four best practices, the right model, hardware, data centre and location, can cut ML training energy up to 100× and emissions up to 1000×.
Patterson et al. (Google / UC Berkeley), 2022
CorroboratedFollowing those practices, machine learning held steady at under 15% of Google's total energy use for three years, even as its use grew.
Patterson et al. (Google / UC Berkeley), 2022
Company-reportedSoftware efficiency and clean-energy procurement cut energy per median Gemini text prompt 33× and its carbon footprint 44× over a single year.
Elsworth et al. (Google), 2025
Company-reportedBLOOM's final training emitted about 24.7 tonnes of CO₂e counting dynamic power only, and 50.5 tonnes including equipment manufacturing and idle consumption.
Luccioni et al., 2022
Peer-reviewedText generation uses a median 0.042 kWh per 1,000 inferences, about the same as charging two smartphones.
Luccioni et al. 2024, cited in Carbon in Motion (ACM), 2024
Single-sourceThe median Gemini Apps text prompt uses 0.24 Wh, less energy than watching nine seconds of television, and far less than many public estimates.
Elsworth et al. (Google), 2025
Company-reportedThe same median prompt consumes the equivalent of five drops of water, 0.26 mL.
Elsworth et al. (Google), 2025
Company-reportedIf data centres cut water-use intensity 10% a year, the new water capacity they need through 2030 falls to 227–604 million gallons a day, from 697–1,451.
Han, Li, Wierman & Ren, 2026
Single-sourceNo verified figure exists for the emissions AI tools help avoid elsewhere in the economy. Claims of net benefit rest on modelled scenarios, not measured outcomes.
No verified figure exists for how much clean power bought for AI is new generation rather than existing supply relabelled. Contracts are public; the counterfactual is not.
Old benchmarks, new water forecasts, images that cost more than text and hardware that outweighs its use.
Training one large Transformer with neural architecture search was estimated at 626,155 lbs of CO₂e, about five times a car's lifetime emissions including fuel.
Strubell, Ganesh & McCallum, 2019
Peer-reviewedGlobal AI demand is projected to account for 4.2–6.6 billion cubic metres of water withdrawal in 2027, more than the total annual withdrawal of 4–6 Denmarks or half the United Kingdom.
Li, Yang, Islam & Ren, 2023
Single-sourceGenerating images costs far more than generating text, and video is newer still and barely measured.
Luccioni et al. 2024, cited in Carbon in Motion (ACM), 2024; Luccioni, Jernite & Strubell, 2023
Single-sourceA 1,024-GPU cluster cooled by evaporation in the UAE would consume over 30 million litres of water a year, in a country rated at extremely high water stress.
Syed et al. (IEEE), 2026
Single-sourceThe share of life-cycle emissions from making the hardware rose from 49% for the iPhone 3GS to 86% for the iPhone 11: as use gets cleaner, manufacturing dominates.
Gupta et al., 2020
Peer-reviewedNo verified figure exists for the footprint of video generation at scale. The systems are new, and the studies so far measure small samples.
No credible figure exists for data-centre water consumption by site. Operators report aggregate withdrawal, not consumption, and not by location.
Water per chat, carbon built into the hardware, cooling, data growth and compute sited where water is scarce.
GPT-3 needs to “drink” a 500 ml bottle of water for roughly 10–50 medium-length responses, depending on when and where it runs.
Li, Yang, Islam & Ren, 2023
Single-sourceAcross Apple's products, manufacturing accounts for 74% of total emissions and hardware use 19%; the hardware life cycle is over 98%.
Gupta et al., 2020
Peer-reviewedTraditional air conditioning typically accounts for over 40% of a data centre's electricity, which is why operators turn to water-based cooling.
Carbon in Motion (ACM), 2024
Single-sourceTraining data for two Meta recommendation models grew 2.4× and 1.9× in two years, to exabyte scale, raising data-ingestion bandwidth 3.2×.
Wu et al. (Meta), 2021
Company-reportedGovernments across the Global South announced over $200 billion in sovereign AI commitments between 2024 and 2026, often in water-stressed or flood-prone places.
Syed et al. (IEEE), 2026
Single-sourceCounting the water used to generate its electricity as well as on-site cooling, training GPT-3 in Microsoft's US data centres could consume 5.4 million litres in total.
Li, Yang, Islam & Ren, 2023
Single-sourceThe market for AI was projected to grow tenfold within five years, into hundreds of billions of dollars, with infrastructure growing to match.
Wu et al. (Meta), 2021
Company-reportedBangladesh's sovereign AI plans call for centralised GPU procurement but do not say where data centres would go, in a country where more than a fifth of the land floods in an average year.
Syed et al. (IEEE), 2026
Single-sourceNew demand the size of a city's water supply, emissions few papers report, and a share of energy nobody can pin down.
If 2024 water-use intensity persists, US data centres could need 697–1,451 million gallons a day of new water capacity by 2030: comparable to New York City's whole daily supply.
Han, Li, Wierman & Ren, 2026
Single-sourceMeasured against one passenger jet's round trip between San Francisco and New York (180 tCO₂e), T5's training came to about 26%, Meena 53%, GShard-600B 2%, Switch Transformer 32% and GPT-3 about 305%.
Patterson et al. (Google), 2021
CorroboratedAsked for the training details behind 500 published models, authors supplied enough for 95 models from 77 papers: a 15.4% response rate.
Luccioni & Hernandez-Garcia, 2023
Peer-reviewedEstimates made for papers that did not publish their emissions have been off by 100× to 100,000×, which is why the authors want emissions reported in every ML paper.
Patterson et al. (Google / UC Berkeley), 2022
CorroboratedData centres used close to 1% of global energy in 2018, yet what proportion of that goes to training AI is not known.
Dodge et al. (FAccT), 2022
Peer-reviewedBoth sides of this page are drawn to the same standard on purpose. Every figure carries its source and a confidence label; where no verified figure existed, the page says so rather than estimating. Whether efficiency outruns growth is not ours to hand you.
The IEA projection is cited via You et al., and Luccioni et al.'s per-inference energy via the Carbon in Motion study, as those sources state them.
