Introduction
Artificial intelligence's environmental footprint centers on two resources: electricity and water. Training and running AI models takes enormous computing power, and that computing power runs in data centers that draw electricity from the grid and water for cooling. In 2024, data centers worldwide consumed roughly 415 terawatt-hours of electricity — about 1.5% of global electricity use — a figure projected to more than double to around 945 TWh by 2030, which would exceed Japan's entire current electricity consumption [1]. U.S. data centers alone used 183 TWh in 2024, over 4% of the country's total electricity, with growth increasingly driven by AI specifically [2]. At the same time, AI has real potential to help address environmental problems: modeling climate systems, optimizing power grids, and finding efficiency gains humans might miss. Understanding AI's environmental impact means holding both realities at once — what AI consumes, and what it might help us conserve. This page lays out the current debate, points to research, directions worth watching, and gives you tools to work through it yourself.
The Debates
What concerns researchers most isn't the current 1.5% global share, but the growth rate: data center electricity use has climbed about 12% a year since 2017, four times faster than overall electricity demand [1]. That growth is already reshaping local grids — data centers draw an estimated 26% of Virginia's electricity and over 20% of Ireland's [1][2][3]. AI's advocates counter that this same computing capacity is what makes advanced climate modeling, materials research, and grid optimization possible at a scale people can't match by hand.
U.S. data centers used an estimated 228 billion gallons of water in 2023 [5]. Individual claims about water use per AI query vary enormously and haven't held up well under scrutiny. A widely cited 2023 estimate suggested 20 to 50 ChatGPT queries consumed about half a liter of water [4]; more recent measurements put a single prompt closer to a fraction of a milliliter up to about 17 milliliters [5]. When OpenAI's CEO claimed 38,000 ChatGPT queries used the same water as growing one almond, fact-checkers found the real ratio was closer to 1,000–10,000 queries per almond — a real cost, but off by a factor of roughly 4 to 40 [5]. That gap is itself a lesson: precise-sounding numbers about AI's impact often rest on shakier methodology than they let on.
Beyond electricity and water at the point of use, AI's footprint includes chip and server manufacturing, the emissions profile of whatever power source runs a given data center (coal, solar, and nuclear vary enormously by region), and hardware disposal. This full picture is harder to measure and rarely shows up in headline statistics, which tend to focus on a single query or training run.
Emerging Approaches
Researchers, companies, and policymakers are pursuing several strategies to reduce AI's footprint — each at a different stage of maturity, with its own tradeoffs.
Not every task needs the largest available model. Model distillation trains smaller models to approximate larger ones' performance, and recent research suggests that routing routine queries to smaller, specialized models instead of general-purpose ones can cut energy use substantially with little quality loss [6][7]. This is already happening commercially, though it depends on companies choosing efficiency over raw capability.
Neuromorphic computing designs chips that process information more like biological neurons, with the potential to use far less energy than the graphics processors AI training and inference rely on today [8]. The approach shows promise in early research and small-scale demonstrations, but it isn't running today's large AI models at commercial scale — worth watching, not yet a deployed solution.
Major tech companies have announced deals to power data centers with renewable energy and next-generation nuclear reactors, including small modular reactors (SMRs) [9]. Critics point to real near-term limits: no U.S. SMR is currently under construction, timelines that promised 5–7 years in 2020 have slipped to 2031, and even restarting recently closed nuclear plants would cover only a small fraction of data centers' 2025 electricity demand [10]. Some observers argue these announcements function more as public relations than near-term fixes — itself worth examining critically rather than taking at face value.
Critical Questions to Consider
Compare AI's electricity and water use to other everyday activities or industries — streaming video, cryptocurrency mining, almond farming, or an average U.S. household's water use (about 82 gallons per person per day [11]). What do these comparisons reveal, and what do they leave out? Does putting AI's footprint next to something familiar make the issue feel bigger or smaller — and why might that framing matter in a public debate?
Data center electricity demand is projected to more than double by 2030. Who should be responsible for managing that growth — AI companies, utility regulators, individual users, someone else? What would that responsibility actually look like?
The gap between the almond comparison and what fact-checkers found shows how fast an unverified number can spread. Find a claim about AI's environmental impact currently circulating online. Trace it to its original source. Does it hold up?
If cleaner power sources for AI take years or decades to scale, what should happen in the meantime? Is slowing AI development a realistic answer, an unrealistic one, or something in between?
Research one proposed response — model distillation, neuromorphic computing, or an alternative cooling method. How mature is it, what would it cost to deploy widely, and who bears the tradeoffs?
Curated Resources
International Energy Agency, "Energy and AI" — Advanced / Report. The most authoritative source on global data center electricity trends. [1]
Pew Research Center, "What we know about energy use at US data centers amid the AI boom" — Beginner. Concise, U.S.-focused summary. [2]
Our World in Data, "How much energy do data centers and artificial intelligence use?" — Beginner/Intermediate. Regularly updated charts and context. [3]
UC Riverside News, "AI programs consume large volumes of scarce water" — Intermediate. The original, widely cited water-per-query estimates. [4]
WRAL, fact-check of the AI-vs-almond water claim — Beginner. A model for how to check a viral statistic. [5]
PNAS, on neuromorphic computing and AI energy use — Advanced. Peer-reviewed. [8]
Bulletin of the Atomic Scientists, on nuclear power and AI data centers — Intermediate/Advanced. A skeptical counterpoint to industry announcements. [10]
Note: this is a fast-moving topic. Figures here reflect public reporting as of September 2026 — check publication dates before citing these numbers elsewhere.
Instructor Resources — For Faculty
The following are potential activities to help faculty develop a lesson around this topic.
Give students a well-known AI-environment claim (the almond comparison, "one query = one bottle of water"). In small groups, have them trace it to its original source, identify the assumptions behind the number, and rate it fair, exaggerated, or missing context. Groups report a one-sentence verdict with reasoning.
Assign each group one approach from Emerging Approaches. They research current news on it and report: what problem it solves, how mature it is, who benefits, who bears the cost. Close with a class discussion — given limited time and money, which approach deserves investment first, and why?
References
[1] International Energy Agency, "Energy and AI: Executive Summary," 2025.
[2] Pew Research Center, "What we know about energy use at US data centers amid the AI boom," Oct. 24, 2025.
[3] Our World in Data, "How much energy do data centers and artificial intelligence use?" 2025.
[4] UC Riverside News, "AI programs consume large volumes of scarce water," Apr. 28, 2023.
[5] WRAL, "Almonds or data centers: which requires more water? Fact-checking AI executive's claim about ChatGPT," Sept. 2026.
[6] arXiv, "Small is Sufficient: Reducing the World AI Energy Consumption Through Model Selection," 2025.
[7] ScienceDirect, "Green AI techniques for reducing energy consumption in AI systems."
[8] PNAS, "Can neuromorphic computing help reduce AI's high energy cost?" 2025.
[9] IAEA, "Data Centres, Artificial Intelligence and Cryptocurrencies Eye Advanced Nuclear to Meet Growing Power Needs." Oct. 2024
[10] Bulletin of the Atomic Scientists, "Data centers powered by next-gen nuclear? Don't fall for Big Tech's PR hype," Jul. 2026.
[11] U.S. EPA WaterSense, "Statistics and Facts," citing USGS, 2015.