# AI and Carbon
> [!abstract] What this note is
> An essay on where AI sits in the physical world: the post-industrial story, the energy it runs on, the waste it leaves, and the shared stories it carries. Figures are from single-pass research checked 2026-11-02; most come from the companies themselves or from press coverage of reports that were not opened.
## The post-industrial story
In 1973 the sociologist Daniel Bell described a coming post-industrial society, organised around information and services rather than factories [1]. The same year, the oil embargo showed how much the industrial world still ran on energy. Star Wars came out four years later.
The post-industrial idea stuck, and so did the feeling that information has no weight. It always had weight. Email, search, streaming and now AI run on power stations, mines, chip factories, cooling water and concrete. AI is the information economy at its most energy-hungry, which makes the weight hard to ignore.
## Energy
Data centres used about 485 TWh of electricity worldwide in 2025 [2]. A terawatt-hour is a billion kilowatt-hours, the unit on a home electric bill. An average American home uses about 10,800 kWh a year [3], so 485 TWh would run about 45 million American homes for a year. There are about 135 million households in the US [4], so that is a third of them.
Worldwide, it is still a small share: data centres used about 1.5% of the world's electricity in 2024, and AI on its own about 0.5% in 2025, by one estimate [5]. The growth is what stands out: up 17% in 2025, more than five times as fast as electricity use overall. The IEA expects data centres to roughly double their use by 2030, to about 3% of the world's electricity and more than all the electricity Japan uses today [6], with the part that goes to AI tripling [2]. In the US, about one kilowatt-hour in every 23 went to a data centre in 2023. By 2028, Berkeley Lab projects between one in 15 and one in 8 [7].
For scale on the other side: about 730 million people, most of them in sub-Saharan Africa, had no electricity at home at all in 2024 [8].
Data-centre demand is arriving faster than clean power can be built. New gas plants are going up, some on data-centre sites. At least 15 US coal plants have had their retirements delayed since January 2025, with data-centre demand one of the stated reasons [9]; only about 30% of the coal retirements planned for 2025 happened [10].
The companies' own reporting softens this. Most count renewable-energy certificates against their electricity use ("market-based" accounting). Counted against the grid their data centres actually draw from ("location-based"), a 2024 Guardian analysis found the big companies' own data-centre emissions were about 7.6 times what they reported [11]. Even on their own terms the numbers are rising: Microsoft's emissions went up about 25% in fiscal 2025 [12], and Google reported record emissions, electricity and water use for 2025 [13].
One prompt is small. Google puts a median Gemini text prompt at 0.24 Wh [14]; OpenAI's chief executive has said about 0.34 Wh for ChatGPT, with no method published [15]. Mistral's life-cycle figure, which includes making the hardware, is 1.14 g of CO2e per response [16]. Images cost far more than text [17], and video more again [18]. The problem is not the single prompt. It is the scale and the build-out.
## Waste
AI chips are replaced every few years as faster ones arrive. The world produced a record 62 million tonnes of electronic waste in 2022, and less than a quarter of it was documented as collected and recycled; the UN expects 82 million tonnes a year by 2030 [19]. One study estimates generative AI alone could add 1.2 to 5 million tonnes between 2020 and 2030, and that reusing and recycling the hardware could cut that by 16% to 86% [20].
Data centres also draw water for cooling. All of it lands in a world already dealing with heat, water stress and every other waste stream. It is a growing landfill among many other problems.
## Intermission: frozen in Cloud City
In *The Empire Strikes Back* (1980), Han Solo is frozen in carbonite in Cloud City, a mining colony floating in the clouds of a gas planet. The chamber that freezes him was built for industry, not for people.
*Han Solo in carbonite, the prop from* Robot Chicken: Star Wars. *Photo: Star Wars / Bonnie Burton, [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/), via [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Han_Solo_in_carbonit.jpg).*
It is one of the best-known images in film. People who have never seen the movie recognise it. That is shared meaning: a story carried by so many people that one picture is enough to call it up.
Today "the cloud" is an industrial site too.
## Shared stories have a body
Yuval Noah Harari's *Nexus* is a history of information networks from the Stone Age to AI [21]. Its argument is that information's main job has been to connect people, through shared stories such as religions, money and nations, rather than to tell the truth, and that AI is the first technology in that history that can make decisions and generate ideas by itself.
It is a book about human history: stories, networks and power. The physical cost of the networks is not its subject. That is the gap this note sits in. Every network in that history had a body, in clay, paper, presses and wire, but the body was small next to the stories it carried. The body of the AI network is measured in power stations and tonnes of waste. When the network that carries our shared stories needs its own power plants, its physical cost becomes part of the story.
The carbonite is fiction. The carbon in the machines that keep stories like it in circulation is not.
## Deep time
There is a reassuring way to look at all this. Earth has been through catastrophe before: five mass extinctions, including the asteroid that ended the dinosaurs 66 million years ago. Each time, life recovered and something new grew into the gap. Mammals, and eventually people, came out of the last one. By that measure the planet is not doomed.
That view is right about the planet and wrong about the timescale. After the worst extinction, at the end of the Permian about 252 million years ago, ecosystems took around 10 million years to fully recover [22]. Regrowth on that scale is no use to anyone living through the disruption. The change now is also fast: global temperature has risen faster since 1970 than in any 50-year period in at least 2,000 years, and there is more carbon dioxide in the air than at any time in at least 2 million years [23]. And past recoveries never had to keep eight billion people fed, housed and powered along the way.
So both things are true. The planet recovers on its own clock. People live on theirs, and the choices made now decide how hard the coming century is, and for whom.
## What one person can change
- **Match the tool to the task.** A small model is enough for most short questions. Save the largest models, long reasoning modes and agents for work that needs them.
- **Generate images and video on purpose.** They are the biggest cost per request.
- **Don't leave agents looping.** An agent that retries for an hour is thousands of requests.
- **Read "100% renewable" for the method.** It is usually market-based. Look for the location-based number.
- **Vote, and show up locally.** Where data centres are built, what powers them and who pays for the grid are decided by people on the ballot: state legislators, governors and, in some states, the utility regulators themselves. In the US, the midterm elections are on 3 November 2026. Zoning boards and utility rate hearings take public comment too.
## Learning more
Free to read, no account needed:
- **[Our World in Data: Energy](https://ourworldindata.org/energy)** — charts of electricity, emissions and energy use for every country, free to reuse.
- **[Sustainable Energy – Without the Hot Air](https://www.withouthotair.com/)** — David MacKay's book on making energy numbers mean something, free online.
- **[Carbon Brief](https://www.carbonbrief.org/)** — explainers and fact-checks on climate and energy.
- **[Energy and AI](https://www.iea.org/reports/energy-and-ai)** — the IEA's report; the executive summary is readable on its own.
- **[En-ROADS](https://en-roads.climateinteractive.org/)** — a climate simulator from Climate Interactive and MIT Sloan. Move the levers and watch the warming curve change.
- **[Data Centers in Virginia](https://jlarc.virginia.gov/landing-2024-data-centers-in-virginia.asp)** — a 2024 state report on what data centres mean for one place: power demand, grid costs and the neighbours. Northern Virginia is the largest data-centre market in the world.
At a library:
- Vaclav Smil, *How the World Really Works* (2022) — the energy and materials under the "post-industrial" economy
- Kate Crawford, *Atlas of AI* (2021) — the physical side of AI: mining, energy and labour
- Peter Brannen, *The Ends of the World* (2017) — the five mass extinctions
## Sources
1. Daniel Bell, *The Coming of Post-Industrial Society* (Basic Books, 1973)
2. [Key Questions on Energy and AI — IEA, April 2026](https://www.iea.org/reports/key-questions-on-energy-and-ai) *(figures via press coverage; the report was not opened)*
3. [How much electricity does an American home use? — US Energy Information Administration](https://www.eia.gov/tools/faqs/faq.php?id=97&t=3) *(2022 figure: 10,791 kWh)*
4. [Total households (TTLHH) — US Census Bureau via FRED](https://fred.stlouisfed.org/series/TTLHH) *(about 134.8 million in 2025)*
5. [How much electricity does AI consume? 2025 summary — Hannah Ritchie](https://hannahritchie.substack.com/p/ai-electricity-2025)
6. [AI data centres to consume more power than Japan by 2030, says IEA — AFP via Malay Mail, 11 Apr 2025](https://www.malaymail.com/news/tech-gadgets/2025/04/11/ai-data-centres-to-consume-more-power-than-japan-by-2030-says-iea/172557)
7. [US data-center power use could nearly triple by 2028, DOE-backed report says — Reuters via The Star, Dec 2024](https://www.thestar.com.my/tech/tech-news/2024/12/20/exclusive-us-data-center-power-use-could-nearly-triple-by-2028-doe-backed-report-says), on Lawrence Berkeley National Laboratory's report
8. [Access to electricity stagnates, leaving 730 million in the dark — Climate and Clean Air Coalition, on the IEA's figures](https://www.ccacoalition.org/resources/access-electricity-stagnates-leaving-globally-730-million-dark)
9. [These 15 coal plants would have retired. Then came AI and Trump — DeSmog, 12 Dec 2025](https://www.desmog.com/2025/12/12/15-coal-plant-retirements-delayed-ai-data-centers-trump-doe-orders/)
10. [US coal-fired power plant retirements slowed in 2025 — Industrial Info](https://www.industrialinfo.com/iirenergy/industry-news/article/us-coal-fired-power-plant-retirements-slowed-in-2025-future-is-uncertain--357093), on EIA figures
11. [Big Tech's datacentre emissions much higher than officially reported — Computing, on the Guardian's 2024 analysis](https://computing.co.uk/news/2024/cloud/big-techs-datacentre-emissions-officially-reported)
12. [2026 Environmental Data Fact Sheet — Microsoft](https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2026-Microsoft-Environmental-Data-Fact-Sheet-PDF.pdf)
13. [Google's AI boom sends emissions, power use soaring — Axios, 30 June 2026](https://www.axios.com/2026/06/30/google-ai-boom-emissions-power-use)
14. [Measuring the environmental impact of delivering AI at Google scale — Google, arXiv 2508.15734, Aug 2025](https://arxiv.org/abs/2508.15734)
15. [Notes on Sam Altman's "The Gentle Singularity" — Simon Willison, 10 June 2025](https://simonwillison.net/2025/Jun/10/sam-altman/)
16. [Our contribution to a global environmental standard for AI — Mistral AI, July 2025](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai)
17. [Making an image with generative AI uses as much energy as charging your phone — MIT Technology Review, Dec 2023](https://www.technologyreview.com/2023/12/01/1084189/making-an-image-with-generative-ai-uses-as-much-energy-as-charging-your-phone/)
18. [Video Killed the Energy Budget — Delavande, Pierrard and Luccioni, arXiv 2509.19222, Sept 2025](https://arxiv.org/abs/2509.19222)
19. [Global E-waste Monitor 2024 — UNITAR and ITU, March 2024](https://unitar.org/about/news-stories/press/global-e-waste-monitor-2024-electronic-waste-rising-five-times-faster-documented-e-waste-recycling)
20. [AI will add to the e-waste problem — MIT Technology Review, 28 Oct 2024](https://www.technologyreview.com/2024/10/28/1106316/ai-e-waste), on Wang et al., *Nature Computational Science*
21. Yuval Noah Harari, *Nexus: A Brief History of Information Networks from the Stone Age to AI* (Random House, 2024)
22. [Earth needs 10 million years to recover from mass extinction — University of Bristol, May 2012](https://www.bristol.ac.uk/news/2012/8535.html), on Chen and Benton, *Nature Geoscience*
23. [Climate Change 2021: Summary for Policymakers — IPCC, AR6 Working Group I](https://www.ipcc.ch/report/ar6/wg1/chapter/summary-for-policymakers/)
## Related
- [[When You Need a Human]] — what AI can't stand in for
- [[AI Security for Users]] — the risks section
---
*Drafted with AI assistance from single-pass research; sources were cited by the tool and not opened by a person. Errors: see [[Corrections]].*