Beneficial AI for America
AI and the environment

AI isn't the climate problem. Not building clean power is.

Data centers use about 1.5% of the world's electricity and cause about half of one percent of its energy-related CO2.1,2 That share is growing fast, and the growth is landing on a grid America has underbuilt for decades. The fix is to build clean power, nuclear and transmission at speed, and to get the arithmetic right.

In proportion

Next to the things that actually drive emissions, data centers are small

Climate debates about AI often skip the denominator. Here it is. Each bar below is one source of global emissions, drawn to scale.

Global emissions by source

Million tonnes of CO2 (or CO2-equivalent), latest year available. Hover or tap a bar for details.

  • Coal power~10,000
  • Agriculture6,240
  • Steel~3,550
  • Cement~2,400
  • Oil & gas methane leaks~2,400
  • Aviation942
  • Data centers (all, not just AI)180–286

Years and gases differ by row (2023–2025); each tooltip gives the scope. The rows also overlap, so they don't add up to a total: steel's figure and the data center row both include emissions from electricity, some of which is also counted under coal power. Sources: coal4, agriculture5, steel6, cement7, methane8, aviation9, data centers1,3.

Show as a table
SourceMt CO2(e)Year
Coal power (estimate)~10,0002025
Agriculture, farm level6,2402023
Steel (derived)~3,5502025
Cement~2,4002024
Oil & gas methane~2,4002025
Aviation (CO2 only)9422024
Data centers180–2862024–2025

The world's energy-related CO2, in 1,000 dots

Each dot is 0.1% of global energy CO2, about 38 million tonnes. Data centers fill about 5 dots on IEA figures, or about 7 on the highest independent estimate.

Data centers: 0.47% (IEA, 2024)Everything else

Math: 180 Mt ÷ 38,255 Mt = 0.47% (2024); the high estimate is 286 ÷ 38,400 = 0.74% (2025). For scale, aviation is about 25 dots, cement about 63 and steel about 93.1,2,3

Electricity: air conditioning uses six times more than every data center on Earth

Global electricity use, terawatt-hours per year

  • Electric motors13,462
  • Air conditioning & cooling2,900+
  • Data centers, 2030 projection~950
  • Data centers, 2025~485

Sources: motors10, cooling11, data centers1,2. In the US, home air conditioners alone used 254 TWh in 2020, more than all US data centers in 2023 (176 TWh).12,13

How many AI questions is that?

Google measured its median text prompt at 0.24 watt-hours;14 an independent estimate for ChatGPT is about 0.3.15 Pick an everyday activity to see how many prompts use the same energy.

—text prompts

Sources: streaming16, car17, image and video18. Google's figure is a median it measured itself, for text only, and excludes training. Long inputs, reasoning, images and video use much more.

If you drive one mile less, you save the energy of about 6,000 AI questions. The climate fight is won or lost in power plants, steel mills, cement kilns and cars, not in your chat window.

The honest part: Small today does not mean small forever. Data center electricity grew 17% in 2025, and data centers made up about half of all US growth in electricity use that year.2 The IEA expects their use to roughly double by 2030.1 More than a third of the gas power capacity under development in the US is slated to power data centers directly.19 AI video and agent tasks use tens to thousands of times more energy than a text prompt.18,20 The IEA estimates that, if widely adopted, existing AI tools could cut more emissions than data centers produce, but warns that this isn't guaranteed and that adoption is slow so far.1 Whether AI's growth is clean depends on what we build, which is the rest of this page.
The water math

The scariest AI water numbers came from broken arithmetic

Three of the most-repeated claims about AI and water were off by roughly 10 to 1,000 times. In each case one term in the equation was wrong or out of date: a unit, an energy estimate, or a word. Here are the equations side by side. Hover or tap any line to highlight the same term in both versions. Shaded lines are the ones that changed.

1. "More than a thousand times a town's water"

Empire of AI (2025) compared a planned Google data center near Santiago, Chile, with the water used by the 88,000 residents of Cerrillos.21,22

As printed

Hardback, pp. 288–289 and endnote

  1. Data center cooling water (Google's filing)169 L per second
  2. ×Seconds in a year31,536,000
  3. =Data center, per year5,329,584,000 L
  4. ÷Cerrillos residential use, 20195,097,946 LThe source table said "liters," but the numbers were cubic meters.
  5. =Data center vs. town1,045×
Data center5.33 bn L
Town0.005 bn L

Corrected

Author's own correction, Dec. 2025

  1. Data center cooling water (Google's filing)169 L per second
  2. ×Seconds in a year31,536,000
  3. =Data center, per year5,329,584,000 L
  4. ÷Cerrillos residential use, 20195,097,946,720 L= 5,097,946.72 cubic meters × 1,000
  5. =Data center vs. town1.05×
Data center5.33 bn L
Town5.10 bn L
Changed termBars use the same scale in both versions.

A check anyone can do: divide the town's water by its people and days. The printed figure works out to 0.16 liters per person per day, less than a cup. The corrected figure is 159 liters, a normal amount for a city.

Analyst Andy Masley found the error in November 2025. Karen Hao confirmed it, published a correction and credited him. Her corrected line says the data center "could use more water than the entire population of Cerrillos."21,22 That is how corrections should work. The trouble is that "a thousand times" had already gone around the world.

2. "A bottle of water for every AI email"

A widely shared 2024 Washington Post and UC Riverside estimate: a 100-word email written with GPT-4 uses 519 mL of water.24 We rebuilt the math from the researchers' published water factors, then swapped in a measured energy figure.23

As published (reconstructed)

2024 estimate, US-average water factors

  1. Server energy per task0.123 kWhAn assumption about GPT-4 energy use; no per-prompt figures had been published yet.
  2. ×Water per kWh, on site plus at the power plant≈ 4.2 L per kWh
  3. =Water per email519 mL
Per email519 mL

With measured energy

Google's measured median prompt, 2025

  1. Energy per prompt0.00024 kWh0.24 Wh, about 500 times lower
  2. ×Water per kWh, on site plus at the power plant≈ 4.2 L per kWh
  3. =Water per prompt≈ 1 mL
Per prompt≈ 1 mL
Changed termBars use the same scale in both versions.

This is not a like-for-like correction: these are different models and years, Google's number is self-reported and an email can take more than one prompt. What it shows is that newer measured energy figures suggest far lower water per prompt. The 2024 estimate relied on an energy guess made before any per-prompt figures were published. About 1 mL is a fifth of a teaspoon.14,23

3. "AI will consume half of Britain's water"

A 2027 projection for global AI from a UC Riverside paper, repeated in Empire of AI as water "consumed."22,23

As repeated

Book's original wording

  1. AI electricity, 2027 (projected)94–147 TWh
  2. ×Water withdrawn per kWh≈ 45 L per kWhMostly power-plant cooling water that is returned to the river
  3. =Called "consumed"1.1–1.7 trillion gal
Water1.1–1.7 tn gal

Water actually consumed

Same paper, consumption figures

  1. AI electricity, 2027 (projected)94–147 TWh
  2. ×Water consumed per kWh≈ 4.1 L per kWhWater that evaporates and is not returned
  3. =Consumed100–158 billion gal
Water0.10–0.16 tn gal
Changed termBars use the same scale in both versions.

The paper itself correctly called the big number "withdrawal." The error came in the retelling, and the author has since changed "consume" to "use or withdraw."22 About 98% of the withdrawal figure is water that passes through power plants and goes back to rivers.23

We hold our own side to the same rule

Google reports 0.26 mL of water per prompt. That counts only water used at the data center, not at the power plant.14 Counted the same way as the critics' numbers, it is about 1 mL. That is the figure we use. A pro-AI case built on flattering accounting would deserve the same correction.

America's water, in 1,000 dots

Each dot is 0.1% of the water consumed each day by US farm irrigation and power plants. US data centers consume less than one dot on site.

Data centers, on site: 0.06%Irrigation and power plants

Math: US data centers consumed 17.4 billion gallons on site in 2023, or 0.048 billion gallons a day.13 Irrigation (73.2) and power plants (4.31) consumed 77.5 billion gallons a day in 2015, the latest national survey.25 0.048 ÷ 77.5 = 0.06%. Adding the water used to make data centers' electricity raises the share to under 1%, with some double counting. The years differ, and we say so.

Golf courses vs. data centers

US water, billion gallons per year

  • Golf courses531
  • Data centers, including power plants~228
  • Data centers, on site17.4

Golf is water applied; data centers are water consumed. Sources: golf26, data centers13.

The honest part: Water is a local question, and some local worries are real. Even corrected, Google's permitted figure for its Chilean data center roughly equaled the town's household water use, during a long drought, using drinking water.21,22 Local pressure worked: a 2024 court ruling sent the project back for review, and Google has moved toward air cooling.27 That's the right model. Use air or closed-loop cooling and reclaimed water in dry places, and publish expected water use before approval, counted like for like.
How we got here

America stopped building always-on clean power and the wires to carry it. AI is the bill arriving.

America has kept adding solar and batteries. In 2025 the grid added 53 gigawatts of new generating capacity, the most in any year since 2002, and about four-fifths of it was solar and batteries.28 What it largely stopped building was nuclear plants and high-voltage power lines, and connecting any new project now takes years.

The strain between AI and the grid is not mainly an AI problem. It is what decades of underbuilding look like when demand finally grows again.

0

New nuclear construction permits issued in the US every year from 1980 through 2011.29

259 → 132

US reactors ordered vs. reactors that ever got a full-power license. More than 100 planned reactors were cancelled.29

13

Working reactors, about 10 gigawatts of always-on clean power, closed from 2013 to 2022.30

Nuclear still makes about 18% of America's electricity and about two-fifths of its clean electricity.31 Imagine where we'd be if we had kept building. When reactors close, the gap is not filled by sunshine. It is filled by gas or coal:

  • California, 2012When the San Onofre plant shut down, natural gas filled the gap and CO2 rose by 9 million tons in the first year.32
  • Germany, 2011–2019After Fukushima, Germany shut 11 of its 17 reactors between 2011 and 2019. A peer-reviewed study of those years found the lost power was replaced mostly by coal and imports, at an estimated social cost of €3–8 billion a year, mostly from deaths caused by air pollution.33

We stopped building the wires, too

New high-voltage transmission lines (345 kV and up), miles per year

  • Needed each year~5,000
  • 2013~4,000
  • 2010–14 average~1,700
  • 2024888
  • 2020–23 average~350

Source: ACEG and Grid Strategies, 2025, including the FERC-data addendum.34

And we made new power wait in line

Median time from asking to connect to the grid to starting operation, in months

  • 200822 months
  • 201536 months
  • 202561 months

About 2,061 gigawatts of proposed power plants and batteries were waiting to connect at the end of 2025, about 85% of it solar, storage and wind. Only 13% of the capacity that asked to connect from 2000 to 2020 is running today. Many requests are speculative, but the wait has nearly tripled.35 Local governments in 48 states have also adopted 888 restrictions on wind and solar.36

A short history of not building

  1. 1979
    Three Mile Island accident. Only 2 reactor construction permits are issued that year, and none for the next 32 years.29
  2. 2001–2017
    Cape Wind, America's first proposed offshore wind farm, is fought for 16 years and abandoned.37
  3. 2013–2022
    13 reactors close, mostly because markets paid nothing for clean, always-on power.30
  4. 2017
    V.C. Summer's two new reactors in South Carolina are abandoned after about $9 billion is spent.38
  5. 2019
    New Hampshire's Supreme Court upholds the denial of Northern Pass, a line to bring Canadian hydropower to New England.39
  6. 2023–2024
    Vogtle 3 and 4 in Georgia come online, the first new-build reactors in decades: years late, with a reported cost above $30 billion against an original estimate of about $14 billion. After a 30-year gap, we had to relearn how to build.40
  7. 2024
    The ADVANCE Act to speed up nuclear licensing passes the Senate 88–2 and the House 393–13.41
  8. 2026
    The NRC approves TerraPower's advanced reactor in 18 months.42 Two Republicans and two Democrats introduce a Senate permitting reform bill.43
The honest part: No single villain did this. Reactor cancellations began before Three Mile Island, as demand slowed and interest rates and costs soared. Most recent closures were owners' business decisions under market rules that didn't value clean, firm power.29,30 Plants and power lines were blocked in blue states and red counties alike. Utilities, regulators, Congress, states and local opponents of every stripe all played a part. That's why the fix can belong to everyone.
What to build now

Build clean, build firm, build fast

AI demand is the strongest reason in a generation to fix how America builds energy. If we don't build clean power, gas will fill the gap. The IEA expects gas and coal to meet over 40% of new data center demand to 2030.1 This is a choice, and we should make it on purpose.

  • Pass permitting reform this CongressA bipartisan Senate bill from Sens. Lee, Capito, Heinrich and Whitehouse would speed environmental reviews and federal transmission permits. It would also make large computing loads pay for the transmission built to serve them.43 Every year a line is delayed costs consumers an estimated 15–37% of its benefits.44
  • Build transmission at five times today's paceWe need about 5,000 miles of high-voltage lines a year. We built 888 in 2024.34
  • Keep every safe reactor running and build new onesFollow the ADVANCE Act with steady orders, not one-off megaprojects, so costs fall as crews and suppliers learn. The ten states that still restrict new nuclear should lift those limits, as states led by both parties have already done.41,45
  • Clear the interconnection lineFive years in line is too long. Screen out speculative requests and study projects in clusters, so real projects connect in months, not years.35
  • Make data centers bring new clean powerBig buyers should pay for new clean, firm supply (nuclear, geothermal, storage, transmission) rather than buying up existing plants' output, which leaves everyone else short. See our asks for industry.
  • Publish the numbers, counted honestlyEvery large data center should report its energy and water use, with water counted like for like (on-site vs. power plant, withdrawn vs. consumed), before approval and every year after.
The honest part: Much of the new clean firm power that tech companies have contracted for is years away. TerraPower's first advanced reactor, for example, is targeted for 2030. Until it arrives, rising demand will be met partly with gas.1,19,42 That is exactly why the building has to start now.
Ask your members of Congress to build Read the full case for AI
Sources

Every number, sourced

Research and drafting were assisted by AI, and a person reviewed every claim. Our full research notes list a confidence level for each figure and the claims we chose not to make. Spot an error? Send a correction.

  1. International Energy Agency, "Energy and AI," April 2025 (data centers: 415 TWh and about 180 Mt CO2 in 2024; 2030 projection; gas and coal share of new demand; AI emissions-reduction potential). iea.org/reports/energy-and-ai/executive-summary
  2. International Energy Agency, "Global Energy Review 2026," April 2026 (global energy CO2; data center growth of 17% in 2025; half of US demand growth). iea.org/reports/global-energy-review-2026/co2-emissions
  3. AFP via The Star, "Study: Data centres emitting more CO2 than thought" (Allianz Trade estimate, 286 Mt in 2025), July 1, 2026. thestar.com.my/tech/tech-news/2026/07/01/study-data-centres-emitting-more-co2-than-thought
  4. Ember, "Global Electricity Review 2026" (coal generation, 2025). CO2 figure is our estimate at about 1 kg CO2/kWh. ember-energy.org/latest-insights/global-electricity-review-2026/2025-in-review
  5. Climate Watch, processed by Our World in Data, "Greenhouse gas emissions by sector," 2026. ourworldindata.org/grapher/ghg-emissions-by-sector
  6. worldsteel, "Sustainability Indicators 2025" (1.92 t CO2 per tonne) and 2025 crude steel production release (1,849.4 Mt). worldsteel.org/wp-content/uploads/Sustainability-Indicators-2025.pdf
  7. R. Andrew (CICERO), "Global CO2 emissions from cement production," Zenodo, October 2025. zenodo.org/records/17287474
  8. International Energy Agency, "Global Methane Tracker 2026," key findings. CO2e conversion is ours. iea.org/reports/global-methane-tracker-2026/key-findings
  9. IATA, "2024 Aviation Emissions: Efficiency Gains vs. Rising Totals," November 28, 2025. iata.org/en/iata-repository/publications/economic-reports/2024-aviation-emissions-efficiency-gains-vs.-rising-totals
  10. IEA 4E Electric Motor Systems Annex, policy brief, December 2025. iea-4e.org/wp-content/uploads/2025/12/4E-Policy-Brief-EMSA-9-20251218.pdf
  11. HVAC&R News, reporting the IEA's 2026 cooling briefing, July 21, 2026. hvacrnews.com.au/news/global-cooling-market-on-the-rise
  12. U.S. Energy Information Administration, "How much electricity is used for air conditioning in the United States?" (RECS 2020). eia.gov/tools/faqs/faq.php?id=1174&t=1
  13. Shehabi A. et al., Lawrence Berkeley National Laboratory, "2024 United States Data Center Energy Usage Report," December 2024 (176 TWh in 2023; water, pp. 55–57). eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
  14. Elsworth C. et al. (Google), "Measuring the environmental impact of delivering AI at Google Scale," arXiv:2508.15734, August 2025. arxiv.org/abs/2508.15734
  15. Josh You, Epoch AI, "How much energy does ChatGPT use?," February 2025. epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
  16. George Kamiya, Carbon Brief, "Factcheck: What is the carbon footprint of streaming video on Netflix?," February 25, 2020. carbonbrief.org/factcheck-what-is-the-carbon-footprint-of-streaming-video-on-netflix
  17. U.S. EPA, "Greenhouse Gas Emissions from a Typical Passenger Vehicle," updated June 2026. Energy per mile is our conversion at 33.7 kWh per gallon. epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle
  18. O'Donnell J. and Crownhart C., MIT Technology Review, "We did the math on AI's energy footprint," May 20, 2025. technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech
  19. Global Energy Monitor, "Betting big on data centers, U.S. now leads world for new gas power development," January 2026. globalenergymonitor.org/press-release/betting-big-on-data-centers-u-s-now-leads-world-for-new-gas-power-development
  20. IEA, "Key questions on energy and AI," presentation at the ADB Asia Clean Energy Forum, June 11, 2026. asiacleanenergyforum.adb.org/wp-content/uploads/2026/06/Key-questions-on-energy-and-AI.pdf
  21. Andy Masley, "Empire of AI is wildly misleading about AI water use," November 16, 2025 (with updates through December 18, 2025). blog.andymasley.com/p/empire-of-ai-is-wildly-misleading
  22. Karen Hao, "Changes to Empire of AI water figures," December 17, 2025. karendhao.com/20251217/empire-water-changes
  23. Li P., Yang J., Islam M.A. and Ren S., "Making AI Less 'Thirsty'," arXiv:2304.03271 v5, March 2025 (Table 1 and 2027 appendix). Our reconstruction of the 519 mL figure uses its US-average factors. arxiv.org/abs/2304.03271
  24. Fortune, reporting the Washington Post and UC Riverside estimate, September 23, 2024. fortune.com/2024/09/23/ai-water-usage-droughts-chatgpt-microsoft
  25. Dieter C.A. et al., U.S. Geological Survey, "Estimated Use of Water in the United States in 2015," Circular 1441, 2018. pubs.usgs.gov/circ/1441/circ1441.pdf
  26. GCSAA, "Golf courses reduce water usage by 31 percent, according to national survey," December 30, 2025. gcsaa.org/who-we-are/media/news/2025/12/30/golf-courses-reduce-water-usage-by-31-percent-according-to-national-survey
  27. Semafor, "Chile Google data center gets permit partially revoked," February 28, 2024. semafor.com/article/02/28/2024/chile-google-data-center-gets-permit-partially-revoked
  28. U.S. Energy Information Administration, "New U.S. electric generating capacity expected to reach a record high in 2026," Today in Energy, February 20, 2026 (2025 additions: 53 GW total, 27.2 GW utility-scale solar, 15 GW batteries; the solar-and-battery share is our arithmetic). eia.gov/todayinenergy/detail.php?id=67205
  29. U.S. Energy Information Administration, "Annual Energy Review 2011," Table 9.1, "Nuclear Generating Units, 1955–2011." eia.gov/totalenergy/data/annual/pdf/sec9_3.pdf
  30. Congressional Research Service, "U.S. Nuclear Plant Shutdowns, State Interventions, and Policy Concerns," R46820 (12 closures through 2021; Palisades closed in 2022). congress.gov/crs-product/R46820
  31. U.S. Energy Information Administration, "Electricity in the U.S." (2025 generation shares; nuclear's share of clean electricity is our arithmetic). eia.gov/energyexplained/electricity/electricity-in-the-us.php
  32. Davis L. and Hausman C., "Market Impacts of a Nuclear Power Plant Closure," American Economic Journal: Applied Economics 8(2), 2016. escholarship.org/content/qt0nz9b511/qt0nz9b511.pdf
  33. Jarvis S., Deschenes O. and Jha A., "The Private and External Costs of Germany's Nuclear Phase-Out," Journal of the European Economic Association 20(3), 2022 (sample 2010–2019, covering the closures of 2011, 2015, 2017 and 2019; social cost of €3–8 billion a year across scenarios). econpapers.repec.org/RePEc:ehl:lserod:113634
  34. ACEG and Grid Strategies, "Fewer New Miles: Strategic Industries Held Back by Slow Pace of Transmission," July 2025, with addendum. gridstrategiesllc.com/wp-content/uploads/ACEG_Grid-Strategies_Fewer-New-Miles-2025_vF.pdf
  35. Rand J. et al., Lawrence Berkeley National Laboratory, "Queued Up: 2026 Edition," June 2026. eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf
  36. pv magazine USA, reporting Columbia Sabin Center, "Opposition to Renewable Energy Facilities in the United States," September 8, 2026. pv-magazine-usa.com/2026/09/08/the-u-s-saw-a-14-percent-increase-in-contested-renewable-projects-9-8-gw-in-cancellations-in-2025
  37. Bureau of Ocean Energy Management, "Cape Wind." boem.gov/Massachusetts-Cape-Wind
  38. American Nuclear Society, "SCANA, Santee Cooper halt V.C. Summer nuclear plant construction," 2017. ans.org/news/article-1973/scana-santee-cooper-halt-v-c-summer-nuclear-plant-construction
  39. WBUR, "Northern Pass halted," July 26, 2019. wbur.org/news/2019/07/26/northern-pass-halted
  40. American Nuclear Society, "Vogtle project update: Cost likely to top $30 billion," and "Vogtle 4 enters commercial operation," 2024. ans.org/news/article-5994/vogtle4-enters-commercial-operation
  41. U.S. Senate Environment and Public Works Committee, "Senate sends Capito-Carper-Whitehouse nuclear energy bill to President's desk," June 2024. epw.senate.gov/public/index.cfm/2024/6/senate-sends-capito-carper-whitehouse-nuclear-energy-bill-to-president-s-desk
  42. American Nuclear Society, "NRC approves TerraPower construction permit," March 4, 2026. ans.org/news/2026-03-04/article-7818/nrc-approves-terrapower-construction-permit
  43. Senate Energy and Natural Resources Committee, "Lee, Capito, Heinrich and Whitehouse introduce Bipartisan American Affordability and Jobs Act of 2026," September 30, 2026. energy.senate.gov/press/rep/lee-capito-heinrich-and-whitehouse-introduce-bipartisan-american-affordability-and-jobs-act-of-2026
  44. Zimmerman, Gramlich and Goggin (Grid Strategies) for WIRES, a transmission industry group, "Delaying Transmission Increases Costs and Reduces Benefits for Consumers," November 2025. gridstrategiesllc.com/wp-content/uploads/GS_WIRES-Cost-of-Delayed-Transmission.pdf
  45. National Conference of State Legislatures, "States restrictions on new nuclear power facility construction," updated April 16, 2026. ncsl.org/research/environment-and-natural-resources/states-restrictions-on-new-nuclear-power-facility.aspx