Agent Atlas

Context · read 14 November 2026

The Footprint

One prompt is tiny. Billions of them, and the data centers behind them, are not. Here is the electricity and water next to things you already know, so you can judge for yourself where it sits between awful and awesome.

0.34 Wh

An average ChatGPT prompt, about a second of an oven (Sam Altman).

0.12%

A year of ChatGPT prompts, as a share of the electricity US homes spend on air conditioning. About 1/819. Our estimate.

4.4% → 6.7–12%

US data centers’ share of US electricity in 2023, and the range expected by 2028 (U.S. Department of Energy).

Electricity, to scale

Estimate
From one Walmart to all of America, on one ruler. Each step right is ten times more.
A year of ChatGPT prompts is the electricity of about 28,800 American homes. Each house is 1,000 of them (U.S. Energy Information Administration). Our estimate.

AI and data centers Everything elseTWh a year, log scale

Show as a table, with sources
WhatTWh a yearNote
One Walmart Supercenter0.0034One store in Covina, California, before a retrofit. (Center for Sustainable Energy)
A year of ChatGPT prompts0.31Our estimate: 2.5 billion a day at 0.34 Wh. (TechCrunch; Sam Altman)
Training Grok 4, once0.31Epoch AI’s estimate, not a lab figure. (Epoch AI)
Lighting in US homes812020. (U.S. Energy Information Administration)
Bitcoin mining, worldwide1202023, Cambridge estimate via EIA. (U.S. Energy Information Administration)
US data centers1762023, all workloads, not only AI. (Lawrence Berkeley National Laboratory)
Air conditioning in US homes2542020. (U.S. Energy Information Administration)
Data centers, worldwide4152024, about 1.5% of world electricity. (International Energy Agency)
All US electricity4,2002025. (U.S. Energy Information Administration)

A year of ChatGPT’s prompts lands near a single estimate for training one frontier model. Both are a rounding error next to US data centers as a whole, which run everything from email to video, not only AI. The real question is not one prompt but the growth: the IEA expects data centers worldwide to more than double, to about 945 TWh by 2030 (International Energy Agency).

Where that electricity comes from

A data center draws from the same grid as your house. In 2025 the US grid ran on this mix, about 58% fossil and 24% renewable (U.S. Energy Information Administration):

  • 41% Natural gas
  • 18% Nuclear
  • 17% Coal
  • 11% Wind
  • 7% Solar
  • 6% Hydro
About, rounded by EIA; petroleum, biomass and the rest make up the small remainder (U.S. Energy Information Administration).

Where the water goes

Factual
Most of it evaporates from cooling towers. Some is reused. More is used at the power plant than in the data center.

Withdrawn is not the same as used up

Hydrologists separate two things. A withdrawal is water taken from a river, lake or aquifer.Consumption is the part that evaporates or otherwise does not come back (U.S. Geological Survey). US power plants, for example, withdraw huge volumes for cooling but consume only about 3.9% of it; the rest goes back, warmer (U.S. Geological Survey).

What a data center does with it

  1. Evaporates most of it. A cooling tower sheds heat by evaporating water. About 80% of what it takes in leaves as vapor (Li, Yang, Islam, Ren (UC Riverside, UT Arlington; arXiv preprint); Google (arXiv preprint)). It is not destroyed: it returns to the sky as rain somewhere else, but it leaves the local river.
  2. Reuses the rest a few times. The water that does not evaporate cycles 3 to 10 times until minerals build up (Li, Yang, Islam, Ren (UC Riverside, UT Arlington; arXiv preprint)). Then it is treated and returned to the river (Google).
  3. Sometimes uses water nobody drinks. Google uses reclaimed or non-potable water at more than a quarter of its campuses (Google).
  4. Sometimes uses almost none. Air-cooled sites consume little or no water (Google (arXiv preprint)). Microsoft’s designs since August 2024 cool chips in a closed loop, avoiding more than 125 million liters a year per data center (Microsoft). The trade: dry cooling uses more electricity (Li, Yang, Islam, Ren (UC Riverside, UT Arlington; arXiv preprint)).

The water nobody counts per prompt

Making the electricity takes water too. In 2023 US data centers consumed about 66 billion liters on site, and nearly 800 billion liters at the power plants that fed them: twelve times as much (Lawrence Berkeley National Laboratory). Google’s five drops per prompt count only the cooling water at its own sites (Google (arXiv preprint)). Researchers split AI’s water into three scopes: cooling, electricity, and making the chips (Li, Yang, Islam, Ren (UC Riverside, UT Arlington; arXiv preprint)).

A day of ChatGPT prompts, at 212,500 gallons of cooling water, is what about 710 American families use at home in a day, at 300 gallons each (U.S. Environmental Protection Agency). Our estimate, on-site cooling only.

Water, to scale

Estimate
Data centers next to golf courses, power plants and farms. Farms use the most, by far.

AI and data centers Everything elsebillion gallons a year, log scale

Show as a table, with sources
Whatbillion gallons a yearNote
A year of ChatGPT prompts, cooling0.078Our estimate: 2.5 billion a day at 0.000085 gallons. (Sam Altman)
Training Grok 4, at the power plant0.2Epoch AI’s estimate of 754 million liters, for the electricity. (Epoch AI)
Google’s data centers, worldwide4.32021, consumed on site. (Google)
US data centers, on site172023, 66 billion liters consumed for cooling. (Lawrence Berkeley National Laboratory)
US data centers, at their power plants2112023, nearly 800 billion liters consumed making their electricity. (Lawrence Berkeley National Laboratory)
US golf courses, applied5472020, 1.68 million acre-feet applied, about a fifth of it recycled. (Golf Course Superintendents Association of America)
US power plants, consumed8612020, freshwater, 2,359 million gallons a day. (U.S. Geological Survey)
US irrigation, consumed35,000Our estimate from three USGS and USDA shares; see the arithmetic. (U.S. Geological Survey; USDA Economic Research Service; U.S. Geological Survey)

Most American water goes to three uses. Power plants, farm irrigation and city water supplies made up 90% of all withdrawals (U.S. Geological Survey). Irrigation alone took 47% of the fresh water (USDA Economic Research Service), and about 72% of that is consumed, taken up by crops or evaporated from fields (U.S. Geological Survey). Factories, mines, livestock and fish farms share most of the rest (U.S. Geological Survey).

On that ruler, data centers are small nationally. The catch is that they are not spread out. A typical AI data center uses as much electricity as 100,000 households, and the largest being built will use twenty times that (International Energy Agency). In a dry county, one campus can matter a lot even when the national total does not.

The arithmetic

Liters to gallons: ÷ 3.785. Acre-feet to gallons: × 325,851. Irrigation: 322 billion gallons a day withdrawn × 87% fresh × 47% irrigation × 72% consumed ≈ 95 billion gallons a day, × 365. The shares come from different years (2015 to 2020), so read it as a size. Golf is water applied, not consumed; most applied water is taken up by grass or evaporates, but we did not find a figure for how much.

How bad, how good

Factual
Smaller than the headlines per use. Bigger than it looks in growth and in one place. Improving fast. Still under-reported.

Bigger than it looks

Still unknown

  • No lab publishes energy per agent session. The 0.34 Wh figure comes without a stated method.
  • Labs rarely publish what training a model used. Our training figure is an outside estimate (Epoch AI).
  • Per-prompt water figures count cooling, not the power plant (Google (arXiv preprint)).

What follows for you: per use, AI is not where most of your footprint is. Where it adds up is in growth, in heavy agent work, and in the towns that host the buildings. If it matters to you, ask where a provider builds and how it cools, use the smallest model that does the job, and keep agent sessions short. The running count on Plate IV shows the totals tick.