All concerns Concern 01 · Water

AI uses water. How much should we worry?

It depends where.

Nationally, data centers use far less water than agriculture. Locally, one large facility can put real pressure on a town’s supply.

Start with the concern
First, one basic fact

AI water use comes from three physical systems.

Make the chips

Factories use highly purified water to clean silicon wafers.20

Make the electricity

Some power plants consume water while producing electricity.4

Cool the computers

Some data centers evaporate water to remove heat. Others use different cooling systems.2

The case against AI

A large data center can become one of a town’s biggest water customers.

The strongest concern is simple. A fast-growing industry is placing large facilities in communities that may have little water to spare. Cooling, electricity generation, and chip manufacturing all use water. The companies often disclose too little for residents to judge the tradeoff.15

One third

Google’s three data-center sites used about one third of The Dalles, Oregon’s water in 2024.16

The number establishes a large local share. Residents still need separate evidence about shortages, summer peaks, drought, and who pays for new capacity.

What the evidence supports

Nationally, crop irrigation consumes more than 120 times as much water as every data center combined.

Berkeley Lab estimated that all U.S. data centers consumed 66 billion liters onsite in 2023. It estimated nearly 800 billion more through the electricity they used.2 Those totals cover AI and non-AI work. No public source measures a national AI-only total.

Converted to the same daily unit, the full data-center estimate is about 627 million gallons. U.S. crop irrigation consumed about 75.7 billion gallons per day on average from 2010 through 2020.17 Agriculture gives the national numbers useful scale.

The national picture
120×

The full data-center estimate is large. Other parts of the economy consume much more.

One crop, for scale California almonds alone use more water than all U.S. data centers combined.

The almond figure counts water applied to fields, so this is a rough comparison.1819

A poorly placed facility can still strain one watershed. The national and local claims answer different questions.
What gets exaggerated

A 2023 study estimated one bottle for 10–50 GPT-3 responses.

The amount changed with location and operating conditions. The paper modeled GPT-3, a system released in 2020. News coverage kept the bottle because it was easy to remember. Online retellings often changed the range into one bottle for one prompt.61011

500 mL

The memorable image.

10–50

Modeled GPT-3 responses per bottle.

“One bottle per prompt”

The shortened online version.

My view

Set water rules for facilities, especially where water is scarce.

AI water use is real. National claims often blur AI with every other data-center workload. Local risk can still be serious. Water policy should focus on the facility, its watershed, and summer demand.

  1. IPublish actual annual and summer-peak use.
  2. IIName the water source and reclaimed share.
  3. IIIPlan for drought before construction.
  4. IVMake the developer pay for added capacity.

A town has every right to reject a project that does not fit its water supply. A person using text or coding AI is not the policy target.

The part that gets missed

AI can help us use less water.

Four practical uses are already being tested.

Find hidden leaks

Models can learn the sound and pressure patterns that reveal a leaking pipe.21

Water crops more precisely

Sensor data and forecasts can help farmers irrigate when plants need it.22

Forecast floods and droughts

Better prediction gives water managers and emergency crews more time to act.23

Improve desalination

Models can test plant designs and operating choices that produce more fresh water with less energy.2425

Water used by computing belongs in the ledger. So does water saved.

Research and footnotes

25 sources
  1. International Energy Agency, Energy and AIGlobal current and 2030 water estimates, water-source categories, and the U.S. municipal comparison.
  2. Berkeley Lab, 2024 United States Data Center Energy Usage ReportU.S. onsite and electricity-related water estimates, workload limits, and 2028 scenarios.
  3. U.S. Geological Survey, Estimated Use of Water in the United StatesNational water definitions and category context.
  4. U.S. Energy Information AdministrationGenerator-level withdrawal and consumption data for electricity production.
  5. Siddik, Shehabi, and MarstonPeer-reviewed analysis of direct, indirect, and water-scarcity-adjusted U.S. data-center use.
  6. Li and colleagues, Making AI Less ThirstyThe peer-reviewed source for the modeled GPT-3 bottle estimate.
  7. West Des Moines Water WorksMeasured Microsoft use, monthly shares, and the utility’s explanation of local restrictions.
  8. West Des Moines water-service agreementPeak-use ceiling, infrastructure conditions, and developer contribution.
  9. Google’s 2025 production estimateA vendor estimate of 0.26 mL of onsite water for a median Gemini text prompt.
  10. UC Riverside’s 2023 releaseThe origin of the memorable half-liter framing in university communications.
  11. Associated Press, ChatGPT and water in IowaEarly national amplification and links to company and utility records.
  12. Climate Emergency FundDisclosure of funding and training for anti-data-center campaigns.
  13. AI Now’s data-center policy toolkitAn advocacy toolkit designed to stop, slow, or restrict some projects.
  14. Right to ComputeDisclosure of pro-build fundraising for policy, media, research, and organizing.
  15. OpenAI threat reportAn adjacent electricity and local-impact campaign. The report did not identify a water campaign.
  16. Oregon Public Broadcasting, The Dalles city recordsFacility-level water use reported from city records.
  17. U.S. Geological Survey, Water Use Across the United StatesAverage daily consumption for crop irrigation, public supply, and thermoelectric power from 2010 through 2020.
  18. University of California Agriculture and Natural Resources, crop water requirementsCalifornia acreage and annual irrigation estimates for olives and almonds.
  19. California Institute for Water Resources, almond irrigationCalifornia almond acreage and the need for more precise irrigation.
  20. NIST, environmental assessment of semiconductor fabricationHow chip factories use ultrapure water, typical volumes, and recycling practices.
  21. Liu, Zayed, and Xiao, machine learning for pipe-leak detectionA peer-reviewed test of machine learning on acoustic signals from water pipes.
  22. USDA Agricultural Research Service, AI for precision irrigationResearch using sensor data and AI to estimate crop water stress and guide irrigation.
  23. NOAA Office of Water Prediction, AI forecasting programEvaluation of AI forecasts for water management, flood response, and emergency services.
  24. U.S. Department of Energy, desalination and water-reuse researchA machine-learning project to optimize reverse-osmosis pretreatment.
  25. Chen and colleagues, optimization of seawater desalinationA peer-reviewed model that tested desalination designs and reported lower energy use and higher water-production efficiency.