Superpower Daily
US Data Centers Could Use 1.125 Trillion Liters of Water by 203...London Surgeons Remove 11mm Brain Tumour With Live AI Anatomy M...Nvidia’s 70% Forecast Lifts Hardware; Salesforce and CrowdStrik...Socure Raises $156M at $5.2B, Buying Fravity to Automate Fraud...Kasm Pairs Xeon 6 Workspaces With Local AI, Says Data Stays In-...Trump Posts AI Scenes With Washington and WWII Generals on Trut...US Data Centers Could Use 1.125 Trillion Liters of Water by 203...London Surgeons Remove 11mm Brain Tumour With Live AI Anatomy M...Nvidia’s 70% Forecast Lifts Hardware; Salesforce and CrowdStrik...Socure Raises $156M at $5.2B, Buying Fravity to Automate Fraud...Kasm Pairs Xeon 6 Workspaces With Local AI, Says Data Stays In-...Trump Posts AI Scenes With Washington and WWII Generals on Trut...
Policypublished

US Data Centers Could Use 1.125 Trillion Liters of Water by 2030, Straining Dry Regions

The key question is not a universal water cost per AI query. It is whether new computing capacity is added where heat, drought and fossil-powered electricity make each unit of demand more consequential.

By 2 min read
US Data Centers Could Use 1.125 Trillion Liters of Water by 2030, Straining Dry Regions
Sign in to save and react

Listen to this story

The audio brief

About 1:34
0:001:34
Read transcript
US data centers could consume as much as 1.125 trillion liters of water a year by 2030, according to a forecast from a research group led by Fengqi You. That sounds like a national-scale crisis, but the more important risk is regional: new facilities may cluster in places already facing heat and drought, including Arizona and New Mexico. The numbers also measure two different things. Data-center cooling used about 66 billion liters in 2023, less than one percent of total US water consumption. The larger 2030 estimate adds water used by coal- and gas-fired power plants to generate the electricity those facilities need. So the footprint extends beyond the building and its cooling equipment. That is why water-per-query comparisons can mislead. Shaolei Ren estimated in 2024 that a short GPT-4 email used around half a liter of water, but he says that estimate is already outdated as systems become more efficient. Google estimated that a median Gemini query used roughly five drops in 2025. Those figures describe different systems and dates, not a universal cost for AI. The biggest levers are where facilities are built, how they are powered, and whether cooling fluid can circulate in a closed loop. You estimates those measures could cut future AI water use by up to 86 percent. But Texas shows the unresolved local question: data centers could rise from under one percent of state water demand today to three to nine percent by 2040. The constraint to watch is whether efficiency gains arrive before drought-prone regions absorb the next wave of capacity.

Story brief

3 key points

The main planning risk is regional, not national: data centers may have modest aggregate water use but intensify shortages where facilities cluster. A 2030 estimate of up to 1.125 trillion liters includes water consumed in electricity generation, unlike the 66 billion-liter 2023 cooling estimate. Texas could see data centers rise from under 1% of water demand today to 3–9% by 2040. Renewable-powered sites,...

  1. 01

    The 2030 forecast combines facility cooling with water used by coal- and gas-fired power plants.

  2. 02

    Texas researchers project data centers could account for 3–9% of state water demand by 2040.

  3. 03

    Fengqi You estimates siting and related measures could reduce future AI water use by up to 86%.

A rapid data-center buildout could put new pressure on already dry parts of the United States even though cooling accounted for less than 1 percent of national water consumption in 2023. By 2030, US data centers could consume as much as 1.125 trillion liters annually under a forecast from Fengqi You’s research group.

One footprint, two very different scales

Researchers estimated that US data-center cooling systems consumed 66 billion liters in 2023. The national total can mask the practical conflict: growing capacity may add significant water stress in drought-strapped places such as Arizona and New Mexico, while creating less concern in water-rich regions.

The 2023 figure and the 2030 forecast are not like-for-like measures. The earlier estimate covers cooling; the forecast also includes water used to generate electricity. Coal- and gas-fired plants require water to cool steam after it turns turbines, extending a facility’s water footprint beyond its own cooling equipment.

The choices behind the forecast

Per-query water claims are a poor guide to that system-level result. Shaolei Ren says his 2024 estimate that a short GPT-4 email used 500 milliliters of water is already outdated as models have become more efficient. Google estimated that a median-length Gemini query used about five drops in 2025, but that is a company estimate for one chatbot, not a universal AI measure.

  • Power: Moving toward solar and wind can reduce water use because they require little to no water to operate, unlike fossil-fuel generation.
  • Siting: Locating facilities away from drought-prone areas and near renewable supply could significantly reduce the footprint. You estimates siting and related measures could cut future AI water use by up to 86 percent.
  • Cooling: Liquid systems can recirculate cooled fluid in a closed loop, but extreme heat or humidity can still require extra water-based cooling.

Efficiency does not settle the local question

Texas shows the range of possible outcomes. Data centers now consume less than 1 percent of the state’s water demand, but researchers estimated a 3 to 9 percent share by 2040. Engineers expect efficiency gains to reduce that eventual figure, leaving local climate, electricity infrastructure and facility design as the decisions that will determine the burden.

Sources

  1. knowablemagazine.orgHow much of a problem is AI’s water use?

Related coverage

Continue reading