The rapid expansion of artificial intelligence is putting growing pressure on water supplies in areas where data centres are being built, with concerns emerging from the United States to India.
In Virginia, home to the world’s largest concentration of data centres, around a third of the state faced extreme drought conditions this summer while server farms continued to consume millions of gallons of water each day.
In India’s Uttar Pradesh state, wells in Gautam Buddha Nagar, an emerging data-centre hub, that once reached water at 20 to 30 feet are now being dug to 80 feet or more, according to reporting by Down To Earth.
The United Nations has also warned about the resource demands of AI. Researchers at the UN University estimate that AI-linked data centres could consume enough water by 2030 to meet the basic annual needs of everyone in sub-Saharan Africa.
The problem is particularly acute because AI’s water footprint is concentrated in specific areas rather than spread evenly across countries. Data centres are often built where electricity, fibre-optic networks and tax incentives are available, rather than where water is abundant.
Two sides of AI’s water footprint
The water used by data centres comes from two main sources.
The first is water used directly to cool servers. AI hardware generates substantially more heat than conventional servers. High-density AI racks can consume more than 100 kilowatts, according to engineers cited by IEEE Spectrum.
Many data centres use evaporative cooling systems, in which water is consumed as it evaporates. Researcher Shaolei Ren of the University of California, Riverside, told Bloomberg that data centres typically evaporate about 80 percent of the water they draw for cooling.
The second source is less visible: water used by power plants that generate electricity for data centres.
Thermal and nuclear power plants require water for cooling, and the huge electricity demand of AI makes this indirect water footprint significantly larger than the water used at the data centre itself.
A July 2026 water-policy brief by the Information Technology and Innovation Foundation estimated that indirect water consumption associated with electricity generation is roughly 12 times direct cooling consumption.
Estimates vary widely
There is no single agreed figure for the amount of water consumed by AI.
The UN University’s Institute for Water, Environment and Health estimated that global data centres consumed about 4.5 trillion litres of water in 2025, equivalent to around 1.8 million Olympic-sized swimming pools.
It projected that consumption could rise to 9.3 trillion litres a year by 2030 under a high-growth AI scenario.
The International Energy Agency, however, estimated global data-centre water consumption at about 560 billion litres in 2023, rising to around 1.2 trillion litres by 2030.
The difference reflects variations in methodology, including whether researchers measure water withdrawal or consumption, whether they count only direct cooling or the full electricity supply chain, and whether water used to manufacture computer chips is included.
The UN Secretary-General launched an AI Environmental Transparency Initiative on 23 June, calling on major AI companies to disclose their carbon, water and land footprints using standardised measures.
Companies report rising consumption
Corporate disclosures show that water consumption is increasing as AI infrastructure expands.
Google’s 2026 environmental report says the company consumed 10.9 billion gallons of water in 2025, 34 percent more than the previous year and more than twice its 2021 consumption.
Amazon reported its first absolute water consumption figure in June, saying it used 2.5 billion gallons globally in 2025.
Microsoft has said it is on track to become “water positive” five years earlier than originally planned.
The figures highlight a central problem: companies are becoming more efficient in their use of water per unit of computing, but total consumption continues to rise as demand for AI services grows.
Researchers describe this phenomenon as the Jevons Paradox, in which efficiency gains can encourage greater use of a resource and offset some of the savings.
Virginia faces drought pressure
Northern Virginia’s Loudoun County, known as “Data Center Alley”, has the world’s highest concentration of hyperscale computing facilities.
By late June, 10 of Virginia’s 11 drought-monitoring zones were under drought warnings. The state’s governor, Abigail Spanberger, urged residents to conserve water.
The Sierra Club’s Virginia chapter said the average data centre consumed roughly four million gallons of water a day.
State officials have disputed suggestions that data centres are exempt from drought restrictions. Virginia’s Department of Environmental Quality said groundwater permits already take seasonal conditions into account.
Officials in Henrico County also said residential complexes and hospitals remain larger water consumers than data centres.
The concern, however, is the speed and concentration of the industry’s growth. One analysis projects that data centres’ share of total water consumption in the Washington metropolitan area could rise from 8 percent in 2025 to 25 percent by 2035.
Virginia introduced new water reporting requirements in its 2026 budget. At least 30 data-centre water bills were introduced in 16 US states during the year, with four states — Idaho, South Dakota, Utah and West Virginia — passing new requirements.
Uttar Pradesh’s groundwater under pressure
Gautam Buddha Nagar presents a different challenge.
The district’s groundwater extraction rate was 104.79 percent of natural replenishment, according to India’s Central Ground Water Board’s 2023 assessment.
That means more water is being extracted than naturally replaced, even before accounting for additional demand from new data centres.
Down To Earth’s reporting from Khora Colony, near an AdaniConneX data-centre construction site, found residents whose borewells had reached depths of about 600 feet.
Uttar Pradesh has signed a deal worth about 390 billion rupees ($4.4bn) with data-centre operator Yotta for five additional facilities. The state’s 2026 data-centre policy targets more than 2GW of new capacity.
Local officials have acknowledged that they do not have basic information on how much water the facilities consume or where that water comes from.
Video and image generation increase demand
The environmental cost of AI also depends on what users ask systems to do.
The UN University estimates that generating a single AI image can have an electricity-related water footprint of about 29 millilitres. Generating a complex video can require about 4.1 litres.
That difference matters as AI services increasingly move beyond text to image, audio and video generation.
Researchers estimate that everyday inference — the process of running trained AI models in response to user requests — now accounts for 80 percent to 90 percent of AI’s total energy and water consumption.
One widely used consumer AI service is estimated to process about 2.5 billion prompts each day.
As AI becomes cheaper and more efficient, greater use could offset some of the environmental gains from improved technology.
Cooling technology offers possible solutions
Companies and governments are exploring ways to reduce the industry’s water footprint.
Industry data suggests that 19 percent of data-centre operators were using liquid cooling in 2026, with adoption expected to increase.
Newer direct-to-chip and immersion cooling systems can reduce cooling-related power consumption by 50 percent to 60 percent, according to industry tests cited in the source material.
Closed-loop cooling systems can reduce water use substantially compared with conventional cooling towers. Microsoft has said all its new data-centre designs use zero-water evaporative cooling.
The use of treated wastewater instead of drinking water is also gaining attention. Virginia has linked some state funding for data centres to the use of reclaimed water.
Can AI growth be separated from water stress?
The central challenge is that the companies investing most heavily in AI also have the resources to develop more water-efficient technology.
Google, Microsoft and Amazon point to lower water use per unit of computing and investments in water reuse and closed-loop systems.
But communities in places such as Virginia and Gautam Buddha Nagar are facing immediate pressure on local water supplies.
Both trends can be true at the same time.
The lack of consistent reporting makes it difficult to determine the scale of AI’s water consumption or compare companies and facilities reliably.
That is why researchers and policymakers are increasingly calling for standardised and mandatory disclosure of water use, alongside greater use of reclaimed water and cooling technologies that require less water.
