AI Is Thirsty, and Water Companies No Longer Have a Say

American data centers consumed 17.4 billion gallons of water in 2023. By 2028, this volume will reach between 38 and 73 billion gallons according to projections from the Shehabi et al. (2024) report from the Lawrence Berkeley National Laboratory, commissioned by the U.S. Department of Energy. A guaranteed doubling, a possible quadrupling. While public debate has focused on terawatt-hours, another resource is quietly diminishing in the aquifers beneath the campuses of Microsoft and Google.

Water is the blind spot in the debate over AI’s environmental footprint. It has no meter on electricity bills, it generates no green certificates, it does not appear in the carbon commitments of major technology companies. And yet, the United Kingdom has just created an institutional precedent that should capture attention: by granting data centers the status of critical national infrastructure, London has effectively reduced the capacity of water companies to restrict their supply during drought periods. The physics of cooling is reshaping water law.

The Essentials

  • American data centers consumed 17.4 billion gallons of water in 2023; projections from the Shehabi et al. report (LBNL, 2024) place this volume at between 38 and 73 billion gallons by 2028.
  • In the United Kingdom, the critical national infrastructure status effectively reduces Thames Water’s capacity to restrict data center supply during droughts, according to a 2025 British government report.
  • The thermodynamic dilemma is real: evaporative cooling consumes little energy but much water; closed-loop cooling does the opposite.
  • Microsoft, Google, and a few European players are experimenting with low water consumption systems, but their large-scale deployment remains unproven.
  • Regulatory pressure is mounting: several American states now require data centers to declare water consumption, and the European Union is preparing reporting obligations.

Why AI Drinks More Than Conventional Servers

An ordinary data center consumes water mainly to keep its servers below an acceptable temperature threshold. The logic is simple: the electricity powering the chips is transformed into heat, the heat must dissipate, and evaporated water carries this heat away with it. It’s the same principle as an industrial cooling tower or a car radiator.

What changes with generative AI is the intensity. Training a large language model solicits graphics processors at levels close to their maximum capacity for weeks or months. Inference—that is, answering a question on ChatGPT or generating an image on Midjourney—mobilizes GPU clusters continuously, twenty-four hours a day. Researchers from the University of California at Riverside estimated that each conversation of twenty to fifty questions on ChatGPT-3 consumes approximately 500 milliliters of water. A plastic bottle. Multiplied by millions of daily users, the arithmetic becomes dizzying.

The International Energy Agency’s report on AI and energy confirms the trend: water consumption in data centers should grow faster than their electrical consumption in coming years, precisely because AI acceleration architectures concentrate heat on smaller surfaces and require more intense cooling. Megawatts and gallons do not progress at the same rate, but gallons are accelerating.

The Physical Dilemma That No Decree Can Solve

There are two major families of cooling systems for data centers. The first uses water evaporation: water absorbs heat from servers, evaporates in cooling towers, and this vapor carries heat to the atmosphere. It is efficient from an energy standpoint—it reduces the site’s electrical bill—but it consumes much water, because evaporated water is lost. The second family operates in a closed loop: water cools the servers, then is itself cooled by mechanical systems (chillers, heat pumps) before returning to the circuit. Less water lost, but noticeably higher electrical consumption.

This is not a choice of values. It is an engineering choice with real physical constraints. In a region where water is abundant and electricity expensive, evaporative cooling is rational. In a region where water is scarce and renewable electricity is available, closed-loop cooling is more sensible. The problem is that many data centers were built in areas that were water-abundant at the time of their establishment—and no longer are.

Northern Virginia hosts the largest concentration of data centers in the world, with approximately 35% of American national capacity according to industry estimates. It draws from the Chesapeake Bay watershed and local aquifers. Yet the region is experiencing increasing episodes of water stress linked to recurring droughts. Arizona, another epicenter of the sector, is directly subject to Colorado River water restrictions. These locations were not chosen for their water resources—they were chosen for land availability, favorable taxation, and proximity to fiber optic cables.

The British Precedent and What It Reveals

In 2025, the British government published a report on data center water consumption as part of its digital infrastructure policy. Its operational conclusion was clear: data centers receive the status of critical national infrastructure, which grants them the same protections as hospitals or transportation networks in case of shortage. In practice, this status greatly reduces the likelihood that Thames Water—the water company serving London and much of the South East, already under financial and regulatory strain—will restrict a data center’s supply during a drought episode. No law formally strips it of this power, but the CNI status creates a strong presumption in favor of maintaining the supply.

This choice has economic logic. The United Kingdom decided that data centers are growth engines and that an unplanned outage at a center hosting financial services or health systems would have unacceptable consequences. But it creates an implicit hierarchy in water allocation: servers before gardens, perhaps before farmers if drought worsens. No elected official has stated it this way, but that is what the law says.

The British case is not isolated. In Ireland, where Dublin has become one of Europe’s largest data center hubs, energy and water regulators have had to revise their planning models after underestimating demand growth. EirGrid, the Irish electricity grid operator, imposed a temporary moratorium on new data center connections in the Dublin region between 2021 and 2023, precisely because infrastructure planning had not anticipated the scale of demand. The same tension is emerging with water.

What Major Operators Are Doing—and What They Are Not Yet Doing

Microsoft announced in 2024 its intention to deploy closed-loop cooling systems at its new sites, with the goal of becoming “water positive” by 2030—that is, of returning more water to local watersheds than it consumes in its global operations. Google has formulated a similar commitment. Meta has published detailed data on consumption by site.

These commitments are real. They are also partial. Closed-loop cooling works well on new sites, built ab initio with this constraint in mind. It is far more difficult to deploy as a retrofit on existing infrastructure, the majority of which was designed for evaporation. The question posed by independent engineers is not whether the technology exists—it does—but whether it is deployable at the scale of tens of millions of square meters of data center surface already in operation.

Other approaches are being explored. Direct liquid cooling, which places the heat transfer fluid in direct contact with chips rather than cooling the ambient air in server rooms, significantly reduces water consumption. Companies like Vertiv and Schneider Electric have been commercializing these systems for several years. Immersion in baths of dielectric fluids—non-conductive liquids in which servers are fully submerged—eliminates virtually all water consumption. Operators like Submer and GRC (Green Revolution Cooling) are deploying this technology at pilot sites. The question remains one of industrial maturity and investment cost, which remains significantly higher than conventional approaches.

There are also localized approaches that leverage geography. Data centers located in Iceland, Norway, or Finland use cool outdoor air cooling for much of the year, reducing both their electrical consumption and water consumption. Microsoft announced its data centers in the Stavanger region as early as 2018, with Azure launch in 2019, exploiting this logic. Residual heat from servers is reused to heat adjacent municipal buildings, transforming the data center from a pure consumer into a participant in an urban thermal network. This is a promising path for new construction in Nordic countries. It solves little for data centers in Phoenix or Loudoun County, Virginia.

The question of energy transition is inseparable from that of water transition: both resources are simultaneously under tension on the same sites, and optimizing one without considering the other produces incomplete solutions.

Regulation Is Beginning to Adjust, Slowly

Until 2022, no American state required data centers to publicly declare their water consumption. The situation began to change under combined pressure from environmental associations, some local elected officials, and the rising prominence of the issue in the AI debate. California adopted legislation in 2023 requiring data centers to declare water consumption above a certain threshold. Maryland and Virginia followed with similar, less stringent provisions. These measures are a first step toward transparency—but they impose no limits.

The European Union goes further on paper. The revised Energy Efficiency Directive has required since 2024 that data centers over 500 kilowatts declare their water efficiency (WUE, Water Usage Effectiveness) as well as their energy efficiency. It also introduces obligations for reuse of residual heat. These obligations apply to operators established in Europe, not those hosting services intended for Europe from other continents—a gap the Commission acknowledges and intends to fill.

The challenge of reporting is real. Calculating a data center’s water consumption is not trivial. “Operational” water—that which evaporates in cooling towers—is measurable. But major operators like AWS and Microsoft also host their equipment in colocation centers they do not own, which complicates the chain of responsibility. And “embodied” water in chip and equipment manufacturing—which represents a substantial fraction of total water footprint according to some studies—escapes entirely from current reporting scopes.

AI already governs entire sectors of the digital economy, as shown by data on American productivity. The question is no longer whether this infrastructure will grow—it will grow—but under what conditions and according to what rules for sharing natural resources.

Geography of the Next Constraint

Water consumption by data centers is not uniformly distributed. It concentrates in a few zones where the industry has massively invested: Northern Virginia, Texas, Arizona, Georgia in the United States; Dublin in Ireland; Frankfurt, Amsterdam, and Stockholm in Europe. Some of these zones are already under water stress. Others will probably become so under the combined effect of climate change and growing demand.

This is not inevitable. It is a location decision. Choices made today about where to locate new campuses—Scandinavia rather than Arizona, zones of water surplus rather than zones under stress—will have consequences for twenty or thirty years, the typical lifespan of a large data center. Several actors are beginning to integrate water risk into their location decisions. BlackRock and other institutional investors have introduced water stress criteria into their analysis frameworks for digital infrastructure investments.

Pressure will also come from competition with other water uses. Irrigated agriculture represents approximately 70% of global freshwater consumption according to the FAO. Data centers are still far from this level, but their growth in fragile agricultural zones—like California’s Central Valley or the Texas plains—creates concrete usage conflicts, not merely theoretical ones. Several Virginia counties have already received construction permit requests denied or conditioned on water neutrality commitments.

The technology to significantly reduce data center water consumption exists. Closed-loop cooling, dielectric immersion, cool air cooling, residual heat reuse are all operational solutions. What is missing is the economic or regulatory signal that would make their systematic deployment preferable to the cheapest short-term solution. Water remains very inexpensive in most American jurisdictions. As long as an evaporative cooling tower costs less to build and operate than a closed-loop system, and as long as consumed water is not priced at its scarcity cost, the market alone will not make this transition.

The real question posed by the British precedent is therefore not whether data centers deserve protection in case of drought. It is who decides the hierarchy among water uses—and whether this decision is made in explicit public debate or through accumulation of silent precedents.

Sources

  1. UK Government — Water use in data centre and AI report (2025)
  2. Shehabi A. et al., United States Data Center Energy Usage Report — Lawrence Berkeley National Laboratory (data cited by US EPA)
  3. International Energy Agency — Energy and AI (2025), iea.org
  4. Environmental and Energy Study Institute (EESI) — Data Centers and Energy, eesi.org
  5. Consumer Reports — The Hidden Environmental Cost of AI, 2024
  6. European Directive on Energy Efficiency (2023/1791), Official Journal of the European Union
  7. LBNL – 2024 US Data Center Energy Usage Report (Shehabi et al.)
  8. GOV.UK – Data Centers Designated as CNI (Sept. 2024)
  9. GDSA – Water use in AI and Data Centres (GOV.UK PDF)
  10. UC Riverside – Making AI Less Thirsty (arXiv:2304.03271)
  11. European Commission – EED / Delegated Regulation EU/2024/1364
  12. Microsoft – Water Positive Commitment 2030
  13. Google – Water Stewardship Commitments
  14. IEA – Energy and AI (2025)
  15. Energy Connects / CRU – End of Irish Moratorium (Dec. 2025)