AI Drinks Water — But Saving It Spends Electricity

700,000 Liters and 6.6 Billion Cubic Meters

Inside the abstract of a single paper, two numbers describing amounts of water sit side by side.

The first number is 700,000 liters. Training a language model called GPT-3 once in Microsoft's state-of-the-art US data center can directly evaporate 700,000 liters of clean freshwater. The second number is 4.2 billion to 6.6 billion cubic meters. This is the amount of water global AI demand is projected to draw upon during the single year of 2027. That is more than 4 to 6 times the total water Denmark draws in a year, and more than half of the UK's amount (arXiv:2304.03271v5).

The two numbers differ right down to their units. One cubic meter is water filling a box 1 meter on each side—width, length, and height—which is 1,000 liters. The second number is written on a scale a thousand times larger than the first. Their denominators are different, too. The first is per single training run of one model, while the second is for the entire world for one year.

And the two numbers count different waters. The first is water that evaporated and vanished, while the second is water that was drawn. Water that is drawn does not disappear on the spot. Part of it evaporates, and the rest returns somewhere else.

Mixing these two ruins the math. Roughly 32 percent of the water US thermal power plants draw from rivers or groundwater is water pumped while not generating power. And a 2015 review counts studies warning of water scarcity that relied on drawn water instead of vanished water as their first error. We will examine both in detail in the next section. Yet the sentences we hear all the time lack that distinction.

"AI drinks water."

This article examines that sentence by dividing it into three parts: what and how much it drinks, what gets used instead if it tries not to drink, and why that answer changes from place to place. Along the way, we count water by splitting it into two places: the water used inside the data center, and the water used by the power plants that generate the electricity used by that data center.

There are two things to know before reading. First, the evidence grades of the cited studies are uneven. This topic has many preprints poured out in 2026, and preprint releases are independent of peer review. We also cite results produced solely by simulations, cases from a single factory, and results applying only to specific regions, and we mark them each time. Second, this article does not re-explain the physics of cooling systems. How and where the heat created by computing gets thrown away is explained by "When Air Is No Longer Enough — Where the Heat from Computing Goes". What we look at here is just one way water and electricity swap places with each other. The debate over how much electricity AI uses in total is in "How Much Electricity Does AI Use? — Why Every Source Gives a Different Number".

Pulled Water and Disappeared Water

About 32 percent of the water pulled up by US thermal power plants was pulled up while those plants were not making electricity (DOI: 10.1115/power-icope2017-3763). This figure emerged because the US Energy Information Administration bundled the amount of water pulled up by each plant, their power generation, and equipment configurations into a single public database, making it possible to look inside down to the single generator unit.

The reason they pull up water even when not generating power lies in how they operate. As the authors organized from interviewing plant personnel, peaking power plants run during peak demand and intermediate load power plants running somewhere in the middle sometimes run their cooling systems in advance. This is so they can respond immediately when operators order them to generate power. Considerations to reduce pump wear and manage water chemical status were also cited as reasons.

This number sounds strange because the phrase "using water" actually means two things. Withdrawal is the amount of water pulled from a river, groundwater, or tap, and consumption is the amount of that water which evaporated and did not return to that watershed. Water rinsed from laundry and sent down the drain is withdrawal, and water put in a pot and flown away as steam is consumption. Water pulled up by a non-generating power plant is counted as withdrawal, but that water does not vanish entirely.

Cooling towers do both of these things at the same time inside one device. While continuously reusing water, they lose inventory through two branches, one being evaporation and the other being blowdown (DOI: 10.1115/power2015-49201). The blowdown side keeps appearing in this article, so let's give it a name. Blowdown is water intentionally discarded before equipment gets damaged because the remaining water grows salty while circulating cooling water. It is like boiling soup for a long time until it gets salty, requiring you to pour in more water or scoop out a ladleful. The same paper notes that cooling towers are the facility that reuses water the most across industry and commerce, adding that their performance depends heavily on local climate.

You would think we could just measure consumption then, but that data does not exist. Even though the water balance equation sits right in the middle of hydrology, as of 2018, there were no actual measurements of water actually consumed by humans and the economy (DOI: 10.5194/hess-2018-53, a discussion paper under open review). However, the conclusion of the same abstract points the other way. The author writes that there are broad water management and policy questions that can be answered even with the withdrawal data we already have, and concludes that while waiting for better consumption surveys to be prepared, the withdrawal data we have now is sufficient for several of the most important scientific and practical purposes. Withdrawal data is still useful if you use it knowing what it is counting.

There is even a review whose conclusion flipped entirely because of this single distinction. A 2015 review paper re-read studies that had warned that widespread water shortages were coming soon to various US regions. The conclusion was that power generation would not cause water shortages in most regions. The first reason is precisely this distinction. Those alarmist studies mistakenly relied on withdrawal instead of consumption when measuring the gap between water demand and supply. The author also writes that markets will increase reuse and new supplies, technology will push up conservation and efficiency, and policies targeting price regulations, water rights, and administrative boundaries can reduce or eliminate shortages (DOI: 10.2166/wp.2015.060, a 2015 review limited to the US).

Therefore, every single water figure appearing in this article needs a tag attached to it. Is it water pulled in (withdrawal), or water that vanished (consumption)?

There Is Only One Metric, But Different Ways of Counting

When you gather 83 reported entries from data centres in one place, the range of water use is very wide. Some facilities report that water is virtually unnecessary, while others report that up to 2.5 litres of water go out for every kilowatt-hour of electricity spent on computing (DOI: 10.3390/fluids11080201).

There is a name for the scale that measures this value. Water Usage Effectiveness is the value obtained by dividing the amount of water used by a facility by the electricity that actually went into the computing equipment. It measures how many litres of water went out when 1 kilowatt-hour of electricity was used for computing. It has the same shape as Power Usage Effectiveness. Power Usage Effectiveness is the value obtained by dividing the electricity drawn by the entire facility by the electricity that actually went into the computing equipment among that ("How Much Electricity Does AI Use? — Why Every Source Gives a Different Number"), and Water Usage Effectiveness is just changing its numerator to water.

The reason the range is this wide is that the cooling methods are divided. Evaporative cooling is a method of dumping heat by evaporating water, and dry cooling is a method of dumping heat using only air instead of water. Sweating to cool the body is evaporative, and cooling with only an electric fan's breeze is dry. In the preceding survey, the side that reported water was virtually unnecessary was a dry facility using closed-loop liquid cooling, and the side that went up to 2.5 litres per kilowatt-hour was the evaporative family.

In the same survey, 23 out of 83 places exceeded 0.4 litres per kilowatt-hour. 0.4 litres is the target value suggested by the industry agreement, but the conditions attached to it are three. It must be a new facility, it must use drinking water, and it must be in a water-scarce region. There is also a name for the final condition. Water stress is a value indicating how closely the amount of water people intend to use in that region approaches the amount of water actually present there. It is a value closer to how many people are trying to drink from the cup at the same time, rather than the amount of water remaining in the cup. Therefore, the agreement's 0.4 litres is a target value attached to new facilities using drinking water in a water-stress region, and if even one condition is removed, it is not applied as is.

The authors of the same paper also organise what this metric measures. The reported Water Usage Effectiveness is a facility-level operational metric, and the share of heat given off by servers that is dumped through evaporation determines that value. The authors write that the performance of extracting heat well on the server side is merely a condition that can make Water Usage Effectiveness low, and does not determine that value. This organisation attaches a tag to Water Usage Effectiveness. If what determines the value is the share dumped through evaporation, the numerator of this scale is close to water that has disappeared (consumption).

It is also necessary to look at the size it occupies in the total. Water consumption by US data centres is 1.7 billion litres a day, which is small for the data centre share compared to the total US water consumption of 1.218 trillion litres a day (DOI: 10.1038/s41545-021-00101-w, 2021 review). However, the abstract does not split and write whether this 1.7 billion litres is only for direct cooling or includes even the water the power plant used to generate that electricity. Two different numbers are attached to the same abstract. In some cases, 57 percent of the water used directly for cooling came from drinking water, and fewer than one-third of data centre operators are measuring water consumption. Data centre water is a small share in the national total, but even the operators measuring that value are few.

There is also the point that the metric itself is shaking. The EU requires data centres through the revised Energy Efficiency Directive to measure, report, and disclose facility-level energy and water performance. One paper argues that the two metrics approved by this directive—Power Usage Effectiveness and Water Usage Effectiveness—can be twisted to create the illusion that efficiency has improved. The authors call this the "efficiency paradox." To get good scores, it becomes advantageous to remodel into a larger AI data centre, and that value is paid for in power supply and residents' water access. The authors' policy proposals are three: measures to reveal and certify efficiency improvements, measures to document trade-offs attached to metric improvements, and a framework to monitor diminishing returns and backfire effects over time (arXiv:2607.22604v1, preprint/EU policy argument). A separate review asking whether Power Usage Effectiveness still plays its part in the AI era also goes in the direction of looking at Water Usage Effectiveness, Carbon Usage Effectiveness, and equipment utilisation rate together (DOI: 10.38126/jspg280101).

The rebound effect is a phenomenon where the share saved because efficiency improved is filled back up because that much more is used. The preceding efficiency paradox has points resembling the rebound effect seen in electricity by "How Much Electricity Does AI Use? — Why Every Source Gives a Different Number". Getting better efficiency scores and having the total amount used decrease are not the same thing.

Save One, Use More of the Other

A pharmaceutical plant in São Paulo, Brazil, calculated three cooling configurations side by side. All three configurations use a chiller, which is a refrigeration machine that makes cold water to cool equipment. Two of them dump the chiller's heat into cooling tower water and differ only in the source of their makeup water, while one dumps it using only air. The first is a chiller cooled by cooling tower water, the second is the same chiller configured to reuse reverse osmosis concentrate as makeup water, and the third is a chiller cooled only by air. Reverse osmosis concentrate is the mineral-rich water left behind after a water purification system filters out clean water, and makeup water is the water freshly added to replace what the cooling tower loses through evaporation and drainage.

The air-only configuration eliminated cooling tower water consumption. Instead, it used about 29 percent more electricity, an amount equal to 1.7 gigawatt-hours per year (DOI: 10.2139/ssrn.7412964, single case study, pharmaceutical plant, Brazil). Where water went to zero, electricity stepped in. This is the backbone of this article.

The same study also measures the price paid by the side that saves water. Mixing in reverse osmosis concentrate raised the silicon concentration. As a result, the number of cycles a single drop of water could make before being discarded dropped from 7.4 to 4.3, and the cooling tower blowdown volume increased from 1.51 to 2.97 cubic meters per hour. Even so, across the entire system, drinking water consumption fell by 16 percent, wastewater generation fell by 41 percent, and electricity practically did not increase. Annual operating costs for the three configurations were 5.39 million, 5.01 million, and 5.13 million Brazilian real, respectively. The authors' conclusion is this: under the evaluated conditions, concentrate reuse achieved the best balance between water and energy, and air cooling maximized direct water savings but paid for them with greater electricity demand.

The same trade-off appears in data centers. There is a study that placed a 10-megawatt baseline facility across twelve data center hubs in the United States and compared dry cooling with evaporative cooling. Evaporative cooling is a method that lowers the temperature by spraying water onto the air entering a dry radiator. Under operational conditions where water is continuously sprayed, evaporative cooling reduced the electricity used to reject heat by a median of 30.5 percent and up to 41 percent. The price paid was about 8 to 56 gallons, or 30 to 210 liters, of on-site water consumed for every kilowatt-hour of electricity saved, and this range varied greatly by region (DOI: 10.2139/ssrn.7238869, preprint, simulation, ASHRAE design conditions).

The authors' judgment runs counter to intuition. The volume of on-site water is roughly an order of magnitude larger than the water that would have been saved at the power plant if that same 1 kilowatt-hour had not been used. Therefore, the authors conclude that evaporative cooling does not reduce operational water use, but rather shifts the water burden from the grid to the site.

The same abstract also notes results in the opposite direction. Under the most severe design conditions in Phoenix, Las Vegas, and Tucson, dry cooling failed to reach the target cooling water output temperature of 41.7 degrees Celsius entirely, making it unusable as the primary cooling technology. Out of thirty-six cases, only the twenty-nine cases where both technologies were feasible were compared, and the impossible cases named by the abstract are these three. Cost rankings were driven more heavily by the ratio of electricity prices to water prices than by climate. The scope of the evidence is also made clear. These twelve locations were chosen from the start as places where evaporative cooling would work well, and humid climates were left for the authors' upcoming follow-up research. The authors propose a four-axis framework pairing design conditions, grid carbon, water stress, and the electricity-to-water price ratio as the basis for site selection, noting that this is also at stake in the ongoing state-level moratorium debates over new data centers.

On the flip side, the side that drives water to zero also shows the same trade-off. A study that optimized urban infrastructure expansion plans alongside hourly climate data ran simulations comparing two strategies: evaporative cooling and closed-loop air-cooled chillers. When run on a representative urban microgrid, optimizing solely to make on-site water use zero shifted the resource burden toward the local power grid. That price required significant increases in low-carbon generation, energy storage, and total infrastructure costs (DOI: 10.2139/ssrn.6722099, preprint, simulation).

A study comparing six heat rejection configurations for liquid cooling facilities across four U.S. hubs concludes that no single configuration leads across all metrics, and the best choice depends on the climate and which metrics are weighted. When resource indicators are given equal weight, dry cooling with a chiller scores the highest at 0.90 to 0.95. By contrast, operating costs were lowest for water-cooled families at 0.100 to 0.135 dollars per IT kilowatt-hour. This means dry cooling demands a substantial financial premium under current electricity prices in exchange for saving resources. And as servers transition to generations that tolerate higher temperatures, both energy and water performance improve, making chillerless wet configurations financially superior across all hubs investigated (DOI: 10.2139/ssrn.7071714, preprint, simulation, 4 U.S. hubs).

This scale did not originate with data centers. A 2011 study comparing coal, nuclear, and natural gas combined-cycle power plants across five sites notes that, based on optimal design criteria under nearly all conditions, wet cooling is the cheapest and boasts the highest output and efficiency. Dry or hybrid cooling drastically cuts water, but those savings cost 7 million to 10 million dollars annually, resulting in a break-even water price of 3 to 6 dollars per 1,000 gallons (DOI: 10.1115/imece2011-63135, 2011, power plant case study). If water is cheaper than that, wet cooling wins; if it is not cheaper, dry or hybrid cooling wins.

It Takes Water to Make That Electricity Too

There is a map that matches the locations of 472 American hyperscale data centers—meaning massive facilities built by giant cloud companies—with power regions, river basins, and water stress data. Under a baseline assumption, operational water consumption came out to roughly 300 billion liters per year, and ranged between 205 billion and 451 billion liters depending on the scenario. Three-quarters of this total was not water leaving those facilities (arXiv:2607.02531v1, preprint).

The rest of the water left from power plants. Indirect water use is the water that power plants spent on behalf of the facility to make the electricity it used. It does not show up on the electricity bill, but it is water that has already evaporated and vanished at distant power plants. On this map, three-quarters of the data center water consumption was this indirect water.

The second finding of the same study is more important. On-site water and indirect water cluster in different places on the map. On-site cooling burdens pile up in water-stressed basins in the West and South-Central regions, while electricity-linked water piles up in a few grid regions in the East with a high share of fossil generation. A balancing authority is an operational unit in charge of matching electricity demand and supply within a single region, and three out of the 24 balancing authorities that house data centers accounted to 59 percent of the electricity-related water. The authors conclude that separating the pathways reveals where and which decisions matter. Cooling design and water sourcing must be handled locally, while power planning and procurement must be handled regionally.

There is a point that appears to clash with the results of the previous section. Indirect water accounts for three-quarters of the total, yet in the comparison of twelve cities, the water saved at power plants was only one-tenth of the water used on site.

Since the two values are answers to different questions, their magnitudes cannot be directly compared. The three-quarters is a distribution of the shares showing which pathways the water of 472 already operating facilities takes. One-tenth is a one-step exchange ratio measuring how much each of the two items moves when one cooling method is changed. The accounting standards differ as well. Three-quarters is a value that the abstract pins down as 'consumption', while the exchange ratio is a value where the abstract does not state whether on-site water and upstream water are withdrawals or consumption. And the exchange ratio came from twelve hubs chosen for dry climates in the first place. In the end, which of the two becomes larger depends on how much water the electricity of that region uses, and that is not a value data center designers can choose. This contrast was put together by us by placing the two papers side by side, and is not a sentence written down in either paper.

You might think that decarbonizing electricity would therefore also reduce indirect water, but that may not be the case. There is a tool that tracked electricity flows across Europe in 15-minute intervals to calculate carbon footprints and water footprints together. A water footprint is the amount of water used in association with an activity. According to the results of this tool, there is a trade-off between decarbonization and water use, and reservoir-type hydropower plays a prominent share in driving up the water footprint even in low-carbon grids. The same study adds that ignoring cross-border electricity trade and temporal variations leads to heavily skewed estimates for both footprints (arXiv:2601.11623v2, preprint).

A literature review collecting water intensity coefficients for China's power sector presents a parallel picture. Coal accounted for 50 percent of power-sector water consumption, and the water consumption intensity—meaning the average water consumed to produce a single unit of electricity—was highest for hydropower. These two are values on different axes and must not be read by overlapping them. The former is where the total went, and the latter is the intensity per unit. The same abstract notes that there was simply no research addressing the water consumption intensity of natural gas power generation, and studies on the withdrawal side were also scarce (DOI: 10.3390/su10041181, literature review, China-limited).

Making carbon and water better at the same time is not easy either. In the problem of scheduling computing tasks across geographically scattered data centers, carbon and water pull against each other, so optimizing for just one makes the other worse. The study that discovered this puts forward a scheduler that optimizes both together (arXiv:2501.17944v2, preprint).

As for how much one would gain by adjusting task placement and power plant operation while taking water into account, the answer given by one simulation is on the scale of a few percent. There is a framework that directly plugs water flows into dispatch optimization—which determines which power plants to run and how much. On IEEE 30-bus and 118-bus test systems, which are standard simulated power grids for research, this framework reduced power-related freshwater withdrawals by about 3 to 5 percent under water-constrained conditions (arXiv:2605.25854v1, preprint, test system simulation). While meaningful for a free gain, it is not a magnitude that changes the digit.

So It's Not About the Total Amount, But Where

When measuring the burden placed on community water supply facilities across ten data center sites in the United States, the gap between the lightest and heaviest sites spanned three orders of magnitude, ranging from 0.2 percent to 134 percent of the host capacity in those regions (arXiv:2606.21760v1, preprint, 10 US sites). The abstract measures this burden as a "water consumption impact index" and compares it to "host capacity," though it does not further clarify what that capacity specifically represents. The 134 percent figure means that the burden on that site exceeds the baseline capacity. The same paper considers that three things have been studied separately but form a single feedback loop: cooling requires water, water scarcity constrains site location, and AI tools can improve the efficiency of water systems.

Therefore, the weight of the exact same 1 liter changes depending on where and when it was used. Research targeting this point points out that prior studies have largely focused on reducing total usage. It then proposes a stress adjustment framework that reflects water availability on a county-by-month basis, incorporating even the water tied to electricity into this framework. On top of that framework, task scheduling, rainwater harvesting, and dry cooling as a water-free alternative are examined together (arXiv:2607.22617v1, preprint). 1 liter used in a place and at a time where water is scarce is heavier than 1 liter used in a place and at a time where water is abundant.

What hits locally first is not the annual total, but the amount of water that can pass through the pipes during a midsummer peak. A study examining direct water withdrawals by US data centers from the perspective of public water systems estimates that if 2024 water intensity levels remain unchanged, between 697 million and 1.451 billion gallons of new water treatment capacity per day will be needed by 2030 (arXiv:2603.02705v2, preprint). New York City's average daily supply is roughly 1 billion gallons. Under an optimistic scenario where water intensity improves by 10 percent each year, this drops to 227 million to 604 million gallons. Still, if IT load grows significantly, capacity capable of feeding half of New York City for most of the year may be required. The cost of that new capacity is on the scale of 10 billion dollars, reaching up to 58 billion dollars in the high-growth scenario.

These numbers do not represent the amount used over the course of a year, but rather the thickness of the pipes—in other words, the capacity—needed to endure the heaviest usage day. And here, the two axes of this article, water and electricity, meet again. As the same study points out, operators caught by water capacity constraints may end up relying on less efficient dry cooling, which then places heavier pressure on the power grid during the summer peak. The same paper also notes that this impact is heavily concentrated on the communities that host data centers. The authors' recommendations are fourfold: disclosure of peak water usage, agreements between corporations and communities, a "capacity neutral" approach that allows regions to preserve remaining water supply capacity, and joint planning that puts water and electricity together.

The burden is not distributed evenly either. Data centers do not arrive one by one but are proposed in clusters, so the impacts placed on local water utilities, power grids, and land-use frameworks overlap and multiply (arXiv:2602.10526v1, preprint, expert panel report). The same report also puts forth a policy roadmap to help the state of Georgia strike a balance among growth, sustainability, equity, and community protection. There are also simulation results showing that the conventional practice of splitting tasks and sending them to data centers in multiple regions can actually heighten interregional environmental disparities. The same study shows that placement explicitly incorporating equity drastically reduces regional gaps in both carbon and water (arXiv:2307.05494v2, preprint, 10-data-center trace simulation).

Calculating with a hypothetical placement makes the disparity even starker. Placing a cluster of 1,024 GPUs with evaporative cooling in the United Arab Emirates would consume over 30 million liters of water annually. That country is classified as having "extremely high" water stress (arXiv:2607.13443v1, preprint, virtual placement model). The authors present a triple dilemma where no nation can simultaneously maximize AI sovereignty, minimize environmental impact, and maintain resource access that citizens can afford. They then propose mandatory water-use efficiency reporting and climate-vulnerability location assessments as design principles.

There is also water that can be moved. Training is the process of building a model, while inference is the act of using a finished model once. The two have different location constraints. Inference must be close to users so that responses are not delayed, but training has no such constraint, meaning a 500-megawatt training cluster in rural Nebraska produces the exact same model as one of the same size in Ashburn, Virginia (DOI: 10.2139/ssrn.6823640, preprint). What this fact means from the perspective of water is singular: the share of water used by AI that is tied to training is water that can move across the map.

Choices That Change the Value

A petrochemical complex in Tarragona, Spain, mixed urban wastewater-treated reclaimed water up to 40 percent into the Ebro River water, feeding it into the cooling tower makeup for an ethylene cracker. The number of turns a drop of water makes before being discarded increased from 4 to up to 7, while total makeup water demand decreased by 22 percent, and the concentration cycles decreased by 50 percent. Reclaimed water contains fewer anti-corrosion components, but by tailoring chemical treatments and real-time monitoring, corrosion was kept within target limits. Mils—a thickness unit where a thousandth of an inch measures how much metal wears away annually—kept corrosion for carbon steel and copper alloys below 1 mil and 0.24 mils per year, respectively. Legionella bacteria were also not detected (DOI: 10.2166/wpt.2026.290, full-scale single case, petrochemical, Spain).

There is research that moves this choice to the data center side and evaluates it across the entire life cycle. Reuse also has a cost. The wastewater reuse scenario contributed roughly twice as much to global warming as using freshwater, with about 80 percent of that difference coming from the energy used to treat wastewater. Moving to a treatment configuration using ultrafiltration and reverse osmosis requires more than five times the energy of freshwater. However, this is offset as the number of water turns increases, reducing concentration cycles and chemicals (DOI: 10.31224/7240, preprint, life cycle assessment).

The latter part of the same paper shifts direction. When the system is recalculated under fully decarbonized conditions, the greenhouse gas emissions from reuse drop significantly, leaving chemical production as the main emission source. The authors suggest that for a cooling facility completed today and operated until 2050, most of its lifespan is spent in a system where the greenhouse gas penalty of reuse is negligible, and water savings accumulate fully from the start. Reclaimed water replaces freshwater on a 1-to-1 basis. The indirect water penalty caused by increased electricity and chemicals on the reuse side is roughly 0.93 liters per cubic meter, which is less than 0.1 percent of the direct replacement benefit. The abstract does not specify whether this value is withdrawal or consumption.

Yet reclaimed water is not used everywhere because the conditions for it to work are demanding. A study using the Microsoft Quincy case as a benchmark and comparing it with Google Douglas County and Amazon Web Services examines three axes: technology, regulation, and finance. The technology axis covers water quality, compatibility with cooling systems, supply reliability, monitoring and reserve capacity, and residue management. The regulation axis covers permitted uses and discharge requirements, jurisdiction, public acceptance, and governance. The financial axis covers capital and operating costs, tariff design, long-term off-take agreements, reliable demand, cash flow certainty, and risk allocation. The author's conclusion is short. Any single axis alone is not enough. All three axes must align simultaneously to turn reclaimed water from a technically possible resource into a public service that can recover its costs (DOI: 10.2139/ssrn.7292218, multi-case study).

And depending on what is reused, the costs are paid differently elsewhere. At a pharmaceutical plant mixed with high-salinity reverse osmosis concentrate, water turned less and concentration cycles increased. At the Tarragona complex mixed with reclaimed water filtered through reverse osmosis, conversely, the number of turns increased and concentration cycles decreased, and the life cycle assessment also calculated that configurations using ultrafiltration and reverse osmosis increase the number of turns and decrease concentration cycles. In return, the reuse side requires energy for treatment.

What We Must Ask

So a single question—"How much water does this data center use?"—does not yield an answer. What we can answer are the following five questions.

① Is this number water withdrawn or water consumed? The two values can diverge significantly at the same facility, and one review points out that reading withdrawals as consumption was the first error of past water shortage warnings (DOI: 10.2166/wp.2015.060).

② In which watershed and in which month did that water leave? The weight of the exact same 1 liter varies by location and season (arXiv:2607.22617v1), and depending on the site, the burden placed on the region ranges from 0.2 percent to 134 percent of host capacity (arXiv:2606.21760v1).

③ If on-site water is made 0, how much does electricity increase, and how much water does that electricity use? Comparing three cooling configurations at a pharmaceutical plant in Brazil showed that air cooling used about 29 percent more electricity (DOI: 10.2139/ssrn.7412964). There is a simulation showing that optimizing solely for zero on-site water shifts the burden toward the power grid (DOI: 10.2139/ssrn.6722099), and a map showing that indirect water use accounts for three-quarters of the total (arXiv:2607.02531v1).

④ In that region, what runs out first—water, electricity, or pipeline capacity? The three are different resources and crowd each other out. If hitting a water capacity limit leads to relying on dry cooling, that puts further pressure on the electrical peak (arXiv:2603.02705v2).

⑤ Who is measuring it, and are those figures made public? Fewer than one-third of operators measure water consumption (DOI: 10.1038/s41545-021-00101-w), and even disclosed metrics can be distorted (arXiv:2607.22604v1).

The final question supports the other four. That is because answering the first four requires someone to measure and disclose them first. Discussions treating the growth of AI data centers in water-stressed regions as a governance issue list three necessities. They are hydrological condition screening, transparency, and safeguards to maintain equity in water allocation. Specifically, they call for place-based regulations that jointly evaluate cumulative water withdrawal, drought sensitivity, emergency reductions, community participation, and mandatory reporting (DOI: 10.4018/979-8-2600-2761-5.ch007).

A study synthesizing the environmental and economic impacts of AI data centers in the U.S. reaches the same conclusion. Impacts are not determined by facility design alone; the electricity, water, and land systems where that facility is placed determine them together (arXiv:2608.09882v1, preprint). The same study also notes that improvements in energy efficiency, cooling configurations, and operational strategies can reduce this impact, while adding that their effectiveness also depends on grid conditions.

Water and electricity are two pans on the same scale. Pressing one down often causes the other to rise, and even cases where both were lowered together come with conditions. Wastewater reuse did under that plant's conditions, and high-temperature server generations did in a simulation. What we can choose is usually where to place the scale.

Therefore, the sentence "AI drinks water" is an unfinished sentence. It becomes a sentence that can finally be answered only when you fill in where, in what season, and in place of what it drinks.

Source Papers

Papers published on arXiv, SSRN, and Research Square are marked as "preprint," and whether they were subsequently published in academic journals was not separately verified. Results dedicated to simulations, single cases, or specific regions are noted in the text each time.

• "Making AI Less 'Thirsty'" (2023, preprint) — arXiv:2304.03271v5 — 700,000 liters of freshwater that GPT-3 training can directly evaporate, and a projected 4.2 billion to 6.6 billion cubic meters of global AI water withdrawal in 2027.
• "Effects of Cooling System Operations on Withdrawal for Thermoelectric Power" (2017) — DOI: 10.1115/power-icope2017-3763 — Roughly 32 percent of thermoelectric power plant withdrawal occurs while not generating power.
• "Water Consumption of Cooling Towers in Different Climatic Zones of the U. S." (2015) — DOI: 10.1115/power2015-49201 — Cooling towers lose inventory through evaporation and blowdown while recirculating water, and their performance depends on the climate.
• "When is water withdrawal data enough?" (2018, discussion paper under open review) — DOI: 10.5194/hess-2018-53 — Noting the lack of consumption measurements, with the conclusion that current withdrawal data is sufficient for several important purposes.
• "The electricity–water nexus: is a crisis imminent?" (2015, review, U.S. only) — DOI: 10.2166/wp.2015.060 — Pointing out that alarmist studies relied on withdrawal instead of consumption, and the opposite conclusion that power generation will not cause water shortages in most regions.
• "A Review of Thermal Management in Modern Data Centres" (2026) — DOI: 10.3390/fluids11080201 — Water usage effectiveness of 83 reported items. Dry closed loops are practically anhydrous, evaporative ones reach up to 2.5 L/kWh, and 23 sites exceed 0.4 L/kWh.
• "Data centre water consumption" (2021, review) — DOI: 10.1038/s41545-021-00101-w — U.S. data center water consumption is 1.7 billion liters a day out of a total 1,218 billion liters, 57 percent of direct cooling is potable water, and fewer than one-third of operators measure it.
• "The Fallacy of Sustainable Generative AI" (2026, preprint, EU policy argument) — arXiv:2607.22604v1 — The "efficiency paradox" where power usage effectiveness and water usage effectiveness can be twisted to create efficiency illusions, along with three policy proposals.
• "Is Power Usage Effectiveness (PUE) Still Fit for Purpose?" (2026, integrated review) — DOI: 10.38126/jspg280101 — A re-examination of power usage effectiveness suitability in the AI era and a synthesis of supplementary metrics.
• "Enhancing Industrial Water Security through Reverse Osmosis Concentrate Reuse" (2026, single case, pharmaceutical plant, Brazil) — DOI: 10.2139/ssrn.7412964 — Comparison of three cooling configurations. Air cooling eliminated cooling tower water and used about 29 percent more electricity.
• "When Does Adiabatic Cooling Outperform Dry Cooling?" (2026, preprint, simulation) — DOI: 10.2139/ssrn.7238869 — Adiabatic cooling cuts heat rejection electricity by a median of 30.5 percent, at the cost of 8 to 56 gallons of on-site water per kWh, which is an order of magnitude larger than the upstream savings. Three cities including Phoenix cannot use dry cooling.
• "Uncovering the burden shifting of data center cooling" (2026, preprint, simulation) — DOI: 10.2139/ssrn.6722099 — Optimizing solely for zero on-site water shifts the burden to the regional power grid.
• "The Energy-Water-Carbon Nexus of Liquid-Cooled Data Centers" (2026, preprint, simulation, four U.S. hubs) — DOI: 10.2139/ssrn.7071714 — No single configuration leads across all metrics. Resource metrics favor dry cooling, while operating costs favor liquid cooling.
• "Cost/Performance Comparisons of Water Conserving Power Plant Cooling Systems" (2011, power plant case study) — DOI: 10.1115/imece2011-63135 — Under almost all optimal design criteria, wet cooling is the cheapest and has the highest output and efficiency. Break-even water cost is $3 to $6 per 1,000 gallons.
• "The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era" (2026, preprint) — arXiv:2607.02531v1 — Mapping 472 sites. Operational water consumption is about 300 billion liters annually, electricity-related water accounts for three-quarters, and 3 out of 24 zones account for 59 percent of it.
• "Wattnet" (2026, preprint) — arXiv:2601.11623v2 — Tracking 15-minute resolution power flows in Europe. Reservoir-based hydropower increases the water footprint even in low-carbon grids.
• "Operational Water Withdrawal and Consumption Factors for Electricity Generation Technology in China" (2018, literature review, China only) — DOI: 10.3390/su10041181 — Coal accounts for 50 percent of power generation water consumption, and hydropower has the highest average consumption intensity. There are no studies on gas power generation intensity.
• "WaterWise" (2025, preprint) — arXiv:2501.17944v2 — Carbon and water conflict, so optimizing for only one makes the other worse. Proposing a co-optimization scheduler.
• "From Accounting to Coordination" (2026, preprint, test grid simulation) — arXiv:2605.25854v1 — Internalizing water into economic dispatch saves about 3 to 5 percent of power-related freshwater withdrawal.
• "AI Data Centers and the Water Use Feedback Loop" (2026, preprint, 10 U.S. sites) — arXiv:2606.21760v1 — Community burdens range from 0.2 percent to 134 percent of host capacity. Cooling water, location constraints, and AI's water efficiency contributions form a single feedback loop.
• "Balancing Bits and Drops" (2026, preprint) — arXiv:2607.22617v1 — County- and monthly-level stress correction framework and expansion to water tied to electricity.
• "Small Bottle, Big Pipe" (2026, preprint) — arXiv:2603.02705v2 — 697 million to 1.451 billion gallons per day of new water utility capacity by 2030. Impacts are heavily concentrated in communities hosting data centers.
• "The Infrastructure Equation" (2026, preprint, expert meeting report) — arXiv:2602.10526v1 — Cluster colocation increases impacts on water mains, power grids, and land.
• "Towards Environmentally Equitable AI via Geographical Load Balancing" (2023, preprint, trace simulation) — arXiv:2307.05494v2 — Conventional load balancing can increase environmental inequality, while equity-aware placement narrows the gap.
• "The Environmental Cost of Digital Sovereignty" (2026, preprint, virtual placement model) — arXiv:2607.13443v1 — Placing a 1,024-GPU cluster with evaporative cooling in the United Arab Emirates consumes over 30 million liters annually.
• "Follow the Curtailment" (2026, preprint) — DOI: 10.2139/ssrn.6823640 — Inference must be close to users due to latency constraints, but training has no such constraint.
• "Wastewater reuse for industrial cooling towers: a full-scale case study" (2026, full-scale single case, petrochemical, Spain) — DOI: 10.2166/wpt.2026.290 — Up to 40 percent reclaimed water blending increases cycles of concentration from 4 to 7 and reduces makeup water by 22 percent.
• "Reclaiming Cooling" (2026, preprint, life cycle assessment) — DOI: 10.31224/7240 — Reclaimed water roughly doubles greenhouse gas emissions, but the penalty becomes negligible in decarbonized grids.
• "Technical, Regulatory, and Financial Conditions for Bankable Water Reuse Projects" (2026, multiple case study) — DOI: 10.2139/ssrn.7292218 — Technical, regulatory, and financial axes must align simultaneously for reclaimed water to become a recoverable service.
• "Governing AI Data Center Growth in Water-Stressed Regions" (2026, book chapter) — DOI: 10.4018/979-8-2600-2761-5.ch007 — A regulatory framework combining hydrological condition screening, disclosure, and equity safeguards.
• "Environmental and Economic Implications of Artificial Intelligence Data Centers in the United States" (2026, preprint) — arXiv:2608.09882v1 — Impacts are determined by power, water, and land systems rather than facility design, and the effects of efficiency, cooling, and operational improvements also depend on grid conditions.