A 100-megawatt data center can use more than 2 million gallons of water per day, WSLS reported on Oct. 26. The number sounds absurd until you remember what the building does. It runs tens of thousands of chips around the clock, and every watt those chips draw eventually leaves as heat.
That heat is the whole story behind why data centers use so much water. Water is not an ingredient in a server, and nobody bottles it into a product. It is a heat sink, and evaporating it is one of the cheapest ways to carry thermal energy out of a building. Our earlier look at AI cooling and water demand covered one angle of the debate. This piece explains the mechanics underneath it.
The topic is trending for a reason. AI buildouts keep pushing rack power higher, local governments are asking operators to disclose their usage, and the public arguments mix up three different numbers: what a cooling tower evaporates, what a power plant withdraws, and what a closed-loop system recirculates. Below, we separate them, show the math, and look at what operators and cloud buyers should expect next.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Why Data Centers Use So Much Water: It Starts With Heat
Servers and other computing equipment generate heat while they work. That is physics, not a design flaw. Nearly all of the electricity a processor consumes ends up as thermal energy, and if the building cannot remove it, chips throttle or fail. Eric Sjostedt, a Virginia Tech researcher, put it plainly to WSLS: “At its core, data centers produce a lot of heat.”
Sjostedt also explained why the water debate gets so heated. In his words, “When people say data centers use a lot of water, it’s all attributed to the cooling system.” In other words, the water is not consumed by computing. It is consumed by the machinery that keeps computing from cooking itself.
Water became the default coolant because it moves heat far better than air. A fan pushes air across hot components and that air carries away a modest amount of energy per cubic foot. A water loop carries away much more per gallon. When the water also evaporates, the effect multiplies, because turning liquid into vapor absorbs a large amount of energy at once. Operators in hot climates lean on that effect because it cuts the electricity needed for chillers.
The upshot: any question about data center water usage is really a question about heat density. More chips per rack means more heat per square foot, and more heat means more water or more electricity to remove it. Nothing in the AI era has made the chips cooler.
Evaporative Cooling: How Water Becomes a Heat Sink
Conventional evaporative cooling removes heat by evaporating water. In a typical design, warm water from the building flows through a cooling tower, where a portion evaporates into the air and takes heat with it. The cooled remainder returns to the loop. Operators then add fresh makeup water to replace what vanished, and that makeup water is the number that shows up on the utility bill.
Data Center Knowledge reported this week that evaporative cooling remains popular because it is generally more energy-efficient than dry cooling, especially in hot climates. That efficiency is why the technology has stuck around even as water scrutiny has grown. A dry system avoids the evaporation but needs more fan and compressor power, and it can struggle on the hottest afternoons.
A paper hosted on Wiley Online Library, titled “Data Centers Water Footprint: The Need for More Transparency,” estimates that cooling systems can evaporate roughly 0.26 to 2.4 gallons of water per kilowatt-hour of server energy. The spread is wide because climate, humidity, water quality and equipment design all change how much a tower must evaporate to hold a given temperature.
Two points matter for readers comparing claims online. First, “water use” in a cooling context usually means water that evaporates and leaves the local system. Second, an individual site’s number depends on the weather. A facility in a dry, hot region may evaporate far more per unit of computing than one in a cool, humid region, even with identical servers.
The 100-Megawatt Math Behind the Headlines
The WSLS figure of more than 2 million gallons per day for a 100-megawatt site lines up with a scaling range Data Center Knowledge attributed to a report by Media Analytics Ltd. That report puts a 1 MW IT load using evaporative cooling at roughly 20,000 to 50,000 gallons per day. Scaling it linearly gives the table below.
| Facility IT load | Estimated daily water use (evaporative cooling) | Source basis |
|---|---|---|
| 1 MW | 20,000 to 50,000 gallons | Media Analytics Ltd., via Data Center Knowledge |
| 10 MW | 200,000 to 500,000 gallons | Same range, scaled |
| 100 MW | 2 million to 5 million gallons | Same range, scaled (WSLS: more than 2 million) |
| 1 GW | 20 million to 50 million gallons | Same range, scaled |
A cross-check using the per-kilowatt-hour range supports the picture. A 100 MW site running all day consumes 2.4 million kilowatt-hours. Multiply that by 0.26 and 2.4 gallons per kilowatt-hour and the result is roughly 0.6 million to 5.8 million gallons (our arithmetic, not a published figure). The WSLS number sits comfortably inside that band.
Treat the table as an upper-bound sketch for evaporative designs, not a forecast for any named facility. Sites that use closed-loop liquid cooling or dry cooling can fall far below it. Sites in extreme heat can exceed it.
From 5.6 Billion to 17 Billion Gallons: The Direct Footprint
U.S. data centers directly used approximately 17 billion gallons of water in 2023, compared with 5.6 billion gallons in 2014. That is roughly a threefold jump in nine years. The National Association of Counties, which published a primer titled “Water matters for data centers,” cites the same trend and rounds the 2023 figure to 17.4 billion gallons.
The timing explains the growth. Between 2014 and 2023 the industry moved from enterprise server rooms to hyperscale campuses, and then to GPU clusters built for machine learning. Each shift packed more power into each rack. Cloud demand rose, buildings multiplied, and the cooling load rose with them.
| Metric | Figure | What it measures |
|---|---|---|
| U.S. direct data center water use, 2014 | 5.6 billion gallons | On-site use, mostly cooling |
| U.S. direct data center water use, 2023 | About 17 billion gallons | On-site use, mostly cooling |
| Freshwater for electric power, seven states (Ceres) | About 3.4 trillion gallons a year | Withdrawals, not consumption |
| Water per 100 MW site (evaporative) | More than 2 million gallons a day | Typical on-site cooling estimate |
Context helps here. Seventeen billion gallons is large in absolute terms, but spread across the country it is small next to agriculture or thermoelectric power. The concern is local. A single campus can concentrate millions of gallons of daily demand on one municipal system or one aquifer, and that is what triggers permit fights.
The Hidden Water Behind the Watts
Data centers also use water indirectly through electricity generation. Thermal power plants boil water, drive turbines, and need cooling of their own. Every kilowatt-hour a server draws carries a share of that upstream water. A Ceres analysis summarized by Yahoo News found that electricity demand is driving the sector’s growing water footprint.
The Ceres report, titled “The Water Behind the Watts,” estimated that data centers in seven states use about 3.4 trillion gallons of freshwater annually for electric-power generation. The figure is dramatic, and it deserves a careful reading.
Withdrawal Is Not Consumption
The 3.4-trillion-gallon number concerns water withdrawals. A withdrawal means water taken from a river, lake or aquifer, and it can include water that returns to the source or to storage after use. Consumption means water that leaves the system, typically by evaporation. Comparing a withdrawal total to a consumption total, as some viral posts do, produces nonsense.
Menon, quoted in coverage of the issue, made a related point about closed-loop coolers. Upstream water consumption, Menon said, “can be about a factor of like 10-x” greater than on-site use. So a facility that advertises very low on-site water use may still sit on top of a large water footprint at the power plant. Whether that matters depends on the grid mix. Solar and wind use little water for generation, while thermal plants use a lot.
This is also why storage and clean-power buildouts connect to the water story. Our report on Tesla’s Megapack demand tied to AI data centers shows how operators are chasing power supply, and the water attached to that power comes along with it.
What Experts and Operators Say
The verified record from engineers and operators agrees on the basics and differs on emphasis. Here are the statements we could confirm, with their sources named so you can read them in full.
- Google, on its data center water page (datacenters.google/water): “Data centers generate heat and need to be cooled to operate.” Google adds that whether water gets used depends on local conditions and the selected cooling technology.
- Google, in its water stewardship announcement on the company blog: “Because data centers generate heat from the servers and chips that power the digital world, they often rely on water to cut back on energy use for cooling.” The company frames water cooling as a way to save electricity compared with air cooling.
- Christopher Intagliata, NPR science journalist, in an NPR transcript on AI infrastructure: “And by consume, I mean that the water evaporates and doesn’t necessarily return to the local watershed.” That definition of consumption is the one most researchers use.
- Tim Shedd, a data-center water and cooling expert, on Latitude Media’s Catalyst podcast: “But for those select data centers, it did consume, relatively speaking, a lot of water because I’m cooling megawatts, right?” He ties high water use to high power density at evaporative sites.
- Microsoft, on its sustainable datacenter blog: “By adopting chip-level cooling solutions, we can deliver precise temperature control without water evaporation.” Microsoft contrasts this with conventional evaporative cooling.
Read together, the quotes describe a consistent trade. Operators who use water do it to save power. Operators who skip water spend the savings on chips and cooling hardware. Nobody in this group disputes that servers make heat.
Closed-Loop Liquid Cooling and Nvidia’s 80°C Chips
The newest buildouts try to break the link between heat and evaporation. Closed-loop liquid cooling circulates a coolant from the servers to a heat-rejection system and back again, so the fluid is reused instead of evaporated. TechRadar’s coverage of Nvidia’s modular AI factory design describes a closed-loop system that circulates liquid directly through servers, close to chips whose temperatures can rise above 80°C (176°F).
Direct-to-chip cooling works because a cold plate sits on the processor, so heat transfers into liquid at the source. Stantec, an engineering firm, describes current deployments that pair cold plates with coolant distribution units, using fluids such as a propylene glycol mixture of about 25% in a recirculating loop. The coolant never leaves the circuit, so it is not evaporated.
A 2026 estimate cited in University of Texas at Arlington coverage says closed-loop and direct-liquid designs can use 10 to 20 times less water than traditional evaporative facilities. Note the caveats. These systems still need periodic fluid replacement and some form of heat rejection, and the electricity they draw still carries upstream water. They shrink the on-site number. They do not erase the footprint.
Reliability matters too. When cooling fails, servers shut down fast. A recent cooling failure at a GW data center showed how much depends on the plant that moves heat away, whether it uses water or not.
The Water vs. Energy Trade-Off Operators Face
Andy Masley, an independent researcher who covers AI and data centers, summed up the design problem: “There’s a pretty direct trade-off between how much water is used and how much energy is used.” Every cooling choice slides along that line. Use more water, and the electric bill drops. Use less water, and the electric bill climbs.
Dry cooling can approach zero direct water use, according to the research behind this report, but it can require more electricity and may perform worse than evaporative cooling in very hot climates. Operators who choose it in a hot region accept both a bigger power draw and a bigger upstream water footprint if the grid is thermal-heavy. That is the catch buyers rarely see in a marketing claim.
| Cooling approach | Direct water use | Electricity demand | Main trade-off |
|---|---|---|---|
| Evaporative (cooling towers) | High: evaporates water continuously | Lower | Efficient in hot climates, but needs steady makeup water |
| Dry cooling (air-cooled) | Near zero | Higher | Can struggle in very hot weather |
| Closed-loop liquid, direct-to-chip | Low: fluid recirculates | Moderate | Needs new hardware and periodic fluid replacement |
| Upstream (power generation) | Varies by grid mix | Not applicable | Per Menon, can be about 10x on-site use for closed-loop sites |
The table shows why no single design wins everywhere. A closed-loop plant in a cool region on a clean grid looks excellent. The same plant in a hot region on a thermal-heavy grid may trade one water problem for another. Siting, grid mix and chip density decide the outcome.
Where the Water Comes From: Local Stakes and Disclosure
National totals hide the real fight, which is local. KUNR in Nevada reported in April that data centers’ water use is hard to track, raising concerns in the drought-prone West, citing a study. The headline describes the core complaint: communities cannot always see how much a campus draws, from which source, or in which season. You can read that report at KUNR.
The Congressional Research Service published a frequently-asked-questions brief on data centers and water for lawmakers, which signals how far the issue has climbed. Legislators now ask the same questions this article answers.
Siting disputes already shape projects. Our coverage of Maryland’s data center moratorium debate shows how quickly local limits can threaten a pipeline. In Europe, an order to halt forest work at a Google site in Finland shows that land and environmental permits can stall a campus as readily as water rules.
The pattern is clear enough. Operators who publish site-level water data, source type and seasonal peaks tend to get a hearing. Operators who guard those numbers invite suspicion, and in drought regions suspicion turns into moratoriums.
Market Impact: What It Means for Cloud Operators and Buyers
Water is now a line item in the economics of cloud capacity. Evaporative cooling cuts power costs but ties a site to a water permit. Liquid cooling raises capital spending on hardware and plumbing but lowers permit risk. Operators weigh these costs against the speed of getting capacity online, and speed usually wins in the AI race.
For buyers, the implications are practical. Enterprise customers with sustainability targets increasingly ask providers for region-level water metrics, not just carbon. Procurement teams can add water-use reporting to their cloud contracts, much as they do with emissions. Teams already tracking spend with FOCUS-style billing data can pair it with provider disclosures to see which regions carry the biggest footprint. Our look at the AI infrastructure race frames the capacity pressure behind these choices, and the broader cloud computing section tracks how providers respond.
Hardware vendors benefit from the shift. Companies that sell cold plates, coolant distribution units and heat-rejection gear gain from every operator who decides to cut evaporative cooling. Nvidia’s pitch in its modular designs, as TechRadar describes it, includes a promise of more energy-efficient data centers. By contrast, towers and water-treatment suppliers face a slower market in water-stressed regions.
Historical Context: How Cooling Got Here
Early data centers leaned on air conditioning. As facilities grew, many operators added water-cooled chillers and cooling towers because evaporation cut energy costs at scale, and Google’s own statement says water cooling can reduce energy use compared with air cooling. Electricity was the cost that operators watched, and water use was a secondary concern.
That ordering is flipping. As the figures above show, direct use rose from 5.6 billion gallons in 2014 to about 17 billion in 2023, and AI chips now run hot enough that air alone cannot keep up. Denser GPU clusters push operators toward liquid on the rack no matter what the water policy says.
In other words, the same chip trend that raises the cooling load also makes liquid cooling nearly unavoidable. Water use at the building edge may fall even as total heat grows. The open question is how much upstream water the extra electricity drags along.
Five Predictions for the Next 12 to 24 Months
These are our analysis, not reported facts, and they rest on the trends documented above.
- Disclosure rules spread. Expect more counties and states to require site-level water reporting as a permit condition, following the transparency push that researchers and reporters have documented.
- Closed-loop becomes the default for new AI halls. Chip temperatures above 80°C make direct-to-chip cooling the standard design for new GPU buildings, whatever the water rules say.
- The debate shifts upstream. As on-site use falls, attention moves to the power plant, which Menon’s factor-of-ten comment already hints at.
- Water-stressed regions see more moratoriums. Drought-prone areas will use pauses and permit limits, as Maryland’s debate suggests, to slow campuses that cannot show low-water designs.
- Buyers add water clauses to cloud contracts. Large customers will ask for region-level water metrics next to carbon data.
One caution applies to all five. Policy moves slower than hardware, and none of these outcomes is guaranteed.
What to Watch Next
Several signals will show whether the water story keeps growing. Watch for updated federal estimates beyond the 2023 data point, since the 17-billion-gallon figure is now nearly three years old and AI buildouts have accelerated since. Watch for operator disclosures that split water into withdrawal and consumption, because that split is the difference between a clean number and a misleading one.
Also watch the hardware cycle. If new GPU generations ship with cooling specs that assume direct-to-chip liquid, retrofit demand for older evaporative halls will rise. And keep an eye on local permit hearings in the drought-prone West, where the reporting gap that KUNR described is most acute.
Frequently Asked Questions
Why do data centers use so much water?
Servers generate heat, and conventional evaporative cooling removes that heat by evaporating water. A 100-megawatt facility can use more than 2 million gallons per day, according to WSLS. The water carries heat out of the building. It is not used in computing itself.
How much water do U.S. data centers use in total?
U.S. data centers directly used approximately 17 billion gallons in 2023, up from 5.6 billion gallons in 2014. Indirect use through electricity generation is separate and larger, though measured differently.
What is the difference between water withdrawal and water consumption?
Withdrawal is water taken from a source, and some of it may return to rivers or storage. Consumption is water that leaves the system, usually by evaporation. The Ceres figure of 3.4 trillion gallons for power generation in seven states is a withdrawal number.
Does closed-loop liquid cooling solve the problem?
It cuts on-site use sharply, with one 2026 expert estimate putting the reduction at 10 to 20 times versus evaporative sites. It does not remove upstream water tied to electricity. Menon said that upstream consumption can be about a factor of 10 greater than on-site use for closed-loop coolers.
Why not just use air cooling everywhere?
Dry cooling can approach zero direct water use, but it usually needs more electricity and can perform worse in very hot climates. Andy Masley described a direct trade-off between water and energy, so cutting one often raises the other.
How hot do AI chips get?
TechRadar’s description of Nvidia’s closed-loop design says the liquid runs near chips whose temperatures can rise above 80°C (176°F). That heat density is why direct-to-chip cooling is spreading.
Can I find out how much water a specific data center uses?
Often not. KUNR reported that data centers’ water use is hard to track, citing a study, and many operators publish only company-wide totals. Local permit filings and utility records are the best starting points.
