A single large data center can drink through as much water in one day as a town of 30,000 people. That is not a typo. Some hyperscale facilities pull 3 to 5 million gallons every 24 hours, mostly to keep rows of servers from cooking themselves. So when people ask how much water does a data center use, the honest answer starts with a range and ends with a long list of “it depends” factors like climate, cooling design, and whether you count the water burned at the power plant feeding the building.
This matters more every year. Data centers now sit in drought-prone counties, next to farms, and near reservoirs that supply drinking water to millions. Communities want answers, and operators are finally publishing numbers. In this guide, you will learn exactly how much water different sizes of data centers consume, how cooling systems actually work, what Water Usage Effectiveness (WUE) means, how AI workloads changed the math, which companies use the most and least, how much water your own ChatGPT prompts or Netflix streams indirectly use, and where the technology is heading next.
The Real Numbers Behind Data Center Water Consumption
A typical large data center uses between 1 million and 5 million gallons of water per day, which equals roughly 300 million to 1.8 billion gallons per year, while smaller enterprise facilities may use only 10,000 to 100,000 gallons daily. The spread is enormous because water use scales directly with how much heat a building must remove and which method it uses to remove that heat.
Think of it this way. Every watt of electricity a server consumes turns into a watt of heat. A 100-megawatt data center produces 100 megawatts of heat that has to go somewhere. Operators can move that heat with air, with water, or with a mix. Water is cheap, efficient, and moves heat about 3,500 times better than air by volume. That efficiency is exactly why so many facilities lean on it.
To put daily usage in perspective, here is how different facility sizes typically compare. These figures assume evaporative or water-cooled systems, which remain the most common design in warm climates.
| Facility Size (IT Load) | Typical Daily Water Use | Annual Water Use | Rough Household Equivalent |
|---|---|---|---|
| Small server room (100 kW) | 1,000 – 3,000 gallons | 0.4 – 1.1 million gallons | 10 – 30 homes |
| Enterprise center (1 MW) | 10,000 – 25,000 gallons | 3.6 – 9 million gallons | 100 – 250 homes |
| Mid-size colocation (10 MW) | 100,000 – 250,000 gallons | 36 – 90 million gallons | 1,000 – 2,500 homes |
| Large hyperscale (50 MW) | 500,000 – 1.3 million gallons | 180 – 470 million gallons | 5,000 – 13,000 homes |
| Mega campus (100+ MW) | 1 – 5 million gallons | 365 million – 1.8 billion gallons | 10,000 – 50,000 homes |
Nationally, the picture is just as striking. United States data centers consumed roughly 66 billion gallons of water directly in a recent reporting year, and that figure climbs sharply when you add the water evaporated at power plants generating their electricity. Some analyses put the combined direct plus indirect figure near 200 billion gallons annually for U.S. facilities alone.
Where All That Water Actually Goes Inside the Building
Water does not disappear into the servers themselves. In almost every case, water leaves the building as vapor through a cooling tower or evaporative cooling unit. Understanding the path helps explain why the numbers get so large.
Here is the basic journey water takes through a typical water-cooled data center:
- City water or well water enters the facility and passes through filtration and chemical treatment to prevent scale and bacteria.
- Treated water flows into a chilled water loop or directly into cooling towers.
- Warm air from the server halls transfers heat into the water through coils or heat exchangers.
- The now-hot water sprays over cooling tower fill material while fans push air across it.
- A portion of the water evaporates, carrying heat away into the atmosphere. This evaporation is the main consumptive loss.
- Remaining water, now more concentrated with minerals, cycles back through the loop several times.
- Once mineral concentration gets too high, operators dump that water as “blowdown” into the sewer and replace it with fresh makeup water.
Roughly 80 percent of the water a cooling tower takes in evaporates. The other 20 percent leaves as blowdown wastewater. That is why water experts separate two different measurements: withdrawal, which is total water taken from a source, and consumption, which is water that never returns to the local watershed. Data centers consume a much higher percentage of what they withdraw than power plants do, because power plants often return most of their water to rivers.
Beyond Cooling: Smaller Water Uses
Cooling dominates, but it is not the only draw. Facilities also use water for humidification, since server halls need controlled humidity to prevent static discharge and corrosion. Bathrooms, kitchens, landscaping, and fire suppression systems add a small slice too. Together these non-cooling uses usually account for less than 5 percent of total consumption at a large site.
Cooling Methods and How Dramatically They Change Water Use
The single biggest factor in a data center’s water bill is the cooling design its engineers chose. Two identical buildings with identical server loads can differ by 100 times in water use depending on this one decision. Here is how the main approaches stack up.
Evaporative and Cooling Tower Systems
These are the water hogs, and also the electricity savers. Cooling towers use evaporation to shed heat, which is thermodynamically brilliant but water-hungry. A 20-megawatt facility using open cooling towers might evaporate 250,000 to 400,000 gallons a day during summer. Operators pick this design because it slashes power bills, and in many regions electricity costs more than water.
Air-Cooled and Chiller-Based Systems
Air-cooled chillers use refrigerant loops and fans instead of evaporation. They consume almost no water on site, sometimes literally zero. The tradeoff is higher electricity use, typically 10 to 30 percent more energy for cooling. Since power plants also use water, the facility often just moves its water footprint upstream rather than eliminating it.
Free Cooling and Outside Air Economization
In cool climates, operators simply open the building to outside air for much of the year. Facilities in Ireland, Scandinavia, and the Pacific Northwest run on free cooling 80 to 95 percent of the time. Water use drops to near zero for most months, spiking only during rare hot spells.
Liquid and Immersion Cooling
Direct-to-chip liquid cooling pipes coolant straight onto processors. Immersion cooling submerges entire servers in non-conductive fluid. Both use closed loops that recirculate the same liquid, so consumption is minimal after initial fill. AI hardware is pushing this technology forward fast because modern GPU racks generate far too much heat for air alone.
| Cooling Method | Water Use (On-Site) | Energy Efficiency | Best Climate |
|---|---|---|---|
| Open cooling towers | Very high | Excellent | Any, best in dry heat |
| Adiabatic / hybrid | Moderate | Very good | Mixed climates |
| Air-cooled chillers | Near zero | Fair | Water-scarce regions |
| Outside air economization | Low to zero | Excellent | Cool, dry climates |
| Direct liquid cooling | Low (closed loop) | Excellent | Any, needed for AI racks |
| Immersion cooling | Very low | Outstanding | Any |
Understanding WUE: The Metric That Makes Comparisons Possible
Engineers measure data center water efficiency with a number called Water Usage Effectiveness, or WUE. You calculate it by dividing annual water use in liters by annual IT energy use in kilowatt-hours. The result tells you how many liters of water the facility burns per kilowatt-hour of computing.
The industry average sits around 1.8 liters per kilowatt-hour, though this varies wildly. Best-in-class facilities hit 0.1 to 0.2 L/kWh. Fully air-cooled sites can report 0.0. Older or poorly located facilities sometimes exceed 3.0 L/kWh.
Here are benchmark WUE figures worth knowing:
- 0.0 L/kWh – Fully air-cooled or closed-loop facility with no evaporative cooling
- 0.1 to 0.3 L/kWh – Leading hyperscale operators using advanced controls and cool climates
- 0.4 to 0.9 L/kWh – Well-designed modern facilities with hybrid cooling
- 1.0 to 2.0 L/kWh – Industry average range for evaporatively cooled sites
- Above 2.5 L/kWh – Older designs or hot, dry locations running towers hard
Here is a practical example of how to use WUE. Say a colocation provider runs a 15-megawatt IT load at 85 percent utilization, which works out to about 112 million kilowatt-hours per year. At an industry-average WUE of 1.8, that facility consumes roughly 201 million liters, or about 53 million gallons annually. Cut the WUE to 0.4 through better cooling, and consumption drops to 12 million gallons. Same computing, one quarter of the water.
One warning about WUE. It only counts on-site water. A facility can report a perfect 0.0 WUE while its electricity comes from a coal plant that evaporates enormous volumes of water. Smart analysts use “total WUE” or “source WUE,” which adds the power generation footprint. That number often lands between 1.5 and 2.5 L/kWh even for air-cooled sites on a fossil-heavy grid.
The Hidden Water Cost of Electricity Generation
Direct water use tells only half the story. Every kilowatt-hour a data center pulls from the grid carries a water footprint from the power plant that made it. Thermoelectric plants, whether coal, gas, or nuclear, use steam turbines that need cooling, and that cooling evaporates water.
Water intensity varies enormously by generation source. This is why the electricity mix of a region matters as much as the cooling design of the building.
| Power Source | Water Consumed (gallons per MWh) | Notes |
|---|---|---|
| Nuclear (cooling tower) | 600 – 800 | Highest consumption per unit |
| Coal (cooling tower) | 500 – 700 | High withdrawal and consumption |
| Natural gas combined cycle | 200 – 300 | Roughly half of coal |
| Concentrated solar thermal | 750 – 900 | Very high in dry regions |
| Wind | 0 – 1 | Essentially zero |
| Solar photovoltaic | 1 – 25 | Panel washing only |
| Hydropower | Highly variable | Reservoir evaporation, often large |
Run the math on a 50-megawatt data center. It consumes roughly 438,000 megawatt-hours per year. On a grid averaging 300 gallons per megawatt-hour, that is 131 million gallons of indirect water use. Add 200 million gallons of direct cooling water and total consumption approaches 331 million gallons annually. Switch that same facility to wind and solar power, and indirect use falls near zero.
This connection explains why major operators buy renewable energy so aggressively. Cutting carbon and cutting water go hand in hand. A facility running on wind power with air-cooled chillers can genuinely approach a near-zero water footprint, something impossible with a coal-powered grid no matter how clever the cooling design.
How AI Workloads Changed the Water Equation
Traditional servers drew 5 to 10 kilowatts per rack. AI training racks packed with GPUs draw 40 to 130 kilowatts, and next-generation designs push past 200 kilowatts. That heat density broke air cooling, and it multiplied water demand at facilities that stuck with evaporative towers.
Researchers have tried to estimate water use per AI interaction. Early studies suggested a conversation of 20 to 50 questions with a large language model consumed around 500 milliliters, roughly a small bottle of water, counting both on-site cooling and power generation. Newer estimates from model providers put a single short query closer to 0.3 milliliters, since efficiency improved dramatically and hardware got better. The truth depends heavily on model size, data center location, and whether you count power plant water.
Training is where the big numbers appear. Estimates for training one large language model range from 700,000 liters to several million liters of water, depending on the facility. One widely cited study calculated that training a GPT-3 scale model in a Microsoft U.S. data center evaporated about 700,000 liters of clean freshwater. Newer, larger models require considerably more compute, though efficiency gains partly offset the growth.
Here is what AI demand means practically for water:
- Higher rack density forces a shift toward liquid cooling, which actually reduces water use compared to evaporative towers
- Total facility power grows, which raises indirect water use at power plants
- Inference at massive scale now outweighs training in total resource use for popular consumer AI products
- New AI campuses cluster in specific regions, concentrating local water demand rather than spreading it
- Operators increasingly site AI facilities near abundant power and cool climates to control both bills
The irony worth noting: AI is pushing the industry toward closed-loop liquid cooling, which on a per-chip basis uses far less water than the evaporative systems it replaces. The problem is scale. Even efficient cooling multiplied across gigawatts of new capacity adds up.
Real-World Examples From Major Operators
Public disclosures give us concrete numbers rather than estimates. Here is what the largest operators actually report and what those figures mean.
Google’s global data center fleet consumed roughly 6.1 billion gallons of water in a recent reporting year across dozens of campuses. The company reports a fleet-wide WUE near 1.0 L/kWh and claims to replenish more water than it consumes in several watersheds. Its Council Bluffs, Iowa campus and The Dalles, Oregon facility have drawn local attention for their withdrawals, with The Dalles using around 355 million gallons in one year, roughly a quarter of the town’s total water supply.
Microsoft
Microsoft reported total water consumption around 1.7 billion gallons across its operations in a recent year, up sharply as AI capacity expanded. The company committed to “water positive” operations and has begun deploying closed-loop designs in new facilities that use zero water for cooling after initial filling.
Meta
Meta reports one of the better WUE figures in the industry, around 0.20 L/kWh, thanks to heavy use of outside air economization and indirect evaporative cooling. Its total consumption still exceeded 800 million gallons in a recent year given its fleet size.
Amazon Web Services
AWS reports a WUE around 0.15 L/kWh and pursues water positive goals by funding recycling and replenishment projects. The company uses recycled municipal water at multiple sites, including facilities in Virginia and Oregon.
A Community Case Study
In Mesa, Arizona, city officials approved data center projects with water allocations of roughly 1.25 million gallons per day, in a state where the Colorado River supply keeps shrinking. Similar tensions have played out in Chile, Uruguay, the Netherlands, and Georgia, where residents pushed back against permits during drought. These conflicts pushed several operators to switch new builds to air-cooled or closed-loop designs specifically to defuse local opposition.
Common Misconceptions People Get Wrong
Public conversation about data center water gets muddled fast. Sorting out the myths helps you read headlines with a sharper eye.
Myth one: all data centers use enormous amounts of water. Plenty use none at all. Air-cooled and closed-loop facilities exist in large numbers, especially in cooler regions and in newer builds designed for AI. Lumping them all together produces misleading averages.
Myth two: the water is destroyed forever. Evaporated water rejoins the water cycle as vapor and eventually falls as precipitation. The real issue is local and timing based. Water evaporated in Arizona might rain down in Missouri three days later, which does nothing for the aquifer it left.
Myth three: withdrawal equals consumption. A facility might withdraw a million gallons and consume 800,000, returning the rest. Reports that cite only withdrawal overstate the impact, and reports citing only consumption understate the strain on treatment infrastructure.
Myth four: cutting water is always good. Switching from evaporative to air cooling can raise electricity use by 20 percent or more. If that power comes from a thermoelectric plant, total water use might actually rise. The right answer depends entirely on the local grid mix and water scarcity.
Other frequent mix-ups worth clearing up:
- Data centers rarely use drinking-quality water by necessity. Many now run on reclaimed, gray, or industrial water.
- Server chips never touch water directly in most designs. Even liquid cooling uses sealed loops with treated coolant.
- Water use does not scale with data stored. It scales with power consumed, so a storage archive uses far less than an AI training cluster of the same size.
- Agriculture still dwarfs data centers in total U.S. water use, though local concentration makes data center demand feel much larger in specific counties.
How Operators Cut Water Use: Best Practices and Tools
The good news is that water reduction is one of the more solvable problems in this industry. Operators have a deep toolbox, and many strategies pay for themselves.
Design and Siting Strategies
The cheapest gallon is the one you never need. Building in a cool climate, raising allowed server inlet temperatures from 68°F to 80°F or higher, and choosing hybrid cooling systems that only spray water on the hottest days can cut consumption by more than half before a single server ships.
Water Source Substitution
Instead of using less, some operators use different water. Options include reclaimed municipal wastewater, harvested rainwater, industrial process water, condensate captured from air handlers, and even treated seawater in coastal locations. Several large campuses now run entirely on non-potable supply.
Operational Optimization
Small tweaks add up. Raising cycles of concentration in cooling towers from 3 to 6 cuts makeup water needs substantially. Better water treatment chemistry allows more recirculation before blowdown. Leak detection, real-time flow monitoring, and AI-driven cooling control systems all squeeze out additional savings.
Here is a practical priority order operators typically follow:
- Measure everything first with submeters on every water line, then calculate real WUE monthly
- Raise server inlet temperature setpoints to the top of the recommended envelope
- Maximize free cooling hours by expanding economizer operating ranges
- Increase cooling tower cycles of concentration through better treatment
- Switch to reclaimed or non-potable water sources where available
- Deploy direct liquid cooling for high-density racks
- Buy renewable power to eliminate indirect water at generation
- Fund watershed replenishment projects for remaining unavoidable use
Useful resources for anyone digging deeper include The Green Grid’s WUE standard, ASHRAE TC 9.9 thermal guidelines, the Uptime Institute’s sustainability reports, the EPA WaterSense program, and the annual environmental reports published by Google, Microsoft, Meta, AWS, and Equinix. The Lawrence Berkeley National Laboratory reports on U.S. data center energy and water use offer the best independent national estimates.
Frequently Asked Questions About Data Center Water Use
People arrive at this topic with very specific questions. Here are direct answers to the ones that come up most often.
How much water does a single Google search or AI prompt use?
A standard web search uses a tiny fraction of a milliliter. A short AI chatbot response has been estimated anywhere from 0.3 milliliters to about 25 milliliters depending on model size, facility, and whether power plant water counts. A longer conversation with a large model can approach half a liter under older estimates.
How much water does streaming a movie use?
Streaming one hour of high-definition video consumes roughly 0.2 to 0.8 liters of water across the whole chain, counting data center cooling and network power generation. A two-hour movie lands somewhere near a standard water bottle.
Do data centers use drinking water?
Many still do, because municipal supply is convenient and already treated. However, the trend is moving fast toward reclaimed and non-potable sources. Several operators now report over half their cooling water comes from non-drinking sources.
Which uses more water, a data center or a golf course?
An average 18-hole golf course in a hot climate uses about 300,000 to 500,000 gallons per day in summer. That falls right in the middle of a mid-size to large data center’s daily use. A mega campus uses considerably more than several golf courses combined.
Can data centers reuse their water?
Partially. Closed-loop and liquid-cooled designs recirculate almost everything. Evaporative systems cannot recover evaporated water, but they can recycle blowdown after treatment and reuse condensate captured from humid air.
Does cold weather really eliminate water use?
Mostly, yes. Facilities in Nordic countries or northern U.S. states run economizers for the vast majority of the year and use water only during summer peaks. Some report annual WUE below 0.1 L/kWh.
What Comes Next for Data Center Water
The direction of travel is clear even if the timeline is not. Regulation, technology, and public pressure are all pushing the industry toward lower water intensity, and several trends will shape the next decade.
Closed-loop cooling is becoming the default for new construction. Microsoft, Google, and several large colocation providers have committed to designs that use zero water for cooling in new facilities, filling the loop once and recirculating it for the building’s life. Liquid cooling adoption, driven by AI rack density rather than water concerns, accelerates this shift as a happy side effect.
Transparency requirements are tightening too. The European Union now requires data centers above a size threshold to report water use annually. Several U.S. states and counties have added water disclosure conditions to permits and tax incentives. Expect more mandatory reporting, which will replace guesswork with real numbers.
Watch for these developments over the coming years:
- Heat reuse projects that pipe warm data center water to district heating systems, apartments, greenhouses, and swimming pools
- Seawater and geothermal cooling at coastal and volcanic sites
- Two-phase immersion cooling that removes far more heat with sealed dielectric fluids
- On-site nuclear and gas generation, which shifts water demand back on site and complicates the accounting
- Water positive commitments backed by verified watershed replenishment rather than accounting offsets
- Siting decisions driven by water risk maps, pushing new builds toward water-abundant regions
One tension will not resolve easily. Data center capacity is growing faster than efficiency improves. Even if average WUE falls by half, tripling total capacity still increases absolute water use. The industry’s real test is whether efficiency gains and better siting can outpace raw growth in demand.
So how much water does a data center use? Anywhere from essentially zero to five million gallons a day, and the difference comes down to cooling design, climate, power source, and choices made years before the first server rack arrived. A large hyperscale facility typically runs through 300 million to 1.8 billion gallons annually when you count evaporative cooling, while a closed-loop, wind-powered facility in a cool climate might use less water than the office building next door. The industry average WUE of about 1.8 liters per kilowatt-hour gives you a useful yardstick, but remember that on-site numbers hide the water evaporated at power plants, which often doubles the real total.
Understanding these numbers matters because you now live in a world where digital services and local water supplies genuinely compete in some communities. The encouraging part is that this problem has real solutions already in hand: liquid cooling, reclaimed water, renewable power, smarter siting, and honest public reporting. As AI drives the biggest data center buildout in history, the operators who treat water as a design constraint rather than an afterthought will build facilities that neighbors welcome instead of fight. Keep asking for the numbers, keep checking the WUE, and expect the answers to get better every year.
