In one month during the summer of 2022, a cluster of data centers in West Des Moines, Iowa drank about 11.5 million gallons of water while helping train an early version of GPT-4. That single facility took roughly 6% of the district’s entire water supply for that month. Numbers like that explain why so many people now ask the same question: does AI use a lot of water, and should we worry about it? The honest answer sits somewhere between the scary headlines and the reassuring corporate blog posts.
Water is the quiet resource behind every chatbot reply, image generation, and recommendation engine. Servers get hot, heat has to go somewhere, and water is one of the cheapest, most efficient ways to carry it away. In this guide, you will learn exactly where the water goes, how many milliliters a single AI prompt really costs, how training differs from everyday use, why your location matters more than your prompt count, and what tech companies are actually doing about it. You will also see how AI’s thirst stacks up against beef, cotton, golf courses, and your own shower.
What People Really Mean When They Talk About AI’s Water Footprint
AI does not literally pour water on chips. The water shows up in three separate places: cooling the buildings that hold the servers, generating the electricity those servers run on, and manufacturing the chips in the first place. When you add all three together, the footprint looks big. When you look at just one, it can look tiny. That gap explains almost every argument you have ever read about this topic.
So does AI use a lot of water? Yes, AI consumes billions of gallons a year worldwide, but the total is still a small slice of national water use, and the real problem is concentration, not the global sum. A single hyperscale data center can rival a small town’s water demand, and many of them sit in dry regions where every gallon already has three claims on it. That local squeeze matters far more than the abstract global figure.
There is one more distinction that trips up almost everyone, and it changes the numbers by a factor of ten or more. Water withdrawal means water pulled from a river, lake, or aquifer, some of which returns. Water consumption means water that evaporates or otherwise leaves the local watershed for good. A power plant might withdraw enormous amounts and return most of it warm but usable. A data center cooling tower is different, because evaporation is the whole point, so most of what it takes never comes back.
Here are the categories experts use when they measure AI’s water demand:
- Scope 1 (onsite): water evaporated in cooling towers, chillers, and humidifiers at the data center itself.
- Scope 2 (offsite energy): water consumed at power plants generating the electricity the servers use, including evaporation from hydropower reservoirs and thermal plant cooling.
- Scope 3 (supply chain): ultrapure water used to manufacture GPUs, memory, and other hardware, plus water tied to construction materials.
- Withdrawal vs. consumption: two different meters that often get quoted as if they were the same number.
- Potable vs. non-potable: drinking-quality water versus reclaimed wastewater, industrial water, or seawater.
Once you keep those categories straight, most contradictory headlines stop contradicting each other. One study is counting evaporation at the building. Another is counting the power plant. A third is counting the chip factory in Taiwan.
Where the Water Actually Goes Inside a Data Center
Think of a rack of AI servers as a very expensive space heater. Nearly every watt of electricity that goes in comes back out as heat. A modern GPU rack can pump out 40 to 130 kilowatts of heat, which is like running dozens of hair dryers nonstop in a closet. If that heat stays put, chips throttle, then fail. So operators move it outside, and water is a brilliant heat mover because evaporating it absorbs a huge amount of energy.
Here is the basic process in a water-cooled facility:
- Servers heat the air or a liquid loop running right against the chips.
- That heat transfers into a closed water loop circulating through the building.
- The warm loop passes through a cooling tower or evaporative unit.
- Some water evaporates into the outside air, carrying the heat away with it.
- Operators top up the loop with fresh makeup water to replace what evaporated.
- They also flush out a share of the remaining water, called blowdown, because minerals concentrate as water evaporates and would otherwise scale up the pipes.
Steps four through six are the whole story. Evaporation plus blowdown equals real, permanent consumption. A large facility might evaporate hundreds of thousands of gallons on a hot afternoon and almost nothing on a cool, humid night.
Cooling Methods and How Thirsty Each One Is
Not all data centers cool the same way, and the choice changes water use dramatically. Evaporative and adiabatic systems use the most water but the least electricity. Air-cooled chillers use almost no water but burn more power, which pushes water use upstream to the power plant instead. Closed-loop liquid cooling circulates the same fluid over and over and only loses water if the outer loop uses evaporation.
Immersion cooling, where whole servers sit in a bath of non-conductive fluid, is the newest option and can nearly eliminate onsite water. Direct-to-chip cold plates sit in between, pulling heat straight off the hottest components with a sealed liquid loop. Because AI chips run so hot, the industry is shifting hard toward liquid cooling for performance reasons, and lower water use is a happy side effect when those loops reject heat to dry coolers instead of towers.
The Metric That Tracks It: WUE
Water Usage Effectiveness, or WUE, measures liters of water consumed per kilowatt-hour of IT energy. The industry average sits near 1.8 liters per kilowatt-hour, according to operator surveys. Google reports a fleet-wide average closer to 1.0 to 1.1 liters per kilowatt-hour, and some new air-cooled designs claim numbers near zero. If you ever want to judge a data center’s thirst, WUE is the single most useful number to ask for, though a low WUE paired with high energy use can just shift the burden to the grid.
How Much Water Does a Single AI Prompt Use?
This is where the estimates diverge wildly, and it is worth understanding why. The widely cited research from the University of California, Riverside estimated that a conversation of roughly 10 to 50 responses with GPT-3 consumed about 500 milliliters, a standard water bottle. Google later published a technical paper claiming a median text prompt to its Gemini model used about 0.26 milliliters, roughly five drops. OpenAI’s leadership has cited a figure around 0.000085 gallons per query, about one-fifteenth of a teaspoon.
Those numbers differ by a factor of a thousand or more. The differences come from newer, far more efficient chips, better cooling, models that are much cheaper to run per token, and, importantly, what each estimate counts. Company figures often focus on onsite cooling at their most efficient sites and may exclude or minimize the water used to generate their electricity. Academic figures usually include both and often assume older hardware and less efficient facilities.
| Estimate source | Water per typical text prompt | What it includes |
|---|---|---|
| Academic estimate (older GPT-3 era hardware) | Roughly 10 to 50 mL | Onsite cooling plus power generation, average US grid |
| Google technical disclosure | About 0.26 mL (median text prompt) | Onsite consumption at Google facilities |
| OpenAI public statement | About 0.32 mL per query | Onsite cooling estimate |
| Image generation (estimated) | Several times a text prompt | Higher compute per output |
| Long video generation (estimated) | Hundreds of times a text prompt | Very high compute per second of output |
A reasonable takeaway: a single text prompt today probably costs somewhere between a few drops and a few teaspoons of water, depending on the model, the data center, and whether you count electricity. Multiply that by billions of prompts a day and it becomes real. Multiply your personal usage of, say, 50 prompts a day, and you get maybe a few glasses of water. That is less than one flush of a modern toilet.
The honest framing is this: your individual AI habit is not the problem. The aggregate build-out of data centers in specific watersheds is.
Training an AI Model vs. Running It Every Day
AI has two very different water bills. Training is the one-time, enormously intensive process of teaching a model, running thousands of GPUs flat out for weeks or months. Inference is what happens every time you type a question, and each event is tiny but happens constantly.
The research that started this whole conversation estimated that training GPT-3 in Microsoft’s US data centers evaporated about 700,000 liters, or 185,000 gallons, of clean freshwater just for onsite cooling. That is roughly enough to produce a few hundred cars, or to fill about a quarter of an Olympic swimming pool. The same paper noted that training the same model in less efficient facilities in Asia could have tripled that figure. Larger, newer models take far more compute than GPT-3 did, so training footprints have grown even as efficiency per chip has improved.
Picture the West Des Moines example again. Microsoft’s cluster there supported a supercomputer used for OpenAI’s training work, and local reporting showed the site pulling about 11.5 million gallons in a single hot month. That is not a rounding error to a city water utility. It is the kind of demand that shows up in municipal planning documents and in neighbors’ conversations about lawn watering rules.
Here is the surprising part, though: over a model’s lifetime, inference usually wins the water contest. Training happens once. Inference happens billions of times a day for years. Industry estimates suggest inference now accounts for somewhere between 60% and 90% of the total energy, and therefore water, tied to a popular model. That flips the common assumption that training is where all the damage happens.
- Training: concentrated, visible, happens in a handful of sites, measured in hundreds of thousands to millions of gallons per model.
- Inference: diffuse, continuous, spread across many regions, and growing every time a new feature ships.
- Fine-tuning: much smaller than full training but repeated often across many customers.
- Idle and redundancy: standby capacity, cooling, and humidity control still consume water even at low utilization.
The Hidden Water in Electricity and Chip Manufacturing
Even a data center that uses zero water onsite still has a water footprint, because electricity itself is thirsty. Thermal power plants, whether coal, gas, or nuclear, boil water to spin turbines and then cool the steam. Hydropower reservoirs lose staggering volumes to evaporation off their surfaces. On average, generating a kilowatt-hour in the United States consumes on the order of 1 to 2 liters of water, though the range is enormous depending on the mix.
That single fact reframes the whole debate. A facility that brags about air cooling has simply moved its water use from the cooling tower to the power plant, unless it also runs on wind or solar. Wind and solar photovoltaic generation consume almost no water during operation, which makes clean-energy contracts one of the most powerful water-saving tools a data center operator has, even though people rarely describe them that way.
Then there are the chips. Semiconductor fabrication is one of the most water-intensive manufacturing processes on Earth, because it requires ultrapure water to rinse wafers between dozens of process steps. A single advanced fab can use millions of gallons per day, and leading foundries in Taiwan have faced production risk during droughts serious enough that the government trucked in water and restricted farm irrigation. Every GPU in an AI cluster carries a slice of that footprint before it ever powers on.
To put the national picture in perspective, a Lawrence Berkeley National Laboratory analysis estimated US data centers directly consumed roughly 66 billion liters, about 17 billion gallons, of water in a recent year, with indirect water tied to their electricity running several times higher. Google alone reported consuming more than 6 billion gallons in a year across its operations, and Microsoft has disclosed water consumption in the billions of gallons as well, with both figures rising sharply as AI capacity grows. Those totals are still under 1% of US water use, but they are climbing at double-digit rates while agriculture stays flat.
Putting AI’s Thirst in Context With Everyday Water Use
Comparisons help, as long as you do not use them to dodge the issue. Agriculture accounts for roughly 70% of global freshwater withdrawals. Thermoelectric power and irrigation dominate US withdrawals. Data centers, including all AI workloads, remain a small national share, though they are the fastest-growing industrial user in several regions.
| Activity | Approximate water footprint |
|---|---|
| One AI text prompt | A few drops to a few teaspoons |
| One 8-minute shower | About 65 liters / 17 gallons |
| One load of laundry | About 60 to 100 liters / 15 to 25 gallons |
| One quarter-pound beef burger | About 2,400 liters / 630 gallons |
| One cotton t-shirt | About 2,700 liters / 700 gallons |
| One pair of jeans | About 7,500 liters / 2,000 gallons |
| Training a large language model (onsite cooling) | Hundreds of thousands to millions of liters |
| One golf course, per day in a dry climate | Over 1 million liters / 300,000 gallons |
Read that table honestly and two things pop out. First, skipping one burger saves more water than thousands of AI prompts. Second, a single training run still equals hundreds of burgers, and one large data center campus can rival a golf course or a small town year-round. Both statements are true at the same time.
The useful conclusion is not “AI is fine” or “AI is a disaster.” It is that individual guilt is the wrong lens. Where the servers sit, what powers them, and how they cool matter far more than how many questions you ask a chatbot this week.
Myths and Misunderstandings Worth Clearing Up
This topic generates confusion faster than almost any other tech story, partly because the numbers are genuinely slippery and partly because both critics and companies have reasons to pick flattering framings. Let us straighten out the most common mix-ups.
- Myth: The water is destroyed. Water never disappears. Evaporated water returns as rain, just usually not in the same watershed or the same season. The local loss is real even though the global stock is unchanged.
- Myth: Every AI query wastes a full bottle of water. That figure came from an early estimate covering a whole conversation on older hardware, and it spread far beyond its original context. Current per-prompt numbers are far smaller.
- Myth: Water use and energy use are the same problem. They trade off. Cutting cooling water often raises electricity use, which raises water use at power plants. Good design optimizes both together.
- Myth: All data centers use drinking water. Many operators now run on reclaimed wastewater, industrial water, or seawater. Some sites use zero potable water for cooling.
- Myth: Withdrawal equals consumption. A facility that withdraws 10 million gallons but returns 9 million to the same river has a very different impact than one that evaporates all 10 million.
- Myth: It is only about AI. Cloud storage, video streaming, search, gaming, and enterprise software share the same buildings and the same cooling systems. AI has accelerated growth, but it did not create data center water use.
One more nuance deserves attention. Corporate “water positive” pledges usually mean replenishing more water than a company consumes, often through wetland restoration or irrigation efficiency projects. Those projects can be genuinely valuable, but replenishing a river in one state does not help a town in another state whose aquifer is dropping. Watershed matching matters, and the best disclosures now report it by region rather than as one global number.
Real Places Where AI Water Use Became a Local Story
Abstract debates get concrete fast when a data center lands in your town. Several cases show what happens when large compute demand meets a stressed water supply, and they are the best evidence for why location beats every other variable.
- The Dalles, Oregon. After a legal fight over public records, reporting revealed that Google’s data centers there used nearly a quarter of the city’s total water, with plans to expand. The city and company later negotiated agreements involving water rights and infrastructure upgrades.
- Montevideo, Uruguay. During a severe drought in 2023, residents protested a planned Google data center as tap water quality dropped and reservoirs shrank. The project was eventually redesigned to use air cooling.
- Santiago, Chile. Community groups challenged a data center proposal in a drought-hit region, and courts required a fresh environmental review. The project was later revised toward air cooling as well.
- Maricopa County, Arizona. Several hyperscale campuses operate in a region under formal Colorado River shortage conditions, prompting cities to require water-efficient designs or reclaimed water commitments in exchange for approvals.
- West Des Moines, Iowa. The AI training cluster mentioned earlier prompted the local utility to signal it would not approve new data center projects unless they demonstrated cutting-edge water conservation technology.
- Memphis, Tennessee. A large AI supercomputer build-out drew scrutiny over aquifer use, leading the operator to fund a greywater recycling facility to supply cooling water instead of drinking water.
Notice the pattern. In nearly every case, public attention pushed the project toward air cooling, reclaimed water, or reduced draw. Transparency changed outcomes. That is why researchers push so hard for facility-level, watershed-level disclosure rather than one global corporate total.
There is also a fairness dimension. Data centers often land in rural or lower-income areas that offer cheap land, cheap power, and tax incentives. Those same communities frequently have the least capacity to fight for water rights or to fund new treatment infrastructure. The benefits, meanwhile, flow to users everywhere. Balancing that trade is now a core part of siting negotiations.
How Companies Are Cutting AI’s Water Use, and What Comes Next
The good news is that water use per unit of AI work is falling fast, and the engineering levers are well understood. Operators do not need a breakthrough. They need to deploy what already exists and to report honestly on results.
The most effective practices in use today include:
- Closed-loop and direct-to-chip liquid cooling that circulates the same fluid for years instead of evaporating fresh water.
- Reclaimed and non-potable water sourcing, including treated municipal wastewater and industrial process water.
- Air-cooled designs in cool or humid climates, where outside air does the work most of the year.
- Higher server inlet temperatures, since letting the room run a few degrees warmer sharply cuts cooling demand.
- Seasonal and time-of-day flexibility, running evaporative cooling only during extreme heat and switching to dry cooling otherwise.
- Renewable power contracts, which slash the invisible water tied to electricity generation.
- Waste heat reuse, piping warm water into district heating networks, a practice already used in parts of Denmark, Finland, and France.
- Workload shifting, moving flexible training jobs to regions and hours where water and power are least stressed.
- Model efficiency work, including distillation, quantization, caching, and smaller task-specific models that answer routine questions without a giant model.
That last item may be the biggest lever of all. Routing a simple question to a small model instead of a frontier one can cut the compute, and therefore the water, by an order of magnitude. Providers already do this behind the scenes, and it improves speed and cost at the same time, which means the incentives point the right direction.
Looking ahead, expect four shifts. First, mandatory disclosure is coming, with the European Union already requiring data center reporting on water and energy and several US states considering similar rules. Second, liquid cooling will become standard rather than exotic, because AI chip densities leave no alternative. Third, siting will move toward cold climates, coastal areas with seawater cooling access, and places with abundant reclaimed water. Fourth, water will start showing up as a line item in procurement, so enterprise buyers can compare providers on liters per million tokens the way they compare price and latency today.
None of this means the total goes down automatically. Efficiency gains have a way of getting eaten by growth, a pattern known as the rebound effect. If AI demand triples while water per query drops by half, the total still climbs. Watching absolute consumption, not just efficiency ratios, is how you tell whether progress is real.
Quick Answers to Common Questions About AI and Water
Should I feel guilty about using ChatGPT?
Not really. If your daily AI use costs a few teaspoons to a few cups of water, one skipped burger or one shorter shower per week more than covers a year of prompts. Your leverage is much stronger as a voter, employee, or customer asking companies about siting and disclosure than as an individual rationing prompts.
Do image and video models use more water than text?
Yes. Water tracks compute, and generating an image takes far more compute than a short text reply. Video generation is dramatically heavier still, since the model produces many frames that must stay consistent. If you want to reduce your footprint, generating fewer unnecessary videos matters far more than typing fewer questions.
Is running AI on my own laptop better for water?
Sometimes, but not always. Your laptop uses no cooling water directly, though your electricity still carries water from the grid. Large data centers run far more efficiently per unit of work, so a hyperscale facility on renewable power with reclaimed water cooling can beat local hardware on total water. Small local models for simple tasks are a reasonable choice regardless.
Which uses more water, AI or streaming video?
Streaming has historically used far more total data center capacity than AI, so its aggregate footprint has been larger. AI is closing the gap quickly because each AI request is much more compute-intensive than delivering a cached video file. Within a few years, AI workloads are expected to dominate new data center water demand.
How can I find out about a data center near me?
Start with your local water utility’s annual report and large-customer disclosures, then check city or county planning documents for development agreements, which often specify water commitments. Company sustainability reports increasingly break out consumption by region. If a facility refuses to share numbers, that itself is useful information for public comment periods.
Will AI ever run without water?
Fully water-free operation is technically possible today using closed-loop liquid cooling, dry coolers, and renewable electricity. The catch is that dry cooling costs more energy and money, especially in hot climates. As disclosure rules tighten and water gets scarcer, more operators will pay that premium, and some already advertise designs that consume no water for cooling at all.
So, circling back to the question that started all this: AI does use a lot of water, but the phrase “a lot” needs context. Globally, AI’s share of freshwater use is small next to farming, power generation, and everyday household demand. Per prompt, the cost is a few drops to a few teaspoons. Yet a single training run can evaporate hundreds of thousands of liters, one campus can rival a small town’s demand, and the fastest growth is landing in some of the driest places on the map. Both truths belong in the same sentence.
The most useful thing you can do with this information is shift your attention from personal guilt to system design. Ask where data centers get built, what powers them, whether they use reclaimed water, and whether operators publish facility-level numbers instead of comforting global averages. The technology to run AI with dramatically less water already exists, and public pressure has repeatedly pushed projects toward those better designs. Keep asking the question, keep reading the local numbers, and AI’s water story can end up as an example of an industry that fixed a problem early rather than one that ignored it until the wells ran dry.
