What's the Impact of AI (Part 3): How much water does AI use?

This post first appeared in our weekly Make Waves Mondays email series on September 8, 2026.

---

Hello hello friend!

As promised, we are back today with Part 3 in our Impact of AI series 💪

If you haven’t read the first two parts in this series yet, you can read both Part 1 and Part 2 over on our blog. I’d definitely recommend reading (or listening to!) both of those before this one! Part 1 lays the foundation for the rest of the series, and Part 2 covers the energy consumption of AI.

Today’s Make Waves Monday is going to be all about the water consumption and water impacts of artificial intelligence.

And similar to last week’s MWM with energy consumption, finding exact water consumption data for AI was essentially impossible, and even finding estimates proved challenging, which did end up making sense the more I got into the research and began to understand how exactly water is used in relation to AI and data centers, and we’ll dive deeper into why that is very shortly.

But in the meantime, I want to start off with a ‘lil Editor’s Note, similar to last week’s:

All of the information included in this Make Waves Monday, including calculations and estimations, is based on my research and data analysis across dozens of articles and studies, using the best available data we could find. These calculations and estimations should be viewed as contributory insights rather than definitive facts. There are many assumptions that need to be made when conducting an analysis like this, and I’ve done my best to make clear those assumptions throughout this series. Given how much this analysis relies on various assumptions and how quickly Artificial Intelligence is developing, the numbers included today may be out of date tomorrow, but the goal of this MWM is to provide context and overarching trends surrounding the environmental impact of AI, based on the most recent available data.

And so with that…let’s dive in 👇

How does AI use water?

Okay so before we can get into how much water AI and data centers use, we first need to look at how they use water. And it’s…a lot.

But first, a couple of definitions:

💡 Water withdrawal is “freshwater taken from the ground or surface water sources, either temporarily or permanently, and then used for agricultural, industrial, or municipal uses.” (4)

💡 Water consumption is “the amount of water ‘evaporated, transpired, incorporated into products or crops, or otherwise removed from the immediate water environment.’ Water consumption reflects the impact on downstream water availability.” (4)

So basically — water withdrawal returns the water to the water cycle, whereas water consumption removes water from the water cycle.

Okay now that we’ve got those definitions sorted, there are three “scopes” of water use for data centers:

Scope 1: On-Site Data Center Cooling

Every time you type a prompt into an AI chat bot, you’re triggering thousands of calculations within the server to “determine the best words to use.” (Because again, as we learned from Part 1, Artificial Intelligence isn’t actually “intelligent”; it’s using math to string words together that sound plausible.) 

All of these calculations produce heat, which needs to be cooled in order for the systems to continue running properly. This heat gets transferred away from the servers using either air or liquid cooling methods, though liquid is much more common. (3, 4)

Most liquid cooling is kinda like sweating — fresh water transports heat from a server to a cooling tower and out of the building through evaporation.

On average, data centers with this type of evaporative, or open-loop, liquid cooling system consume anywhere from 1 to 9 liters* of water per kilowatt-hour (kWh) of server energy that’s produced. 

*1 to 9 is a really large window, so for context: the 1 liter per kWh figure is based on “Google’s annualized global on-site water efficiency,” while the 9 liters per kWh figure is based on “a large commercial data center during the summer in Arizona.” (4)

This kind of system is actually the more energy-efficient system, and typically used in areas with strained power grids. But on the flip side, it obviously requires a lot of water, which is often drawn from municipal water sources. (4, 87)

Alternatively, closed-loop cooling systems operate by recirculating the same water (with some additions) through heat exchangers or cooling loops. In these systems, only about 5–10% of the water that’s used is consumed through evaporation or leaks, which reduces overall water use by about 70% compared to open-loop systems, and are therefore often used in areas with limited water supply. However, these closed-loop systems require much more energy use than open-loop systems. (87)

It’s a terrible catch-22 where the less water we use, the more energy we use, and the less energy we use, the more water we use.

Now, if a data center is located in a cooler region, that data center may “use ‘free’ outside air to directly reject the heat to the outside environment.” (4, 87) While outside air cooling is more energy-efficient than liquid cooling, when the outside air temperature rises above ~85°F, the data center then needs to employ water evaporation cooling systems. And if the outside air is too dry, additional water is then needed to add humidity control. All-in-all, about 70% of this water is consumed, and can add stress to the local water supply on the very same days that it’s needed most by the local community. (4)

This whole situation feels a whole heck of a lot like what we talked about a few weeks ago with regards to the climate change impacts of air conditioning… A space needs to be cooled down, which is done by moving heat from inside to outside, which in turn causes the local environment to heat up.

Scope 1 water use is the most commonly-reported metric for AI water use — especially in reports from the AI companies themselves — yet on-site cooling is only responsible for about 15% of the water use related to an AI query. (92)

Scope 2: Off-Site Electricity Generation

As we know from Part 2, data centers need constant electricity to keep them running every hour of every day, and many regions around the world, including most of the United States, rely on thermoelectric power — which includes coal, oil, gas, and nuclear. (98) All of that power generation requires immense amounts of water to cool the equipment, similar to the data center cooling itself. (4) According to a 2019 study out of Duke University, about 40% of all water use in the US goes to cooling thermoelectric plants. (99)

On average, in the US, water withdrawal for energy generation purposes is about 43.8 liters per kilowatt-hour, and water consumption is about 3.1 liters per kilowatt-hour. (4)

But it’s also important to remember here that data centers are often built in areas with dirtier energy grids, which would increase the rate of Scope 2 water consumption in these areas. 

According to that same Duke study, solar and wind energy only use about 1–2% of the water that coal or natural gas use. (99) So while current data center power is likely pulling more than 3.1 liters of water per kilowatt-hour (and in fact, Meta has reported that its Scope 2 water consumption is about 3.7 liters per kilowatt hour), if our entire electric grid was renewable, that number would drop to just about 31–62 milliliters per kWh.

It’s alllllll connected.

Scope 3: Server Manufacturing in the Supply Chain

While the numbers aren’t as readily available for this one, producing the chips and servers needed for AI to operate requires huge amounts of water, which often contains toxic chemicals and/or hazardous wastes when discharged from the manufacturing plants. (4)

And the rate of water recycling for Scope 3 water is very low.

For example, in Singapore, the water recycling rate at wafer plants is about 45%, and at semiconductor plants just 23%. (4)

Apple has also reported that Scope 3 water use accounts for about 99% of the company’s entire water footprint (4), and it’s probably pretty safe to assume that a similar number is true for most tech companies.

How much fresh water is being used for AI?

In 2021, before ChatGPT was released, all of the data centers in the United States consumed about 449 million gallons of water per day. (87)

And while I couldn’t find a straightforward number like this for 2026, what I could find was that by 2028 — just barely more than one year from now — data centers in the United States are projected to consume nearly 74 billion gallons (280 billion liters) of water per year. (86) That’s an increase of more than 16,000% in 7 years.

About 57–90% of all of that water is coming from lakes, rivers, and aquifers — aka the same places that we get our drinking water from. And making matters worse…data centers tend to be located in areas that are already experiencing water stress, which means that the substantial increase in water consumption as a result of increased AI use will worsen prolonged droughts that we’re already seeing. (4, 18, 22, 67, 80, 87) Some people are already finding that their taps are literally running dry because data centers are using all of the groundwater where they live. (20)

Now, some data centers do use recycled water, rather than freshwater, but that’s still very much the exception, not the norm.

Amazon, for example, uses recycled water in about 24 of their US data centers, and announced last year that by 2030, they plan to use recycled water at about 120 of their data centers in the US. (88) Now, Amazon doesn’t publicly release how many data centers they currently have, but leaked data from last fall suggests that at that time they had more than 900. (89)

Yes, using recycled water at 120 data centers is better than using recycled water at 24 data centers, but when they already have 900 data centers and at the rate new data centers are being built…that’s such a small number. And Amazon has about $267 billion committed to new AI projects that isn’t even shown anywhere on their balance sheets — so they’re planning some massive growth in the next few years. (62)

Google has also publicly set a goal to “replenish” 120% of the water that their AI and data centers consume (3), but according to Google’s sustainability report, in 2025, they only "replenished" about 78% of their consumption. (85)

But what is “water replenishment”? Because water is a finite resource. So how are they possibly planning to “replenish” more water than they use?

According to Google’s website, they’re not actually “replenishing” water (because wtf does that even mean?), but rather they’re funding water conservation projects. This includes things like installing toilet leak detectors, planting native plants, and improving irrigation efficiencies. (78)

And again… YES. These are important projects and it’s good that they’re being done and I’m glad that money is being invested into these valuable and important projects.

But that is not replenishing anything. That’s an offset. Google’s data centers consumed 6.1 billion gallons of water in 2023, and then nearly doubled that in just two years to consume 10.9 billion gallons in 2025. (77) And they will absolutely be keeping up this pace because they have over $900 billion committed to new AI projects not on their balance sheets. (62)

It’s wildly unsustainable, and no amount of detected toilet leaks are going to fix the problem.

How do data centers pollute water?

It’s probably safe to assume that you’ve seen at least one headline about unsafe drinking water in areas with data centers, but if data centers need fresh water for cooling, how exactly is that causing pollution in the drinking water system?

In most cases, the water pollution is caused by chemicals that are added to the water after it’s been withdrawn from the water supply. Because water is corrosive, these data centers need to add anti-corrosion chemicals and biocides to the water used in the cooling systems, and then as the water moves through the cooling process, it picks up minerals and heavy metals, which further contaminate the water.

If the water isn’t treated properly when it’s discharged into the wastewater system, it can put an increased burden on the local water treatment plants, or just directly pollute waterways, depending on how the wastewater system works in that area. (87)

And in some areas, that’s just the tip of the iceberg.

Take Oregon’s Morrow County, for example…

All of the following about Morrow County comes from source [80] — a Rolling Stone article published last fall. It’s quite long, but I ~highly~ recommend giving it a read, especially if you’re interested in learning more in-depth about the shady backdoor deals that are often going on to get these data centers built and how they truly affect communities and public health, while only benefiting a small handful of individuals.

As we’ve already established, oftentimes, water-hungry data centers are built in areas already experiencing water stress, and Morrow County is no exception. Every year, Morrow County experiences periods of drought. It’s essentially a desert, but the extra water use from the data centers isn’t even their biggest problem — it’s the pollution.

In the 1990s, a large irrigation system was built in the area, which meant that “large agricultural companies could use fertilizers in fields year-round.” As a result, some of the largest agricultural companies moved in, including Lamb Weston, who basically supplies all of the potatoes for McDonald’s french fries, Tillamook, and one of the largest dairy operators in the country.

All of that fertilizer resulted in the contamination of the Lower Umatilla Basin. In 1992, the average nitrate concentration in the basin was 9.2 parts per million (ppm), and by 2015 that concentration had risen to 15.3 ppm, with some areas as high as 73 ppm.

The state’s established safe limit is 7 ppm.

Basically, every day, these megafarms and food processing plants send millions of gallons of water to the Port of Morrow, which then pumps it into several lagoons outfitted with systems to trap the solids, turn them into methane gas, then burn off the methane. Through this process, the water that’s left is rich in nitrogen, and is pumped back to the farms.

“It’s a novel recycling process that alleviates the Port’s wastewater burden and offers farms a steady flow of highly concentrated fertilized water to expand their industrial-ag footprints.”

The crops absorb as much as they can, but they can only take up so much water, and the rest seeps through the soil and pollutes the aquifer below.

In 2011, Amazon opened up a 10,000 square foot data center in Morrow County to service Amazon Web Services (AWS). They were given a 15-year tax abatement for each data center they built in the area (a deal worth billions for Amazon), and since that deal was made, Amazon has constructed 7 more hyperscale data centers in the area, with agreements to build 5 more.

Amazon’s data centers have exacerbated the water pollution problem in Morrow County.

The data centers remove tens of millions of gallons of water from the aquifer every year to cool the equipment, and then discharge it into the Port’s wastewater system, which in turn means the Port needs to discard the water from the lagoons into the fields more often, which amplifies how much nitrogen goes back into the aquifer.

To deal with this excess of wastewater, the Port has done things like irrigate the fields during the winter months when nothing is growing, though in October 2025 they said they’d stop doing that. But not before giving the Oregon governor an ultimatum less than a year prior, saying that if the governor didn’t allow them to keep dumping nitrate-laden water on dormant winter crops and contaminating the water supply, that they’d furlough thousands of employees and cause “substantial economic harm to the region and the State of Oregon.” The governor granted the request.

Because of this ongoing snowball effect causing constant increases in pollution, eventually the water that Amazon was pulling from the deepest parts of the aquifer — that was supposed to be clean water — was polluted with nitrate levels as high as 13 ppm.

And adding to this snowball effect is that through the cooling process, as we know, some of the water evaporates, which increases the nitrogen concentration. In some instances, the nitrogen concentration in this discharged water is as high as 56 ppm — 8 times higher than Oregon’s safety limit.

Two county commissioners, Jim Doherty and Melissa Lindsay, recognized this was a problem when Jim began to notice that his neighbors began experiencing miscarriages and cancer rates like he’d never seen before. Doherty and Lindsay fought for the state to declare a state of emergency, which the governor did, and sent about $800,000 in emergency funds (though the commissioners requested $4 million). These emergency funds were used to hire water trucks so residents with wells could fill large jugs of water, as well as to distribute bilingual flyers to residents about the contaminated waters, letting them know that boiling the water would not help.

State agencies and large farm operators argued that it wasn’t actually an emergency, and that the burden of solving the problem should fall to “anyone who uses water or land in this area.” These same farm operators were also telling their employees (many of whom were undocumented immigrants) that the emergency declaration would put their jobs at risk, in order to keep them from speaking up about it.

And at a town hall meeting in 2022 to discuss the problem, no one from the Oregon Health Authority or the Department of Environmental Quality (DEQ), or the governor showed up, despite being invited and the fact that hundreds of homes in the area had nitrate levels above the state’s established safety limit.

Amazon contributed money towards the cause, but upon reviewing hundreds of emails that were obtained through a public records request, Rolling Stone found ample evidence that Amazon was only donating the money to make themselves look good in order to get $2 billion in tax abatements, including this gem of an email:

“Our absence…will be noted not only by the public but also very important stakeholders who 1) hold the keys to $2B of tax abatements currently in negotiation for the business 2) gate keep for key land acquisitions to meet supply needs, and 3) facilitate our permitting and water/fiber infrastructure approvals. All of which are very critical for our continued success.”

To which Rolling Stone had this to say:

“It was in their best interest to curry favor with the public, particularly if it meant they could avoid any commitment to altering the wastewater management practices that continued to exacerbate the crisis.”

It all just really brings this to mind…

At the time of the Rolling Stone article’s publication, the Port was awaiting a federal grant approval that would help them construct additional wastewater storage facilities to help reduce the nitrate levels in the wastewater.

In the meantime, Amazon is paying the Port to cover the cost of permit violations issued from the DEQ and “making contributions to the community, including $850,000 to the SAGE center, a local museum that features an Amazon learning environment called the AWS Think Big Space.”

As I mentioned above, all of this nitrate contamination has had real health effects on the community — from miscarriages to cancers.

One community member had his voice box removed due to a cancer that only smokers get, but he’d never smoked a day in his life.

Commissioner Dougherty’s wife told Rolling Stone, “There’s 14 people that live on my road, and I think nine of them have cancer right now.”

The Dougherty’s son lives a few miles away on land he and his wife just purchased in 2018 in a house they built new, with a well that they drilled that hadn’t tested positive for nitrates when they moved in. Two years later, they suffered a miscarriage. When his mom suggested testing their water, the son responded, “Mom, we just built the place, it’s all undetectable.” She told him to “check the goddamn water.”

It came back at 27 ppm, almost 4 times higher than the Oregon limit.

How much water does a single AI query use?

Okay so as I mentioned at the top, finding accurate (or even estimate) numbers for the amount of water used by one AI query was almost impossible to find.

We’ve all heard the claim that “a single ChatGPT prompt uses a bottle of water,” but how true is that, really?

Well. The reason that these numbers are so impossible to find is that the amount of water used depends HEAVILY on where the data center is located, what type of cooling system the data center uses, and even what’s included in the reported number (i.e. Did the author of that number include Scope 3 water consumption?, etc.).

But here’s what I do know:

• Google claims that each Gemini prompt uses about 0.26 mL of water (63), and Sam Altman claims that each ChatGPT query uses about 0.3 mL of water (24, 90). However, both of those estimates only include Scope 1 water usage (90), and only about 15% of the water used for an AI query is attributable to on-site cooling. The other 85% is from training and electricity generation. (92)
  
  
  • If we extrapolate Sam Altman’s 0.3 mL claim to include training and electricity, that number increases to 2 mL per query.
    
    
  • Similar to Google and Sam Altman’s energy use numbers, I have a very hard time actually believing that these are accurate.
    
    
• According to a 2025 study out of UC Riverside, in the United States, every 29.6 ChatGPT-3 queries uses about 500 mL of water, and each query uses about 16.9 mL. In Washington State, each query uses about 47.5 mL of water, so it only takes about 10.5 queries to use 500 mL of water. (4)
  
  
  • These numbers from UC Riverside do not include the training phase of AI, but they do include both Scope 1 and Scope 2 water use.
    
    
  • These numbers are also based on GPT-3, and as we know from Part 2, we’re on GPT-5.5 now, which uses significantly more energy than GPT-3, or even GPT-4 (which our calculations were based on in Part 2). Water used for cooling tracks closely with energy use, so as energy use goes up, so does water use. (92)
    
    
• According to the Washington Post, if one in ten working Americans use ChatGPT-4 just once per week to write one single email, over a year, that alone would consume 435 million liters of water. This number includes water use across the entire AI lifecycle. (92)
  
  
• According to a United Nations University report, creating one AI image uses about 29 mL of water, and creating a “complex AI video” uses about 4,100 mL. (96)

So given all of these numbers, and what we know from our overview of the different ways that AI consumes water, I’ve calculated my own estimates in alignment with the calculations we did in Part 2 for energy use:

• In Part 2, we calculated that a single ChatGPT-4 query consumes about 4.28 watt-hours of energy. Again, this is a conservative estimate, especially considering that we’re now on GPT-5.5.
  
  
• Above, we learned that data centers consume anywhere from 1 to 9 mL of water per watt-hour of energy for cooling purposes.
  
  
• We also learned above that on average in the United States, energy generation consumes about 3.1 mL of water per watt-hour of energy. This is also a conservative estimate, since data centers tend to be located in areas with dirtier grids that require more cooling.

Pulling these numbers together, we can calculate that one ChatGPT-4 query consumes anywhere from 17.6 to 51.8 mL of water.

How does AI’s water consumption stack up against other everyday activities?

So here’s the thing about AI query water consumption: it’s pretty low on the scale of individual water use for everyday actions. 

And you often see that as an excuse to “not worry about it,” or deflect the conversation around AI to something like, “yeah well if you don’t like AI then you shouldn’t eat meat either!” 

I even recently saw someone on Threads try to claim that “no one” is talking about how much water beef production uses and like…hello?? Have you literally ever once talked to a vegan, a vegetarian, or any environmentalist? It’s talked about A LOT. 

And yeah, it’s also a problem.

So if you’ll allow Sassy Krystina™ to have her moment…

I am so flippin’ tired of seeing these red herring fallacies everywhere. I’m tired of seeing so many articles talking about the massive scale of impact that AI has and then closing with, “Yeah man it’s a lot, but hey your individual use doesn’t really matter because it’s such a small number.”

OUR INDIVIDUAL ACTIONS ARE WHAT ADD UP TO COLLECTIVE IMPACTS.

If I gave you one penny, you wouldn’t say that penny made you rich.

But if I kept giving you pennies, one at a time, at exactly how many pennies would you say that you’re rich?

Or put another way…how many drops of water would you say it takes to become an ocean? (There’s a reason A Drop in the Ocean is named what it is.)

You cannot separate individual actions from collective impacts.

I’m over these arguments. I’m just. so. over them.

Because sure, let’s say that one ChatGPT query “only” uses 17.6 mL of water. It’s the low estimate, but sure, we’ll go with it.

ChatGPT gets 2.5 BILLION QUERIES EVERY. SINGLE. DAY. (81)

That 17.6 mL per query adds up to more than 11.6 MILLION GALLONS of water used every single day. For what? Ugly posters and Instagram captions that all sound identical?

And that’s ONLY ChatGPT. That doesn’t include any of the other AI platforms out there.

So I’m gonna include a chart of everyday activities’ water usage here, and individual AI queries are gonna have the lowest consumption. But that doesn’t mean it doesn’t matter.

Because just four years ago, none of this existed.

So much of the zero waste / plastic free / sustainable living movement is asking ourselves, “What did our grandparents do? How can we learn from the way they did things, before plastic was everywhere and manufactured demand and planned obsolescence launched us into a constant state of overconsumption?”

With AI, we literally just have to go back 4 years. We don’t even have to go pre-COVID. How did we do things without AI in 2022? 

We did them just fine, thankyouverymuch.

All of these numbers AT BEST represent where we are now, but in reality it’s probably closer to a year or so ago. But with how rapidly AI is being integrated EVERYWHERE (and how we mentioned in Part 1 that the tech companies are forcing it upon us so hard because unless every one of us integrates AI into every facet of our lives almost immediately they’ll lose TRILLIONS of dollars that they do not have), it’s gonna get worse.

The following quote is from the MIT deep dive I referenced in Part 2:

“In the future, we won’t simply ping AI models with a question or two throughout the day, or have them generate a photo. Instead, leading labs are racing us toward a model where AI ‘agents’ perform tasks for us without our supervising their every move. We will speak to models in voice mode, chat with companions for 2 hours a day, and point our phone cameras at our surroundings in video mode. We will give complex tasks to so-called ‘reasoning models’ that work through tasks logically but have been found to require 43 times more energy for simple problems, or ‘deep research’ models that spend hours creating reports for us. We will have AI models that are ‘personalized’ by training our data and preferences… [A]ll bets are off in the coming years.”

This article was from May 2025, with this quote warning that their estimates very soon won’t fully account for how AI is really being used. And we’re already seeing these things. It’s happening now.

So next week, we’ll wrap up our series with ways that we can all take action and reject the idea that all of this is the inevitability that the tech bros want us to think it is.

I truly do believe that it’s not too late. 

It’s up to each of us to decide what kind of future we want to see.

P.S. Let me know what you’re thinking of this series so far! Is there anything that’s surprised you, or anyone that you’ve shared the series with yet? I want to know your thoughts! Drop a comment below and let’s chat about it!

 

💡 View all of our sources used for this blog post series here.