Process Engineer Explains: The Math Behind "Water-Efficient AI Data Centres" Is Laughably Wrong

[00:00:00] Speaker 1: Across the American West, AI data centers are now competing directly with residents for drinking water, all while the tech industry says we are improving our efficiency. And the data centers that claim to use almost no water, well, they found a way to shift the burden somewhere that you are not supposed to look. But here's a fact that sounds like it was written by someone who failed their thermodynamics exam. The more energy efficient AI data centers become, the more total water the industry consumes. I know that sounds counterintuitive, and in this video, I'm going to show you the phenomenon discovered by 19th century economists that prove that this is actually true, and that why nothing the tech industry is currently building will change it. There is a significant physical constraint on AI infrastructure. And by using engineering first principles, I will explain it in a way to you that nobody else in this AI commentary space is doing. And by the end of the video, I will give you a simple tool for you to use to fact check the water consumption headline numbers by yourselves. In my previous video, I covered the power and infrastructure constraints on AI, and the single most common pushback that I got in the comments was, but efficiency will improve. So doesn't that change the picture? And yeah, it's a fair question. I'm a chemical process engineer, and I spent years in the energy sector designing process systems, heat exchanges, cooling loops, the infrastructure that moves heat and water at an industrial scale. My PhD research was about what it actually takes to deliver power and water access sustainably at the scale that a real population needs. And so when I look at a hyperscale data center, instead of a compute problem, what I see is a heat rejection problem. And that reframe changes everything about the efficiency question, because the answer is grounded in thermodynamics that no amount of engineering cleverness gets around. And if you happen to live anywhere that shares a watershed with a data center, or anywhere that pays taxes towards the water infrastructure that they need, the equation I'm going to show you is going to be an eye opener. But before I get to the equation, I need to show you the scale of what's already been built because the numbers are genuinely staggering. A study published in 2026 by researchers at UC Riverside modeled US data center water demand through to 2030. Their low estimates for new peak capacity required was 697 million gallons a day. And their high estimate was just under 1.5 billion gallons a day. For context, New York City uses about 1 billion gallons a day. So we're talking about building somewhere between three quarters and one and a half New York City's worth of entirely new water infrastructure in about four years to keep the AI build out running. Now, I need you to hold two things in your head while I go through this. First, the real capacity constraint on building AI data centers isn't the design or supply of chips. It isn't planning permission and it isn't capital. It is heat dissipation because you can throw any amount of money at a data center. If you cannot reject the heat, you cannot build it. That is a universal physical law. It's not a market condition. And second, every efficiency curve in technology history eventually flattens. And it's not because the engineers gave up, but because physics sets a hard ceiling. And I believe that we're approaching that ceiling faster than the press releases are letting on. Now, I'm going to come back to both of those. But first, I want to put a number on why this matters specifically to you. You see, the cost of water infrastructure needed for the AI build out. You have treatment plans, pipelines, permits. It runs into tens of billions of dollars. And that money comes from municipal budgets, which means it comes from your taxes and your utility rates. In Virginia, for example, residential electricity bills have already gone up by about $16 a month just to fund data center power infrastructure. Water bills in several data center corridors are next. And if your retirement savings sits in a default 401k or workplace pension, and for most of you watching, they probably do. You're already financing this build out through your investments in index funds that finances these AI companies. And so the water that these facilities consume isn't just an environmental issue. It's a line item on your bills. And it's a risk that's sitting in your retirement account, whether you know it's there or not. Now, earlier, I told you that there was an equation. So let me show you exactly what it says. Right. So we'll do this from scratch and build a vocabulary first because there are too many people who use these words loosely and that looseness is where the confusion lives. A watt is a joule per second. It's basically a rate. It tells you how fast energy is moving, not how much energy you have. So a gigawatt is a billion of those every second continuously. And when we talk about IT load, that's the electrical power that's actually consumed by the computing equipment itself. So servers, GPUs, storage, networking gear, as opposed to the power that's consumed by everything else that keeps the building running. So things like cooling fans, pumps, chillers, lighting, and power conversion losses. And when a data center operator talks about cooling load, they mean exactly one thing. It's the rate at which heat has to be carried out of the building. And what makes data centers different from a lot of other industrial facilities is the cooling load pretty much equals the electrical load. And so let me walk you through a real example. And I'm going to use a theoretical data center with about 500 megawatts of IT load. Is that big? Not particularly because companies like Google and Microsoft are building facilities that are much larger than that size right now. But it's a clean number and it makes the arithmetic easy to follow. And so step one is the conservation of energy. You see, every single watt of electricity that enters a data center becomes heat. That is not a design choice. It's not a failure of engineering. It's just thermodynamics. The electricity goes in as electrons. The computation produces heat. There is no exit where useful work leaves the building in any other form. It's all heat. And so our cooling load, which I described earlier, we call that Q, equals our electrical load. And we'll call that P. That's the whole of step one. Step two, we have to define the latent heat of vaporization. Now, when water evaporates, so when it changes from liquid to vapor, it absorbs a fixed quantity of energy from whatever it's in contact with. And that's how, for example, your body cools down when you sweat. The water on your skin isn't just carrying heat away. It's absorbing a fixed amount of energy in the act of changing phase from liquid to vapor. And that fixed amount is called the latent heat of vaporization. And this value changes with the temperature of the liquid. Data center cooling towers will evaporate water at ambient environmental temperatures, which for the purpose of this video, we can take to be about 25 degrees Celsius. The latent heat of vaporization of water at this temperature is 2442 kilojoules per kilogram. And so we move to step three, which is the equation. The mass flow rate of water that you need to evaporate equals the heat that you need to reject divided by the latent heat of vaporization. This statement is represented in the equation that you see on your screen right now. One megawatt is 1000 kilowatts and one kilowatts is one kilojoule per second. So one megawatts is 1000 kilojoules per second. We stated that our latent heat of vaporization is 2442 kilojoules per kilogram. And so for one megawatts, the mass flow rate equals 1000 kilojoules per second divided by 2442 kilojoules per kilogram. Watch the units cancel, kilojoules on top, kilojoules on the bottom, gone. You are left with kilograms per second. So that means that your mass flow rate is now 0.41 kilograms per second per megawatts of heat rejected or 0.41 kilogram per second times 86,400 seconds per day, which is circa 35,400 liters per day or approximately 9,352 gallons per megawatt. That is the thermodynamic ceiling. A one megawatt heat load at perfect evaporative cooling efficiency evaporates 0.41 kilograms of water every single second it is running. If we take the typical, hypothetical 500 megawatt data center I mentioned above, the mass flow rate becomes 205 kilograms per second or roughly 4.7 million US gallons per day. Now, real world systems don't operate at this theoretical ceiling. Actual cooling systems are part latent but they are also part sensible. That is, some of the heat leaves through evaporation and some just leaves as warmed air or warmed water without a phase change. This is also known as sensible heat transfer. And also the IT load that a company reports isn't identical to the total heat that the whole facility rejects. And so the typical range across the industry is about 8,800 to 2,900 liters per megawatt hour of IT load depending on ambient temperature or humidity and how the system is actually designed. But they are best-in-class systems that use less with values of about 1,000 liters per megawatt hour. So let's use this value. Our 500 megawatt data center runs continuously. 24 hours a day. So 500 megawatt times 24 hours equals 12,000 megawatt hours per day. 12,000 megawatt hours per day times 1,000 liters per megawatt hour equals 12 million liters per day. If you convert to US gallons at approximately 3.79 liters per gallon, it's roughly 3.2 million gallons of cooling water per day from one facility. And here's the part that sustainability reports bury. Up to 85% of that water evaporates and it never returns to the local water systems. It's just gone. It becomes atmospheric moisture somewhere else. Now, before the sharp-eyed viewers head to the comments to tell me that I'm wrong. Yes, all these numbers depend on climate design choices and how much of the cooling is evaporative versus mechanical. The range does matter but it will never be zero. And here's what I promised you at the start of this video. You now have everything that you need to fact check any data center water headline. What you do is you take the facility's power capacity in megawatts and you multiply it by 9,400 gallons per day or about 10,000 gallons for quick and easy calculation to get the thermodynamic ceiling. Now, note like I just said that real world facilities typically land between 40 and 75% of that depending on design and climate. And so if the published number is well below that range, you have to raise an eyebrow and ask them what they're not counting. For some further context, let me tell you what that kind of water consumption looks like in practice. Google's data center in the Dalles, Oregon consumed about 550 million gallons or roughly 1.5 million gallons per day in 2025. And this was about 40% of that small city's total water supply. When the Oregonian, which was a newspaper, tried to get the actual usage figures through a public records request, the city of the Dalles sued the newspaper to keep Google's numbers secret. Now, this is the part of the video where I ask if you have so far received any decent value from watching this video to please subscribe if you haven't already. It really helps small channels like mine. Now, if you follow the tech industry, you most likely have heard about liquid cooling and how it's going to solve all of this problem, right? But I need to show you exactly what that claim is and isn't doing. You see, when Nvidia talks about liquid cooled blackwell racks or when you hear about direct-to-chip liquid cooling or when Google says that one of their facilities uses closed loop cooling, what they are describing is the coolant circuits. It means that the fluid that touches the chips is recirculated. Nothing evaporates and it doesn't mix with local water. So, problem is solved, eh? Except closed loop does not say anything about where the heat that loop picks up goes. Here's the thing. The coolant loop absorbs heat from the chips and carries it somewhere. And that somewhere is either a cooling water which evaporates water or a dry cooler and chiller which consumes electricity. Liquid cooling doesn't eliminate heat rejection. It is more efficient at getting the heat to the rejection system. However, it doesn't change the fundamental equation. And so, when a data center operator tells you that their facility uses dramatically less water on site, you need to ask them the follow-up question that they'd probably rather you didn't. Where did that water go and what replaced it? In most cases, the honest answer is going to be electricity via dry coolers or chillers. And now, this next bit is fascinating because what happens to the water when you replace evaporative cooling with chillers might actually make the problem larger and not smaller. And in order to understand why, we need to follow the electricity. You see, when a data switches from wet cooling towers to dry cooling, their on-site water use drops dramatically. And some hyperscalers report this as evidence that their facilities are becoming more sustainable. But dry cooling isn't free. Running chillers and mechanical coolers, rather than evaporatively cooling your data center, requires significantly more electricity. And that electricity has to come from somewhere. In the United States, electricity generation has historically averaged around 15 gallons of water withdrawn per kilowatt hour. Although it has recently dropped to just about 12 gallons per kilowatt hour and roughly 0.47 gallons actually consumed, meaning it doesn't return to the source. If you scale that to a whole year of generation, and the United States geological survey data puts consumption at something like 1.2 to 1.4 million gallons a day for every terawatt hour of annual production capacity. And so when the data centers say that they are eliminating their on-site water use by going dry, it does genuinely reduce its own water bill. But they are also increasing their electricity draw, which increases the water consumed upstream at the power plant. And so the water didn't just vanish. It just moved out of their sustainability reports and into someone else's watershed. Therefore, the honest metric should be on-site consumption plus power plant consumption combined. And when you add those up, dry cooling saves water is often a reporting sleight of hand rather than a physical reality. Berkeley lab put total US data center in direct water use at roughly 800 billion liters in 2023, somewhere around 579 million gallons a day. Now, I want to flag this now that this number is contested. And I'll bring you the actual counter argument later on in this video. But even the more conservative estimates don't get you to zero. They get you to, it's smaller than the headline, but you know, it's still not nothing. But here's a detail that I think should bother you more than the number itself. In nearly every case I could find, the exact withdrawal volumes are hidden behind non-disclosure agreements between the operator and the local government. Now, I'm doing this math with public averages because the companies involved won't publish the rule figures. I'll come back to why they hide these numbers and exactly how they do it later on in this video. But there's another layer to this that isn't talked about enough in discussions on data centers and it has to do with where specifically they chose to build. And so where do hyperscalers prefer to build? Arizona, Nevada, New Mexico, Texas. And the reason makes sense from a cooling efficiency perspective. You see, evaporative cooling works by exploiting the gap between ambient air temperature and something that we call the wet bulb temperature, which is roughly how cool the air could get if you evaporated water into it as fast as possible. In dry desert air, that gap is massive. The physics are incredibly favorable and the cooling towers run efficiently. You get great water usage effectiveness numbers, but in humid air, the gap collapses and evaporative cooling barely works at all. So this is why you cool down faster in dry conditions compared to when it's, you know, really humid outside. And that's part of why the industry keeps choosing these locations. The air genuinely does the job better there. But you see, the problem is that when you say something works best in dry desert air, it means, hey, let's evaporate the scarcest water in the entire country, boiled straight into the desert sky in the one region that can least afford to lose it. They are building the world's most water hungry infrastructure in places with the least water to spare. And then they report a low water per kilowatt hour figure and call it sustainable. And the heat that same process rejects doesn't just vanish once it leaves the tower. A field study measured thermal plumes from a Phoenix area data center cluster and found downwind air up to 2.2 degrees Celsius warmer, detectable at 500 meters out from the site. One facility's waste heat alone was measured, exceeding the output of 40,000 households in a desert city that's already running air conditioning at near peak load for eight months of the year. When you add that heat, it increases local cooling demand, which means more power is needed, which means more water is withdrawn upstream of the power stations. The efficiency gains create a vicious feedback loop in exactly the places where the water systems can least afford it. Right, if you've made it this far, you'll remember that at the beginning of this video, I made a claim that the more energy efficient an AI data center becomes, the more water it consumes. And so why does improving the efficiency of these systems make them even thirstier? In my previous video, the single most common pushback I got in the comments was, but efficiency will improve. Doesn't that change the picture? And like I said earlier, it's a fair question. And yes, per unit efficiency is improving, but it really doesn't help. The reason has a name. It's called the Jevons paradox. In the 1860s, the economist William Stanley Jevons noticed that making steam engines more coal efficient didn't reduce coal consumption. It simply reduced the cost per unit of work, which increased demand, which increased total coal consumption. The same principle applies here. Per query per unit water consumption is genuinely falling. Microsoft's own water usage effectiveness figure has dropped from about 0.49 liters per kilowatt hour down to roughly 0.30. Amazon is claiming figures as low as 0.12. And these are real verified engineering improvements. But the total water use is rising anyway, because the volume of compute being run has exploded faster than the efficiency gains could offset it. Google's total water consumption rose 17% in a single year to 6.1 billion gallons in 2023, despite the company simultaneously improving its per unit efficiency. Its single most water intensive US site was averaging 2.7 million gallons a day by 2024. So like I said, per query is cheaper, but in total, it's thirstier. And it sounds like a contradiction, but it's really just what happens every single time you make something cheaper to run. The aggregate goes up even as the unit costs go down. And I want you to note that this isn't just a passive market market force because these companies have every commercial incentive to ensure that demand outstrips efficiency gains. Token prices, for example, have dropped roughly 90% since 2023. However, did that reduce the total cost that enterprise, for example, is paying for AI? No, of course not. It exploded it. Cheaper tokens meant more products, more agents, more always on systems, more use cases that perhaps weren't economically viable at the old price. And there's something else that rarely gets discussed. Some of these AI companies are actively manufacturing demand. Enterprise contracts with minimum token commitments. Product integrations that somehow default to AI whether the user needs it or not. Organizations are running AI agents just to meet their monthly allocation. Token maxing, anyone? No, does this efficiency treadmill just keep running forever or does it eventually hit a wall? Well, there is seemingly this idea that engineers will simply keep improving this indefinitely until the water problem quietly disappears. This was even a sentiment that was shared in the comments in my previous video. But the thing is efficiency gains follow an S-curve. Fast improvements come early while the cheap winds are still on the table and then you have a flattening as you approach the physical limit. The easy gains in data center cooling have already been taken. The industry moved from average water usage effectiveness, so that's a metric that measures data center water efficiency. With figures of around 1.9 in 2015 to around 0.3 for the most efficient hyperscale facilities today. That's a real improvement. And it may seem like there's headroom left. But there isn't. You see, the early improvements, the better airflow management, free air cooling in mild climates, heat exchangers instead of legacy chillers were cheap and effective. But what's left costs disproportionately more per unit of improvement. This is the same shape that every mature energy technology eventually takes. And underneath the S-curve sits something that no amount of engineering cleverness gets you passed. You have two hard physical walls and every design has to pick which one it hits. If you go down the low water power hungry path, so that's dry cooling, chillers, heat pumps, you will eventually hit the Carnot limit. This is the theoretical maximum efficiency of any heat pump and it is set by the temperatures involved. You cannot engineer past it. You can approach it asymptotically with enough capital, but you will never cross it. Blame the second law of thermodynamics for this. If you go down the low power water hungry path, which is evaporative cooling, you will hit the wet bulb temperature and that's the coldest that you can cool anything by evaporation alone. So for example, in Phoenix in July, that's around 21 degrees Celsius, you cannot cool a data center below the local wet bulb temperature without switching to a more expensive system. There's only so much moisture that the surrounding air can absorb before it simply can't take anymore. You cannot engineer past that either. So every facility built today is choosing which wall it's going to run into. No design avoids both walls. The rejected heat from the data centers absolutely has to go somewhere. And both of its only exits are capped by physics that no amount of capital injection will change. So let's scale this up because the national picture is where the numbers stop being abstract. The Wren et al. study puts the infrastructure cost of that 697 million to 1.5 billion gallons per day demand at somewhere between 15 billion dollars and 58 billion dollars in a high growth scenario. 7 to 28 billion dollars in a slower growth one and between 6 and 24 billion dollars for a best case scenario where the industry hits 10 percent annual efficiency improvements going forward. Now I know that sounds like a lot but it's not actually the binding constraints. One of the researchers put it plainly. Money can build treatment plants and pipes but money cannot buy you more snowpack. You can finance your way past almost every constraint in this industry. Cheap shortages, power shortages, land, labor, permitting delays. Unfortunately you cannot finance your way past the drought. The water that isn't in a mountain range in a given year does not exist at any price. And to put some real numbers under the abstraction. Training GP3 alone, so you know the old chap GP3 model, is estimated to have evaporated around 700,000 liters of water, roughly 185,000 gallons, just for that one training run. And projections for global AI water withdrawal by 2027 run as high as 4.2 to 6.6 billion cubic meters, which is roughly half of the United Kingdom's entire annual water withdrawal. The Colorado River supplies water to Arizona, Nevada, California and New Mexico, which is where a significant portion of the planned air infrastructure is being sited. The Colorado River Compact, which was signed in 1922, divided 16.4 million acre feet of water per year between seven states. The river's actual average flow is around 12.4 million acre feet. They divided water that did not exist due to an overestimation of its flow. The states competing for water from a shrinking river, at the same states where they are planning to build the equivalent of a city's worth of new water demand for AI infrastructure. And this isn't just a model, it's already playing out in towns right now and the towns are starting to push back. In August 2025, Tucson City Council voted unanimously to reject an Amazon-linked data center project, referred to internally as Project Blue. The stated reason was water. By early 2026, 12 US states had introduced moratorium bills on data center construction. A Gallup poll found that over 70% of Americans oppose data centers being built near their homes. Let's look at the Great Lakes region specifically because it's where this collides hardest with a resource that these tech companies seem to assume is functionally infinite. 243 data centers are planned in Illinois, 192 in Ohio, 142 in New York. The Alliance for the Great Lakes, which serves 40 million people, has one of 150 billion gallons of water per year of new withdrawals within five years. And inside every one of these stories is the same pattern that I flagged earlier, that the actual numbers are hidden. I told you I'd come back to that. So let me show you what I mean. The infrastructure that the AI buildout needs is running directly into the infrastructure that towns actually live on. You see, the AI industry doesn't just use water. It also obscures how much water they use. There are four layers to this. First, operators report on site water use only, which ignores the water that's consumed upstream at the power plant that feeds them. This is the reporting sleight of hand that I showed you earlier. Second, they optimize for a metric called water usage effectiveness, which measures liters per kilowatt hour. And then they site their facilities in hot, dry climate where that metric looks best, which happens to be where the water is scarcest. Third, the actual withdrawal volumes are buried behind non-disclosure agreements between the operator and the local government, which means that the communities whose water they are drawing can't see how much water is being taken. And fourth, pilot stage designs like Microsoft Zero Water Systems gets announced as solutions to a problem that every facility actually running today still has and will probably not come into full scale production for years. Now, I'm going to do something unusual. I am going to argue against myself because I think there's a serious case for the defense. And I'm going to invite you to provide your own arguments if you have any in the comments. In December 2024, Microsoft announced a closed-loop chip-level cooling design that consumed zero water for evaporation entirely. They claim it saved 125 million liters per year per facility compared to their previous systems and they are piloting it in Phoenix and Mount Pleasant, Wisconsin, with plans to make it standard across all new designs from late 2027. And if this can be executed at scale, then it's a genuine step change. But here's my honest rebuttal to this point I'm just making. It trades water for power. It still has to reject that heat through chillers, which means it's still bound by everything I told you about the carnal limited path and it is a future tense thing because the vast majority of data center capacity actually running today still evaporates water and will keep doing so for years while this design, if it actually works, rolls out. So the water moved. It did not disappear. And it is piloting, not deployed on the 400 megawatt campuses under active construction right now. And remember the national numbers I used earlier in this video for I used earlier in this video. What's the first time I used earlier in this video in this video? Well, the researcher Brian Potter has argued that the Berkeley Lab indirect water estimate I cited, so that 579 million gallons a day figure, likely double counts evaporation from hydroelectric reservoirs. And it doesn't properly account for the fact that most hyperscalers are buying wind and solar power through purchase agreements, which carry essentially zero operational water costs. His adjusted estimate lands closer to 200 to 275 million gallons a day. That's a real methodological critique from someone who did careful work. And after going through his calculations, my honest response is, he's probably right that the headline national figure is inflated. And I will leave a link to his article in the description box. But here's where I think these numbers debate is actually a distraction from the real argument, and where I'll even hand you a point that looks like it undercuts me. Arizona's roughly 400 golf courses reportedly consume significantly more water than the entire US data center industry. So if you're arguing about aggregate gallons, golf courses win that argument against data centers easily. But the primary argument should not be about the national totals. It should be about the evaporative load that's aimed with real precision into specific stressed watersheds such as Arizona, Utah, the Mahomet aquifer, where it goes head to head with the drinking water of actual residents. A number that might look trivial when it's averaged across a whole country can still be devastating when it's averaged across just one basin. And so here's where I land in all of this. And two things can be true at the same time, while neither one cancels the other. On one hand, the water efficient cooling design is real. The progress is measurable and the engineers that are building it really deserve credit. But on the other hand, the AI build-out at this scale, sited in these specific places, on this timeline, still cannot source the water it needs because the fix relocates the burden rather than removing it. And because physics and thermodynamics set a floor that no amount of capital can lower. And because also, as one researcher put it better than I called, the snowpack doesn't negotiate. I spent a significant proportion of my career on the other side of this exact problem. You know, building the infrastructure that's supposed to deliver power and water access sustainably at the scale a real population actually needs. And so me watching an industry with what is effectively an unlimited capital run straight into the same wall that constrains everyone else says a lot. If you want to see the full engineering breakdown on the other side of this, why the power generation timeline, the grid queue, and the cooling load means that this build-out can't be delivered on the schedule that these AI companies are promising. I did that in my previous video. It's the other half of this picture. It will give you the complete toolkit to fact-check any AI data center claim that you see. And it's linked right here. If this video gave you a sharper way to think about what's actually been built in the AI space, please give it a like and subscribe for more like this. I'm out. Bye.