The AI Water Use Controversy: How It's Completely Misunderstood

Ejaaz: You've all heard the headlines. AI is draining our water supply. Ejaaz: Data centers are stealing drinking water from the communities. Ejaaz: ChatGPT is literally drying up the planet. It sounds terrifying and it's also completely fake. Ejaaz: Today we're going to bust one of the biggest myths on the internet and walk Ejaaz: through the actual numbers which reveal something crazy that the world's largest Ejaaz: data center uses about as much water as two burger joints.

Ejaaz: We're going to walk you through exactly how much water is needed for the biggest Ejaaz: data center in the world, Colossus 2, and why the majority of water used is Ejaaz: only a fraction of the water used by your local golf club that your friends go to every weekend. Josh: So part of the feedback that we've seen as we become more of a presence, Josh: a platform in the world of AI, is that there's a lot of narratives that try to take down progress.

Josh: And the newest and hottest topic has been the topic of water consumption, Josh: to the point where when I talk about with my friends about what I'm working Josh: on and what I'm interested in, that frequently comes up as the first rebuttal. Josh: It's like, this is horrible for the environment. It's using so much water, so much energy.

Josh: And this episode, we're going to focus on the water, particularly around the Josh: public perception versus comparing it to the reality and how far off it really is. Josh: We're seeing on the screen a few headlines of people from Utah, Josh: very viral publications and headlines that have been talking about this. Josh: But the reality is, is it's simply not true.

Josh: So we're going to methodically and in a very fun way kind of dissect how wrong Josh: this actually is, starting with the book that kind of spawned it all. Josh: The author's name is Karen Howe. No pun intended. Josh: She really, like Karen, Karen seems like a very fitting name as an author called Josh: the Empire of AI, claiming a Google data center would use 1,000 times the water Josh: of an 88,000 person city.

Josh: Studies projecting AI will consume 1.7 trillion gallons of fresh water by 2027. Josh: These claims are a bit outrageous, but this has been the narrative that people Josh: have been using as they discuss their rebuttals against why AI should exist. Josh: And this simply is not true. And I guess this is kind of where we got skeptical. Josh: We were like, okay, 1.7 trillion gallons, a thousand times more than an 88,000 person city.

Josh: These numbers are outrageously large, but they're being peddled around as if Josh: it is truth in these large publications. Ejaaz: The thing is, like, we work in the AI space, Josh. Ejaaz: So trillion dollars, trillions, We're kind of like used to it at this point. Josh: Trillion here, trillion there. Ejaaz: Yeah, right. But what I wasn't used to is a trillion gallons of water. Ejaaz: And the first thought I had to my head was, what is this compared to?

Ejaaz: Like, I can't quite comprehend how much water this is and how much water is Ejaaz: used in other industries that aren't AI adjacent. Ejaaz: So I started getting skeptical and I came across what I think now is the most Ejaaz: important bit of journalism done on data centers by these guys at SemiAnalysis.

Ejaaz: They are a crazy team of researchers that kind of like dig into all the boring, Ejaaz: nitty-gritty, hard data center stuff to bring us the facts about lots of different Ejaaz: things, including how much water is consumed at the data center level. Ejaaz: And the revealings are super interesting. But what I love most about this, Ejaaz: is, Josh, is that not only is it really informative, it's also hilarious. Ejaaz: Because... Oh, it's great. Josh: I love this post so much.

Ejaaz: Right. They've compared data center water usage Ejaaz: to burger restaurants, specifically In-N-Out. So for some context here, Ejaaz: they took the largest data center in the world, which is Elon Musk's Colossus 2. Ejaaz: It's a data center based in Memphis, and it is the first data center to reach a gigawatt of compute.

Ejaaz: We actually mentioned this on yesterday's episode. If you haven't seen it, Ejaaz: definitely go check that out. It's super cool. Ejaaz: So we're talking about $18 billion worth of GPUs here, Josh. Ejaaz: So as you can imagine, it's probably using a lot of water, right? Ejaaz: So the math that they uncovered was the most revealing part. Ejaaz: So Colossus 2 has an annual water footprint of 346 million gallons of water per year.

Ejaaz: Guess how much the average in-and-out store consumes per year? Josh: A lot. Ejaaz: 147 million gallons per year. Per store. Ejaaz: Per store, sorry, yeah, per store, yes. So that means that the largest data Ejaaz: center in the entire world currently today that everyone's complaining about Ejaaz: consumes the same amount of water as two and a half In-N-Out stores. Ejaaz: Why aren't we protesting In-N-Out? Josh: This is super fascinating to me. And it gets even better than this.

Josh: If you think that's funny, it gets even better. Josh: So basically, the report looked at tokens per burger. So we were able to get Josh: a metric for how you can actually justify output. Ejaaz: Wait, wait, wait. Run that unit by me again. Josh: We are breaking the news right here on Limitless. We are measuring water efficiency by tokens per burger. Josh: And a single burger's water footprint is about 245 gallons of water.

Josh: That equals 2.7 billion AI output tokens, which roughly equates to one burger... Josh: Being equivalent to using Grok 30 times a day for 668 years. Josh: So the numbers are just absolutely astronomical. Ejaaz: Wait, wait, wait. Josh: Before I'm done, wait, one more thing, one final conclusion to this. Josh: There are over 400 in and outs. There's so many of them.

Josh: And there's what, maybe 10 AI data centers? So the scale and the magnitude at Josh: which they are wrong is just so astronomical. Josh: I found it really funny like once you actually get into the numbers you realize Josh: this is so not a problem that it's.

Ejaaz: Not even funny wait dude so this is you're telling me so this is incredibly click Ejaaz: baity basically um the figures are astronomically Ejaaz: wrong and if you compare it Ejaaz: to just a casual restaurant we're talking about burgers only here by the way Ejaaz: we're not talking about the fries sides and everything else just milkshakes Ejaaz: god knows how much one of the milkshakes use we're just talking about the burgers

Ejaaz: including like their restaurant output and the supply chain for this so the Ejaaz: fact of the matter is, this problem isn't really a problem. Ejaaz: It's actually massively overblown. But what I'm curious about, Ejaaz: Josh, is like, okay, is this a realistic take? Ejaaz: Like, I know that data centers use a lot of water. Ejaaz: How are they using it? And like, how much of this is renewable versus is actually Ejaaz: burned and evaporated into the air and we'd never get it back?

Ejaaz: Like, what's the comparison? Josh: Yes, yes, yes. Okay, so the truth, they are using lots of water. Josh: There's hundreds of millions of gallons per year that are going through these data centers. Josh: Some of it gets lost, a lot of it gets preserved. And the way they do it is Josh: there's two types of cooling. Josh: There's dry cooling, and then there's adiabatic cooling, which is the process Josh: where air cools down without exchanging heat with its surroundings.

Josh: So if you remember the old iPhone episode that we did that talked about vapor Josh: chamber, you can imagine scaling a vapor chamber to the size of an industrial data center. Josh: And that's roughly how the adiabatic cooling works. Josh: So it evaporates the water, the water leaves the system, and that's where they Josh: lose about 267 million gallons per year. Josh: And then the second loss function is the flush and discharge.

Josh: So one thing that I learned in preparing for this episode is that, Josh: I mean, there's a lot of mineral buildup in the water that they use. Josh: And 67 million of those get discharged as waste water per year. Josh: But those are the two ways they do it. Josh: It's dry and it's adiabatic. And there are promises in 2027 and 2028 from a Josh: lot of the major AI labs to decrease this waste to about 95%. Josh: Currently, it sits about 90%. So 90% gets recycled, 10% gets lost.

Josh: Those numbers are going to increase incrementally until about 2030 when the Josh: number is actually net positive. Ejaaz: Okay, so it sounds like the classical way of thinking about how data centers Ejaaz: get cooled is you run a bunch of water in pipes through all these different GPUs, Ejaaz: and it removes the heat from these GPUs so they are able to perform at optimal levels, right? Ejaaz: And part of this water evaporates, never to be seen again, and some of this

Ejaaz: water gets kind of recycled over and over again. And the adiabatic system, Ejaaz: I think is what you said, is kind of a hybrid of both of these things.

Ejaaz: But most importantly, it's more of a closed loop system. Ejaaz: So we've kind of got like the majority of the water being renewed. Ejaaz: So there's a comparison between water being renewed and water being consumed, Ejaaz: which means lost forever. Do I have that right? Josh: Yeah, so there's, you could think of a closed-loop system, going back to that vapor chamber. Josh: If a vapor chamber is 100% efficient at being closed-loop, where water reaches

Josh: the processor, it heats up, it evaporates. As it evaporates, it dissipates the heat. Josh: This is that at scale, although with some sort of a loss function at the end, Josh: where some of that evaporated water does currently exit the system. Josh: It sits now at about 10%, and that's where that loss comes from. Ejaaz: But I'm guessing it's not 1.7 trillion gallons of water a year. Yeah, it's a

Josh: Far cry from the 1.7 trillion number. And then there's a second aspect to this Josh: that gets a lot of criticism, which is the actual power generation, Josh: how much water gets used in the generation of energy through these turbines Josh: that are natural gas, some solar. Josh: And the answer to that is, in the case of Colossus 2, 0%. There is no meaningful Josh: water consumption of power generation at all.

Josh: The entirety of it comes from cooling the GPU system, which is closed loop and done with water. Josh: So maybe we just go into where that 1.7 trillion number even came from. Josh: Because this is the source of Josh: a lot of the narrative that we've seen play out over the last few weeks. Ejaaz: Exactly. So I'm showing all of you folks who are watching a scientific paper Ejaaz: from UC Riverside titled Making AI Less Thirsty. Josh: When you say scientific, you have to do air quotes.

Ejaaz: Yeah, sorry, sorry. Pseudoscience. From this university called Making AI Less Ejaaz: Thirsty, uncovering and addressing the secret water footprint of AI models. Ejaaz: Now, this paper is the source of a lot of the clickbait headlines and TikToks Ejaaz: that you watch online. It's this one number. Ejaaz: 1.7 trillion gallons of water will be consumed by AI data centers alone by 2027. Ejaaz: That's around the corner. We're talking about next year here, right?

Ejaaz: And so a bunch of people kind of like threw up their hands and started protesting Ejaaz: data centers because they were like, that is so much water. Ejaaz: We're not going to be, we humans aren't going to be left with enough water to Ejaaz: consume ourselves. Right? Ejaaz: But here's what this study actually says. It claims 1.7 trillion gallons of Ejaaz: water are used for withdrawal. Ejaaz: That's a fancy term of saying recycled.

Ejaaz: So imagine like the water being taken, used to cool down the systems, Ejaaz: the GPs that we just mentioned, and then recycled again and again and again. Ejaaz: So it's not net new water we're talking about here. Ejaaz: Josh, do you want to know the actual consumption that is being like permanently Ejaaz: removed, the ones that we should be trying to protest? Josh: Yeah, well, I mean, based on that, it's what? It's at least a full order of Josh: magnitude off than what is projected.

Ejaaz: No, it's 100 to 158 billion, not trillion, billion gallons of water. That's 10%. Ejaaz: Actually, it's less than 10% if you take the lower bound of the reported number. Ejaaz: So the water that's actually being used by the data centers is only kind of like 10 to 15% of that. Ejaaz: And okay, so some people then go like, what about the drinking water?

Ejaaz: Like what percentage of that is being affected from the study that was being Ejaaz: made? It's only 3% of the headline number. Ejaaz: So everyone took 1.7 trillion gallons of water and assumed all of that water Ejaaz: was being wasted, never to be seen again, used again, and it was pulling from

Ejaaz: other resources when that strictly isn't being true. So the point I'm making, Ejaaz: and I'm going to reiterate it again, Ejaaz: The water, the 1.7 trillion gallons of water isn't being consumed. Ejaaz: Think of it like this. Like imagine diverting a river to run through a mill. We do that today, right? Ejaaz: And it flows back into the river. That's called withdrawal. That's the 1.7 trillion Ejaaz: number that I'm pulling out here.

Ejaaz: So most of the power plant water is being reused over and over again. Ejaaz: So that's only the first major kind of bit of pseudoscience that we needed to bust. Ejaaz: But there's this second thing, and Josh, you mentioned it earlier from, Ejaaz: ironically, Karen Howe, who wrote this famous book called... Josh: Our favorite author. Ejaaz: ...called The Empire of Air. Let me get this up here so everyone can see this book.

Ejaaz: It has been rated over 1,300 times on Amazon, but more importantly, Ejaaz: it has been quoted directly by The Economist, The New York Times, and so much more. Ejaaz: For this specific stack, it says Google's data center will use 1,000 times the Ejaaz: water of an 88,000-person city. Ejaaz: Guess what? It was off by a factor of not 1,000 times, 4,500 times. Ejaaz: Josh, can you run me down these stats? Because it's just insane, dude.

Josh: No, that stat actually makes me sick. It's horrible. So instead of comparing Josh: it to Google, we're going to start with golf courses because that's just... Josh: If you want to come at it, you got to start with the golf course. Josh: The average golf course, 312,000 gallons of water. Josh: Desert golf course is 1 to 2 million. A Google data center in Virginia is 400,000. Josh: Now what does that 400,000 give you?

Josh: That powers Gmail, Google Drive, YouTube, the entire G Suite for billions of Josh: people around the world with the equivalent of 1.2 golf courses worth of water for billions of people. Josh: Google's global data center that powers 4 billion accounts, Josh: equals 43 golf courses. And in the state of Arizona, there are 370.

Josh: So if you're comparing apples to apples here, that is the Lufx comparison. And it's just so off. Josh: This essay that was reported, this book that was published, it's off by like Josh: several orders of magnitude. It's not even close. Ejaaz: I mean, I'm looking at some of the things that they got wrong because someone Ejaaz: did a breakdown of this book that's been quoted so many times and that's behind all these headlines.

Ejaaz: The book reported 5 million liters as the city's annual water use, Ejaaz: but that was misconstrued. Ejaaz: What she actually meant was 5 billion cubic meters. Ejaaz: For understanding here... Josh: Liters and cubic meters are very different things. Ejaaz: Very, very different things. So she confused liters with cubic meters, Ejaaz: which is already 1,000x error. Ejaaz: But then she said the data center would actually use 3% of the municipal water system, not 1,000x.

Ejaaz: That's where we led to the 4,500x factor off that she was. Ejaaz: So basically, it is an absolute incorrect piece of reporting that has been spread Ejaaz: by some of the most important and popular media publications that we've seen. Ejaaz: Josh, I want to kind of take your golf course thing a little further because Ejaaz: I actually quite like that. Josh: Okay, we'll keep going on the golf courses.

Ejaaz: So what's the number one state that's been getting a lot of protests about data centers? Arizona. Ejaaz: Let's just look at the golf courses in Arizona. Arizona has 370 golf courses. Ejaaz: Each golf course consumes about 1 to 2 million gallons of water per day. Ejaaz: So that's around 400 to 800 million gallons per day for an Arizona golf course, Ejaaz: right? For all the Arizona golf course.

Ejaaz: If you compare that to data centers that consume 905 million gallons, Ejaaz: or rather 0.12% of county water, Ejaaz: That's like just in Arizona alone. So golf courses in Arizona, Ejaaz: 29 billion gallons per year. Ejaaz: And a data center, the biggest one in Arizona, or actually collectively all Ejaaz: of these, actually, my correction, is just 0.12% of that.

Ejaaz: 900 million gallons per year. So it's obvious that there's just a lot of misinformation out there. Ejaaz: And I think it's really important to just bust this wide open completely. It's just wrong. Ejaaz: And part of me thinks that a lot of the hatred, if I'm being honest, Ejaaz: Josh, I think comes from people kind of equating AI to enriching people that they don't maybe like. Ejaaz: Like, I get it. Like, AI will be used as a tool to enrich billionaires even further.

Ejaaz: It's so wasteful. I don't use it. What people don't realize is the average activity Ejaaz: that you do on a weekend or the average bit of food that you might consume, Ejaaz: buying a burger, 668 years straight of using Grok 30 times a day. Ejaaz: People don't do that that often. Ejaaz: It's just, it's important to level set, in my opinion. Josh: Yeah, there are a lot of valid criticisms. This is unfortunately not one of them.

Josh: But there are some instances in which it does, or it has in the past actually Josh: affected localized effects where very small towns have actually felt an impact of this. Josh: They date most recently back to 2022. Josh: There actually hasn't been many sources recently that have determined that it Josh: is making a problem, but there have been some instances. Josh: The first one was in Oregon, where a Google data center It consumed 29% of the Josh: town's local water supply.

Josh: There's another one from 2019 in Virginia. Josh: Virginia, famously, that's where a lot of the internet data center runs. Josh: I think a majority of the data center is run from these Virginia AWS servers. Josh: It consumed 63% of Lodon County in Virginia in 2019. Josh: But since then, there really hasn't been that much of an interference with the local water supply. Josh: And there have been solutions proposed from the companies who are most responsible for.

Josh: Creating that strain. And that is Google, Amazon, Meta, and Microsoft, Josh: who all committed to be water positive by 2030 using new cooling techniques. Josh: So that vapor chamber that we were talking about earlier will be done at an Josh: industrial scale and will actually be able to preserve 100% of the closed loop Josh: water supply. And I think that's going to be a really big deal. Josh: That paired with direct to chip cooling is also going to make a big difference.

Josh: If you remember our Vera Rubin episode, where we talked about how the new chips Josh: are cooled, the actual cooling temperature, if I remember right, Josh: it was like 115 degrees Fahrenheit that it could be. Josh: So now you could actually cool these ships with hot water. It requires much less. Josh: Air cooling is getting a little more interesting. Josh: There's a lot of solutions coming along the way that will make this a lot more Josh: resource or a lot less resource intensive.

Ejaaz: Cool. So if I were to summarize some takeaways for this myth busting episode, Ejaaz: there's a few that come to mind. Number one, Ejaaz: The numbers just don't support the panic that people are putting out there. Ejaaz: The fact is, AI data centers currently today, they might change later, Ejaaz: Use a fraction of what golf courses, agriculture, the t-shirt that you're wearing consumes to produce.

Ejaaz: So even if we tripled AI water usage today, it would still pretty much be a rounding error. Ejaaz: Number two, the context matters. You can't confuse 1,000 liters with 5 billion Ejaaz: or whatever the number was, cubic meters of water.

Ejaaz: That is super important. And comparing water consumption for AI data centers Ejaaz: and your average burger joint might just be the comparison that you need to Ejaaz: kind of like set you straight and be like, okay, well, maybe this isn't that Ejaaz: important going forward.

Ejaaz: And then the third thing, which I think is kind of underspoken about quite a Ejaaz: lot is, I think, if I were to guess, we're moving towards a world where we end Ejaaz: up actually using less water for data centers, right? Ejaaz: Part of it is due to the different systems, like the adiabatic system that you mentioned, Josh.

Ejaaz: But also, I think a lot of these AI companies are going to start building water Ejaaz: recycling plants to kind of push that 90% water renewability figure that we Ejaaz: mentioned earlier much, much higher. Ejaaz: I think Colossus 2 and Elon is doing that right now for Colossus 3, Ejaaz: actually. They're building out a water recycling plant. Ejaaz: So I think overall, this is a nothing burger, pardon the pun.

Ejaaz: And we're going to look back on this. Yeah, we're going to look back on this Ejaaz: in the future and realize that we're consuming water in much faster ways in Ejaaz: so many other industries that we aren't currently protesting. Ejaaz: So keep quiet, eat your burger, and let the AI flow. Josh: So if you had to guess what the next narrative would be, that has a negative Josh: spin on it. Do you have any ideas? I think my answer is going to be energy.

Josh: I think they're going to start to converge on the correct argument, Josh: which is the energy consumption on a localized scale, starting to actually impact Josh: the cost per kilowatt of the average person's home. Josh: And how does that get offloaded? Well, a lot more natural gas turbines, a lot more solar panels. Josh: And the process of scaling that up is happening, but it's happening slower than Josh: the scale at which they're consuming.

Josh: So when you take a gigawatt data center like Colossus 2, that is consuming the Josh: equivalent collective output of San Francisco localized to a small town in Tennessee, Josh: there are impacts there that are real. Josh: It's just a matter of time until those kind of get uncovered and then get dealt with. Josh: I mean, they are dealing with it quickly, but there is a real strain happening Josh: on some grids that are localized to where these data centers exist.

Ejaaz: I agree with you. And I actually think the numbers that will be quoted on headlines Ejaaz: about that specifically will actually be closer to home and to the point because Ejaaz: It's simple enough to kind of scale new water techniques to kind of cool stuff down. Ejaaz: Like we reported on a previous episode, I think like two weeks ago, Ejaaz: that they're not even using cold water anymore. Ejaaz: They're using warm water, 45 degrees Celsius or Fahrenheit, which is super warm.

Ejaaz: I think it's like 90 degrees Fahrenheit to cool these systems down. Ejaaz: I think it's a different game with energy where we actually do have limited constraint. Ejaaz: It takes so much more work and expertise to scale that. and we're going to have Ejaaz: to tap into like town supply or city supply. So I agree with you. Ejaaz: I'm looking forward to kind of like unpicking that one in the future. Josh: Yeah, we'd be good at this. We should design the next PSYOP against our own industry.

Josh: I think that would be much better to consult with the experts prior to doing this next time. Ejaaz: That's hilarious. Josh: But yeah, I guess that concludes the Mythbusters episode on the first one that Josh: we'll be dealing with, which is the water consumption and the new metric that Josh: is burger or tokens per burger. Josh: And in the case of our tokens per burger metric, the cost is very low. Josh: And I don't think this is anything to actually worry about.

Ejaaz: That was the end of the episode, but I'm actually curious whether you guys enjoyed Ejaaz: this. We hope you guys learn something new from this. Ejaaz: Josh and I kind of like went back and forth on this, whether we should do this episode. Ejaaz: We realized like a myth-busting series could be really cool because there's Ejaaz: just a lot of myths and false claims out there and we face it every day. Ejaaz: We try and unpack it, spend all of our time figuring this stuff out.

Josh: And I love that show growing up. It was great. Ejaaz: Yes, same. Actually, dude, maybe we need to, we need to come back with some Ejaaz: glasses for the next myth-busting episode. maybe a trench coat, Ejaaz: like a fedora, if we really lean into it. Josh: I'll write in my magnifying glass. Ejaaz: Exactly. But yeah, if you found this informative, if you found this interesting, Ejaaz: and you aren't subscribed to us, which apparently is around 70% to 80% of you,

Josh: Please do so. It's a percentage that is far too high. Ejaaz: It's way too high. It is actually almost as high as these false headlines that Ejaaz: we keep seeing about water usage for data centers. Ejaaz: So if we're describing you currently, it takes two seconds. Please subscribe. Ejaaz: Please turn on notifications. Ejaaz: If you're listening to this on Spotify, Apple Music, or wherever the hell you Ejaaz: listen to this on, please also do the same and give us a rating.

Ejaaz: It helps us out massively and puts our videos out to way more people so that Ejaaz: we get more eyeballs on this and we can keep producing better videos for you. I think that's it, Josh. Josh: Anything from you? Yeah, and just a small reminder about the newsletter. Josh: Today, we just dropped a new piece that coincided with the roundup, Josh: which was a weekly roundup of the five most important, noteworthy things that Josh: you want to be informed on.

Josh: We post that twice a week, once every Wednesday, one's every Friday, Josh: one's a thought piece, one's a recap. Josh: So you can join 100,000 other people who are also subscribed to getting the Josh: info prior to these episodes dropping. Josh: And I think that concludes it. That just wraps it up. So thank you so much for Josh: watching as always. And we will see you guys in the next episode. Ejaaz: See you guys.