[00:00:00] Jacob Haimes: Welcome to MuckrAIkers, where we dig through the latest happenings around so-called AI. In each episode, we highlight recent events, contextualize the most important ones, and try to separate muck from meaning. I'm your host, Jacob Haimes, and joining me is my co-host, Igor Krawczuk

[00:00:17] Igor: Thank you, Jacob. Uh, this week we're gonna add our voices to a debate around, data centers and water usage because there's not enough content about that yet 

[00:00:28] Jacob Haimes: Yeah. And more specifically, there's not enough good content, uh, because it turns out this is actually, like, a, a hard problem. And so I feel at least, and I think Igor also feels similarly, uh, that, uh, most of the coverage is not the greatest. Um, and a big reason for that is just that you sort of need to be an expert in, like, three different areas to, to really understand, uh, and, you know, uh, work with sort of the situation, uh, in an adaptable way without, like, really digging into a given example.

[00:01:04] Jacob Haimes: Um, and I mean, I'm not that. And, uh, so we're not gonna say what we think the def...

[00:01:13] Igor: areas that you would say you need to be expert in? 

[00:01:15] Jacob Haimes: Uh, so, so, like, uh, data center-- or, like, evaporative, uh, cool-- like, cooling power plant technology, so that's sort of like a mechanical engineer or environmental engineer. Uh, uh, but then also you want to know about the, uh, sort of like environmental impacts of...

[00:01:37] Jacob Haimes: So that sort of then goes towards environmental engineer, but then you also need to know about, uh, the, the chips themselves, um, and, like, that portion of it. Uh, so I guess I'm thinking of, like, you kinda want a AI hardware person plus a data center, uh, slash power, um, system usage, plus, uh, like, environmental engineering person.

[00:02:07] Jacob Haimes: And that's very rare.

[00:02:13] Igor: So instead, what we're gonna do is we're gonna give you seven handy questions to ask, we're gonna give you like a worked example, uh, and also use this to like give some of the background knowledge that despite not being experts, you know, we just a bunch and we're now not experts, but we know a bit more than I think most people who open their mouths on, uh, on this stuff. 

[00:02:42] Jacob Haimes: Yeah. I guess I, I did do a little bit of like, you know, the power plant design stuff in, in undergrad, so that's like my limited bit of knowledge, um, here as well. But yeah, really it's, it's just more like breaking down the problem and taking time to split it into the questions that are important, and then acknowledging that we can't know, um, in every single situation, and it's a nuanced question and it's difficult.

[00:03:10] Jacob Haimes: Um, which gets to

[00:03:12] Igor: bit of, of, of a cop-out act- uh, ac- actually. I think, like, we can know. It's just, like, you need to just ask this question, and it is much more difficult than often made out to be. And at the same time, like, if we dig into this, it's not that hard to see that there's, like, shady shit ongoing. 

[00:03:27] Jacob Haimes: Yeah. So when I say we can't know, I mean w- without spending a ton of time on it, um, it is very difficult to tell in a ge- like you can't make a general statement about this. It needs to be situation specific, uh, is more of what I'm saying. Like we can't make a general claim

[00:03:48] Igor: In what direction you mean? I mean, we can make a claim like, you know, like massive build of an AI data center will be drawing down a lot more resources. And then like 

[00:03:59] Jacob Haimes: Yes. But-- So I, I guess that's, uh... I, I do think it's important to think about, like, comparative use and whether you're making a big deal about something because it's new versus making a big deal about it because it's a problem. And, like, maybe it indicates that you should've been making a bigger deal about the things, you know, in the past as well.

[00:04:28] Jacob Haimes: But then it's not a question about, you know, the AI data centers. It's a question about data centers in general. And so I'm-- Y-you can make claims that are, like, directional, right? We can say number will go up if an AI data center is next to you. Um, the question is, what if, uh, what if there was, um, you know, a data center that was not an AI data center that was put next to you?

[00:04:57] Jacob Haimes: Uh, like, how does that compare? Uh, and then also if... Like, is it actually worth it? Um, and, and, like, I guess I'm not-- I, I don't wanna wade into that conversation here.

[00:05:10] Igor: Right, like, like, uh, like I would w-wait a little bit. In at least like I think we should say like a bit like the, like the sides, if there are some. Like, you know, like the caricature of this is that, you know, like there's like one straw man version is like, uh, that an argu- a side would argue data centers are destroying the environment and like the water usage is, uh, completely, um, out of whack, and it's gonna be the thing that exacerbates like, um, water and power scarcity. And then the other side would argue, "Nah, of that, all of the, those claims are false and over-reported, and all of the activists are just NIMBYs who stand in the way of progress." And somewhere in between those two caricatures, there is like more reasonable versions of these claims, and then also like what people are actually arguing 

[00:06:03] Jacob Haimes: But also, like, very importantly, I guess what I'm arguing is You-- One can hold both of these claims and b- and have both of these claims be true for a given data center. Uh, n-not the same given data center, but two different data centers because of a bunch of, you know, these questions that we're about to, to talk about,

[00:06:27] Igor: Yeah. So like, so like for, for, for-- if you can have both of things for different days at hand, but not the same one. Yeah. Vedant Agrawal. 

[00:06:34] Jacob Haimes: Yes

[00:06:35] Igor: Should we go into our magical seven questions then that will guide the audience in navigating the nuance and, uh, complexity of this questio- uh, of this topic? 

[00:06:47] Jacob Haimes: Yeah. Yeah, and I guess we'll also just mention, like, we might add some stuff in the show notes about other, uh ... Like, uh, this is a hot topic. People are saying stuff. We're not, we're not gonna mention what other people are thinking. We're just going to talk about how, like, the, what the questions are that we think are important to em- empower a person to then make an educated decision on do I personally think that this specific data center is good or bad?

[00:07:17] Igor: Or alternatively, that this source or this article that I just read is done in a way that seems trustworthy or, or not. Because if 

[00:07:26] Jacob Haimes: Yeah

[00:07:26] Igor: discuss some of a question that we, uh, will bring up, they might be bad. And, uh, we will not go into it on air, but we do see and we do judge, um 

[00:07:37] Jacob Haimes: So true. So

[00:07:39] Igor: what, 

[00:07:40] Jacob Haimes: do you wanna give the overview?

[00:07:41] Igor: people should-- Ah, yeah, let's do overview. Um, the seven magical questions, uh, we have are: Where is it? How exactly is it being cooled? like where does the water come from? Uh, how is it powered? And where does the power come from? Uh Who owns it? Who pays for it? And it actually an AI data center or is it just a normal data center


[00:08:14] Is It Really an "AI" Data Center?
---

[00:08:14] Jacob Haimes: Yeah. And I, I think we can start with that last one, 'cause it's, it may seem like trivial, um, but it's actually a

[00:08:22] Jacob Haimes: good question to ask because even though there isn't a formal standard or distinction between a data center and an AI data center, the requirements and costs, uh, not just, uh, mo- monetary, but also, uh, environmental, uh, impacts and that sort of thing, are actually significantly different between a, like, sort of traditional data center and an AI data center.

[00:08:50] Jacob Haimes: And from what I could tell, and I, of course, this is, like, fuzzy and, you know, based on information that I have at hand, but, um, at minimum, uh, an AI data center is roughly, like, four times the power use per area of a traditional data center. Um, and

[00:09:13] Igor: maybe to give it maybe kind of like a traditional data center is a data center that hosts like mainly CPU and hard drive, like holding servers. The, um, and these will usually be co-locating many different, um, workloads. So if you think of like, uh, a, a large provider in this is in Europe, it's, is Hetzner.

[00:09:43] Igor: They have a data center in Frankfurt that will power most of Central Europe, and lots of like small sites will have like either like a server running there, um, as in like a physical server is rented out or co-located by one company, or they have like the Hetzner cloud where people share virtual hosts there. then the load on this will vary based on like the customer base of the site being hosted there or like the company using it. there's like decorrelation, which means that might be like a peak on one side, but usually that will not bring the data center to like one hundred percent usage. And this is different to AI data centers, which are mainly GPU centric.

[00:10:36] Igor: And usually if they're rented out by like hyperscalers, like the, the workload will go up in one go because they will have, you know, a section of, let's say like the Groq or let, let's be, let's, let's be more realistic. A section of any AI except Groq might be, uh, go into like high demand. 

[00:10:59] Jacob Haimes: Yeah, yeah, yeah

[00:11:01] Igor: and, uh, and then, uh, even just a single instance of this needs like multiple servers with eight GPUs each, which, uh, which then brings like a massive load up all at once and keeps it up as long as like usage, um, 

[00:11:19] Jacob Haimes: Yeah. So 

[00:11:19] Jacob Haimes: it's it's usage pattern is very like, uh, spiky, I guess, zero or one, uh, which is, is more difficult on the grid and where it's drawing power from. Um, and then also very importantly, like I said, it's using more, and so that generates more heat. Uh, and so to keep the systems running at the speed that you want them to, you need to be able to get the heat out.

[00:11:43] Jacob Haimes: And what this changes is that you can no longer just use an air-based, um, cooling process. In, in typical, uh, serving-- servers, uh, or data centers, um, you just use air. You may have like, um, a high, uh, flow of air, right? So there's a lot of air going in and going out so that it carries heat with it. Um, but it would be manageable with air.

[00:12:14] Jacob Haimes: That is not the case

[00:12:16] Igor: Like you, y- y- usually you have like a hot aisle and a cold aisle and you keep it separate and keep like a high differential because then that gives you like more power efficiency, which is s- something we're g- we're gonna g- uh, go back, uh, go back to. But you don't have any direct water consumption 

[00:12:32] Jacob Haimes: Right. Well, you don't have-- you don't need water. Water isn't a question here, um, necessarily. Um, I guess there's like, you know, potentially if it's related to the power usage or, um, if you're, you know, having a pool party or something, that might cost some water. Uh, but for AI data centers, you need a, a higher, uh, exchange of heat because more power is being used, because more heat is being generated.

[00:12:59] Jacob Haimes: So you need to use a liquid cooling system because the specific heat is higher, which basically means it can absorb more, uh, heat more quickly. Uh, and so instead of having air going through different areas, you have special systems that allow for, you know, pipes of, of the cooling, uh, fluid to go, you know, right next to the, the chips, and then that draws out the heat that way


[00:13:28] How Is It Cooled, and Where's the Water From?
---

[00:13:28] Igor: And so this is when like transition to the second thing was like, how is it being cooled and like where is the water coming from? Is like, like to the water adjustment. I was talking about like direct water usage because like what, what is a common like sleight of hand in this discussion is that like to only look at like one part of a system

[00:13:47] Jacob Haimes: Mm-hmm

[00:13:48] Igor: like if you want to do like, uh, like if you want to assess the environmental impact on like a global level, not only for your community, but like on a global level, then the water usage that goes into the energy production is also a factor that goes up if you have more energy consumption, obviously. So the, the, the, uh, the later question of like how is it being powered actually also influences like the water consumption because if your power source requires water cooling, for example, like in France, they often need to shut down the nuclear reactors if, uh, they don't have enough water, um, because the rivers get too hot. that would be a, be a factor. But, uh, the direct-- Yeah 

[00:14:33] Jacob Haimes: I, I think it's worth pausing there as well because one of the questions I had initially when we were talking about, you know, maybe doing this episode was Why, why is water important now all of a sudden? Why did-- It, it feels like, you know, before AI, this wasn't something that we were talking about at all.

[00:14:53] Jacob Haimes: People didn't really care about water usage, and now all of a sudden...

[00:14:56] Igor: people 

[00:14:57] Jacob Haimes: Well, sure. Uh, but the, the popular knowledge, uh, the zeitgeist, if you will, uh, did not seem to care as much about water usage, and now all of a sudden it is a much bigger deal. So why did this happen? Uh, and the thing that I sort of came across and, and what my now, I guess, hypothesis is, um, is that as you increase or as you decrease energy usage, one way to sort of allow yourself to do that in this, um, power plant, uh, or, or like water cooling setup is to increase the amount of water that you're using.

[00:15:41] Jacob Haimes: So instead of using, uh, a process where-- which is like a lot of what Europe uses, where it's an entirely closed loop system, you don't ever actually use up any water outside of the loop. It's just using, you know, a fluid within the pipe and then some other sort of heat exchanger which is powered by electricity.

[00:16:04] Jacob Haimes: So that is power consumption, right, to, to do that. However, instead of having some sort of heat exchanger that is powered by external electricity, you can have, uh, you know, river water in theory running next to the, uh, you know, coolant loop and that will exchange the, the heat there and then you are no longer using up as much power, but you are putting a bunch of extra heat into the water.

[00:16:43] Jacob Haimes: Another option would be to use evaporative cooling there

[00:16:45] Igor: y-you put it back. Yeah, so but like, so this is important, like you're still withdrawing the water. 

[00:16:51] Jacob Haimes: Yeah

[00:16:53] Igor: And like, uh, like it's not a given that this water can be used for the same purposes like, uh, uh, later on. And you might not be using river waters. Um, so as, as we, we'll see this in our work-work example, you might be using like drinking water because that's like the, the easiest, uh, to access water 

[00:17:10] Jacob Haimes: It's not just that it's the easiest to access, it's also that the pipes are, um... But you don't wanna put fluid that has a bunch of shit in it into pipes because then the pipes degrade faster. So it'd be a lot better if you had liquid that you knew what the kind of content it was and it was, you know, relatively clean.

[00:17:31] Jacob Haimes: Um, and so that's the other prime reason why, why drinking water is used in a lot of these cases

[00:17:39] Igor: And then this is-- So like this is withdrawal and then returning it can be consumed. But, and now there's also consumption, which would be for evaporative cooling 

[00:17:47] Jacob Haimes: Yeah. So instead of having, you know, uh, an electricity-powered heat exchanger or having your data center next to a river, you know, I mean, that's a little bit fantastical, but you get the idea. Um, you could have an evaporative cooling system which takes the, um, you know, liquid cooling loop and uses evaporative cooling, which is actually really cool, uh, because, you know, energy is used up when water evaporates, and so

[00:18:19] Igor: Ba dum tsssh. Very cool 

[00:18:22] Jacob Haimes: Nice.

[00:18:23] Jacob Haimes: Uh, wasn't even trying to do that. I actually just think that the evaporative cooling process is very interesting. Um, but regardless, if you are using this type of, um, cooling process for your, um, liquid coolant, then you are going to have consumptive water usage, and this requires a lot of water because you're essentially taking the water from the source, you're running the, um, you know, coolant loop through it while that coolant loop is superheated, and the evaporative cooling sort of, uh, silo is removing the energy from that, you know, closed loop, uh, because the evaporation is, is, uh, withdrawing the energy.

[00:19:17] Jacob Haimes: So, um Yeah. Anyways, there's a lot of different ways you can configure this, and efficiency is really important as well. So depending on the pipes that you're using, the, you know, the type of liquid coolant that you're using inside the closed loop portion, um, you can significantly impact the efficiency of your, you know, heat exchange process.


[00:19:48] Where Is It?
---

[00:19:48] Jacob Haimes: And that's important as well, or related to another question, which is, like, probably the biggest question, I think, which is where is it? Um

[00:20:03] Igor: Well, uh, d-d-depends, like, like, the-- it's a, it's a question that determines the most how like, the claim that, like, something is harmful 

[00:20:13] Jacob Haimes: Yeah

[00:20:14] Igor: w-uh, uh, w- uh, would be, Except in some extreme cases where it's actually harmful everywhere, which we'll get to, uh, 

[00:20:22] Jacob Haimes: Yeah. Yeah. That's true. Um, but yeah, in, in general, like, uh, the one question you can ask, um, that would give you the most sort of information about, uh, whether a data center, um, is something that you can get behind or not is, like, where are you putting it? Um, and there are three different categories that we think are important here, um, to, to, like, dig into within this question.

[00:20:50] Jacob Haimes: Uh, the first, since we're already on the water, uh, question, is climate. Um, so knowing whether the location is constrained in terms of total water and whether it's constrained in terms of drinkable water, if drinkable water is being used and if it's not. Um, because if you are withdrawing water from the pool of water that might be used for drinking water, then in the event of a drought, you could, you know, put yourself into a situation where now your community is significantly less resilient because you don't have that excess water anymore.

[00:21:32] Jacob Haimes: So,

[00:21:33] Igor: like make, making this m-more concrete, like if, i-if there's like enough rain to replenish what you're taking out be fine, or if there's like enough river water coming through, like, uh, th-that is like, like potable stuff like this 

[00:21:48] Jacob Haimes: Yes. Uh, so that's like one important aspect of it and it's, uh, critical to know, uh, I guess like the water situation more broadly. And that then gets into a whole nother field of sort of like how the water, uh, you know, sharing and, and consumption and, and stuff works within a given place. In addition,

[00:22:12] Igor: f- very fun, yeah.

[00:22:13] Jacob Haimes: yeah

[00:22:14] Igor: you said, you said a thing as well, like I think I would say that's even more important is like, is it like the community? Is like what is like the community that you're next to?

[00:22:26] Jacob Haimes: Yeah. Well, okay, before, before we go to that, 'cause that's sort of one of the s- the second portion of where is it, I do wanna mention that the efficiency of the plant is also impacted by the environmental constraints, uh, or conditions of the location. So, uh, you can't get a power plant to be as efficient in a place that is very humid if you are using evaporative cooling.

[00:22:51] Jacob Haimes: It is like a worked example sort of there of why climate would be relevant to efficiency of the power plant. Uh, but then moving... Sorry, did you have something to say there?

[00:23:06] Igor: The ideal environment is, is like cold and dry outside, right? 

[00:23:12] Jacob Haimes: Um, yes. Yes, that is correct

[00:23:17] Igor: Um, so, so like in general, like you have like a heat differential, uh, for any machine to work like, like, like more and more, uh, efficient. 

[00:23:23] Jacob Haimes: Yes

[00:23:24] Igor: to go to the community aspect now 

[00:23:26] Jacob Haimes: Well, you did too. So

[00:23:28] Igor: Yeah. Um, 

[00:23:29] Jacob Haimes: you 

[00:23:30] Jacob Haimes: can, you can, start it off

[00:23:32] Igor: like there's a tendency of putting v- the loud, obnoxious, horrible thing next to the poor people rich people have lawyers and they can fi-, uh, they can fight you better 

[00:23:44] Jacob Haimes: Yeah

[00:23:46] Igor: and usually that means that like you're putting obnoxious thing next to already vulnerable communities that also might already be facing, you know, like various scarcities like water scarcity, li- uh, uh, li- like problems of, uh, of like, um, cooling themselves.

[00:24:05] Igor: Like we had, we had a heat wave in Europe recently and we also had discussions around data centers because like, um, dumping the heat into the air doesn't necessarily come for free. Like, like, like it can get noticeably hotter right next to a data center you're, if you're at like downwind of one, especially if they produce the power on site and then dump the heat out, like it might be like a miserable place, uh, uh, to live in. Also accounting for like noise, 

[00:24:38] Jacob Haimes: Sorry,

[00:24:38] Igor: noise 

[00:24:39] Jacob Haimes: that again

[00:24:40] Igor: and things like this 

[00:24:42] Jacob Haimes: Uh, say, say that again

[00:24:44] Igor: Um 

[00:24:45] Jacob Haimes: Also accounting for noise pollution

[00:24:47] Igor: It does get noticeably hotter, um, if you're right next to a data center, and especially if they're producing the power on site and you're like downwind of them, can be a quite unpleasant, like, living experience. And that's before you start to talk about things like noise pollu- uh, pollution. There might be like an ongoing like hum or like, uh, 

[00:25:06] Jacob Haimes: Yeah. 

[00:25:06] Jacob Haimes: Well, even if you can't hear it, right? I- there's, uh, some actually really interesting sort of like, uh, work that I think it was just like a YouTuber did- Mm-hmm ... um, that looked at, um, like, uh, readings of, uh, very, uh, like i- infra noise,

[00:25:26] Igor: yeah 

[00:25:27] Jacob Haimes: um, next to data centers. And it's like, you know, incredibly loud infrasound.

[00:25:33] Jacob Haimes: So we as, as humans cannot perceive it, but it is there. Uh, and then that coupled with like health concerns arising, uh, in the communities surrounding the data centers, um, you know, is, is like a, a real thing that's happening.

[00:25:48] Igor: It can also like have an effect on like machines because they are constantly a bit, like a bit, a bit vibrating. Uh, and that's like before the g- the grid effects come into play because if there's like this like up and down, up and down, workloads that can actually like degrade the fridges, uh, uh, uh, uh, uh, because like the large machines in your house, they're coupled to the, to the grid and they actually like get a bit of a like elec-electrical shock if there's like grid variation. So all of this stuff does matter and it's gonna be like next to like poorly built houses usually with a promise of like, oh, there's gonna be good jobs. usually that's not the case. Data centers don't need that many jobs and like the whole name of the game is more and more automation. So it's gonna be like some con-construction for a while and then you have, uh, uh, you're done basically. 

[00:26:40] Jacob Haimes: And this is, I mean, it--

[00:26:42] Igor: you want to talk about? 

[00:26:43] Jacob Haimes: I mean, this is the whole, uh, idea behind, um, like environmental justice, right? So it-- this is not just a data center thing. This is systematically true. Um, I mean, and a, a local example for me is, uh, like Rocky Flats, uh, was a nuclear, uh, trigger power plant, or not power plant, nuclear trigger plant that was near me, and there was a fire there, uh, in the '70s, I believe.

[00:27:10] Jacob Haimes: And so then everywhere around it was designated as, uh, land that couldn't be built on for the next, you know, 150 years or something like that. Um, recently, they removed portions of that, uh, so now there are portions that can be built on, and there is housing that's being built on that land without, uh, it being properly disclosed that it is on nuclear, uh, l- like, uh, land that shouldn't be built on according to, you know, after this thing happened in, in the '70s.

[00:27:45] Jacob Haimes: Um, and who is living there? Well, it's gonna be a lot of people who have less income because those are cheaper houses. So like this is, this is not a single thing. It's-- and we shouldn't be taking it in, uh, i-isolation here. And it-- so it's really important to note, like, I mean, even Colossus, you know, Grok, uh, SpaceX's big data center, the fact that that is in Memphis, which is one of the locations in the US with the lowest cost of living, um, it, like, that, that was for a reason.

[00:28:23] Jacob Haimes: Now, it's also convenient because it's not particularly water-constrained. That doesn't mean that there aren't potential water problems, though. And, uh, like

[00:28:33] Igor: We, we can get to Colossus, uh, like when we have all our questions, uh, uh, 

[00:28:38] Jacob Haimes: Yeah, that's true.

[00:28:39] Igor: like a speed run 

[00:28:41] Jacob Haimes: But I just wanted to, to like mention that this like environmental justice is, is its own thing and it's, it's important to, to know about, uh, that kind of stuff 'cause it's, it's happening a lot more, um, than people expect. Um, also related to where is it is the legal question, um, which I think you had some more specifics on


[00:29:03] Collosus as a Worked Example
---

[00:29:03] Igor: I was gonna go into like first of like powering it because you, you talked about the nuclear plant and, um, and we talked about like the downstream effects. So like, we can maybe also just start also a bit on the work example, like, you know, like, uh, Colossus is, is, is, is, is, is Memph- Memphis. like it is very much an AI da- uh, data center.

[00:29:25] Igor: It is like, it's actually built for this last year in 25, 2025. Um, 

[00:29:32] Jacob Haimes: without the proper permits

[00:29:34] Igor: without proper permits, how is it being cooled? Evaporative cooling. There's lots of like mixed stories, but like, uh, um, they draw water from the aquifer at a rate that is actually like depleting um, and at least to the source that I, that I found on this. Um, and they are talking about building a gray water recycling plant to then use that, it's called make-up water, to it into their, um, evaporative cooling, uh, uh, uh, loop. But as of right now, they're prioritizing finishing the, uh, the power pl- the actual like data center

[00:30:16] Jacob Haimes: I mean, not even finishing the data center, like, like, sorry, not even prioritizing finishing the data center. They have full on stopped construction of the recycling, uh, plant

[00:30:28] Igor: mean, that's what prototyping means, right? Like, you just do one thing and not the other thing 

[00:30:33] Jacob Haimes: Uh, some people prioritize without com- like, you know, completely stopping something

[00:30:38] Igor: No, that's, that's not prioritizing. Like, that, 

[00:30:40] Jacob Haimes: It, yeah. The, I w- I agree with you, but like I just wanna make sure it's clear that they, it's not even like 

[00:30:46] Jacob Haimes: they're they're not working on it and they're, they're like being coy about it. They have announced that they are stopping work.

[00:30:53] Jacob Haimes: So like absolutely zero work is being conducted on this

[00:30:57] Igor: And, um, the, uh, so this is for, for, uh, for, for the cooling. Now Memphis is not a water, um, constraint area in, uh, like in the, in the strict sense and the, and the aquifer is there, but taking water from aquifer c- does come with like risks, uh, uh, uh, because you-- it can lead to like, uh, subsidence, I think is, uh, is, is the term. Like, like the ground can shrink un- under you if it goes on for long enough. this is a thing that should be tracked, and it's very convenient for them to just prioritize one thing and stop, uh,

[00:31:34] Jacob Haimes: It's also particularly bad for, um, community resilience as well because, uh, the reason that we have the aquifers, uh, is so that we can store water for when times are harder. Um, and you don't wanna ever let the aquifers get below a certain level, uh, because if you do, then that messes with all the equipment, and then it makes the whole system sort of not work.

[00:31:54] Jacob Haimes: Um, and if you are reducing your margin for drought by reducing the amount of water that's contained in the aquifer, um, y- you are more likely to put yourself in a situation where because of, uh, you know, a drought that eventually will happen, uh, y- your community isn't able to, you know, sustain itself without external support.

[00:32:22] Igor: And, uh, I have found some, uh, uh, cases where we, we put them into chance of like companies either making an, an, an ac- accidental in their reporting of how much they So this is now withdrawal. We're not talking about con- consumption even. Even the withdrawal, um, takes, uh, some-- like the water in the first place.

[00:32:48] Igor: If it eventually goes back to the aquifer, that's a, that's a separate process, uh, that, is again a case by case it, but like the taking has its own effects and they reported less than they actually took. And in other cases, uh, like they had like agreements that allowed them to basically like arbitrarily scale their demands against way more than was originally agreed on. Um, it was just like in the fine print, so it's like legal, then they-- there was like activists going against it because it would have created exactly the situation that we describe. Um, so that is like is it? It's like Memphis, like close to the border with, um, Mississippi ap- ap- apparently. Um, how is it being cooled? Uh, uh, uh, cooling, very much AI data, data center. How is it being powered? Shady as fuck. draw, they, they draw from the grid in Tennessee, but they have, um, unpermitted, um, for, for, for a good chunk of their operation, gas turbines running on, uh, on like trailers in Mississippi because they're skirting the re- regulations of like when is a power source like stationary and therefore needs, needs a permit versus just like a temporary power source. And there's like exchanges with the EPA or what is left of the EPA and activists around this. But they're basically being powered by like a hodgepodge of power sources, um, where gas turbines are very popular because you can very quickly spin them up. But they're really not, not good for, uh, on like a total like efficiency and emissions basis. Other popular da- uh, power sources for data centers are coal plants. Uh, I think some nuclear plants are also being restarted or prolonged

[00:34:42] Jacob Haimes: In, in theory. Um, but yeah, uh, it's hard to know for sure. Um, there, there are reasons to be, you know, uh, tentatively optimistic about like certain n- nuclear power, uh, things. There are reasons to be skeptical of other aspects and, and so like, uh, u- until it's here, I, I think, uh, and, and like actively being used, 'cause it is being used in some cases, but, um, like we really can't be


[00:35:14] How Is It Powered?
---

[00:35:14] Igor: so, so like to be precise here, like, um, there's a large push for nuclear power for various reasons that are kind of an-- as, as an episode. A lot of these that are actually going live are old power plants being restarted, often called like behind the meter, as in like they're like off the grid, but they're just directly feeding the thing that they're being like, like that they're feeding. 

[00:35:41] Jacob Haimes: Mm-hmm.

[00:35:42] Igor: And there's an active push for like startups that are trying to sell small nuclear reactors, it's being waved about as like a thing that's gonna fix all this. But the highest percentage of new loads, additional load is usually just off the grid, which does cause like increased grid prices, grid instabilities, other, other types of issues for people who are always consuming. Coal plants going ba- uh, back up or being added or gas turbines, because those you can eas- easily add in a given region. These are like the three major ones. Ideally, you want something like hydropower and shit like that, but that's basically only what Norway and, and, uh, uh, China is realistically, uh, using. Um, 

[00:36:34] Jacob Haimes: Yeah, 

[00:36:34] Jacob Haimes: and, and that's more of like a broader, you know, uh, energy question, but it's still relevant to, to this because it is an increase in energy cost. Um, and so where are you getting that power? How is that being generated is still, you know, very much relevant.

[00:36:49] Igor: consume w- water. 

[00:36:50] Jacob Haimes: Right 

[00:36:51] Igor: um, uh, that, like, depends where they are, but, like, that's, uh, that adds to the total water cost. Um, so Colossus uses, um, these, like, trailer gas turbines and doing a lot of sha- shade shit with them.

[00:37:03] Jacob Haimes: And the reason, uh, if that's important just in more general sense, is the legal regulations at, in a given location and the ways in which whoever the owners are can attempt to subvert or, uh, deny or, um, you know, avoid these legal regulations. So, uh, you know, Musk, uh, has gained, uh, a lot by doing shady, uh, stuff that, uh, sort of plays things within the US legal system against itself.

[00:37:41] Jacob Haimes: Uh, so the fact that it's, uh, they're, you know, mobile... technically mobile units that are in another state means that multiple states have a, like, a say in this, but then also that makes it harder to do things. So, like, uh, it's very int- th-this choice is very intentional. Uh, and it is being used to prolong the amount of time that it can operate before, uh, you know, regulators come in and actually say, "Wait a second, you shouldn't be doing this," uh, simply because of, you know, bureaucracy and things like that


[00:38:15] Who Owns It & Who Pays?
---

[00:38:15] Igor: We-- the next two questions is, like, who owns it and who pays? Which are separate questions. In the case of Colossus, it's actually kind of rare in that, like, SpaceX AI, like the mega parent company, actually to flat out own this facility. And usually you will find that, like, it's owned by a, what's called an SPV, special purpose, uh, vehicle, which will be like a company found, uh, found specifically to, like, get a contract from a parent company that they're gonna, like, use all of the compute that this SPV will own, and they will take a promise and raise a bunch of debt banks and other investors, and then construct the data center. And the reason why they do this is to take a debt off the balance sheet of the company, which is its own thing that's gonna set up a whole financial crisis. Um, we'll put something in the show notes there. But, uh, the specific ownership structure also matters because it de-determines who you can sue, is, like, actually, like, uh, the spokesperson or, like, the person you're getting into, into contact with, and that is often quite opaque.

[00:39:27] Igor: In the case of SpaceX, it's relatively clear. The who pays more nuanced because a lot of these, um, data centers are built with, like, incentives or abatements or other types of deals because, you know, in theory, having, you know, like, facility in your community you, like, the chance to raise, like, property taxes.

[00:39:54] Igor: Uh, you might have some jobs, uh, coming into your area. You would have, like, a construction build-out that's gonna be, like, a brief but, you know, meaningful increase in total, like, consumption and so on. so companies do like to play and counties and cities, like, against each other to get these types of incentives. I think in the case of Colossus, there wasn't really much incentivization from, like, Memphis itself. There was maybe some, like, collaboration on some of the, like, infrastructure usage Uh, on this, on the Mississippi side where they do the power, there was an incentive that basically exempted them from all sales and use tax for the equipment and software and so on. Uh, and there's like agreement the county, uh, but it's a bit unclear what exactly, um, it, it goes there. But, um, the unpermitted, um, turbine usage, the overall kind of like laissez-faire attitude that like SpaceX often has with like environmental regulation of like waste dumpage and stuff like that, all of that stuff is not accounted in this.

[00:41:27] Igor: The, the, like externalities are borne just by, by the communities that this stuff is, is, is built around in. other places you might have flat out, you know, like the city foregoing taxes or giving them like millions of dollars to build at their center of, or funding the, funding the cost of infrastructure build out, uh, and stuff like this.

[00:41:49] Igor: It changes the calculus on whether something is actually worth it

[00:41:52] Jacob Haimes: And what's not on our side here, um, is that companies go to great extents to prevent this information from being accessible. Um, uh, I believe, uh, uh, Amanda Ptaszek, um, ha- has, uh, some, some work on this specifically, uh, for, you know, AI companies. But there's also, you know, a number of other groups as well, um, that are, are trying to look into this more.

[00:42:24] Jacob Haimes: The thing is that you can classify some of this information as, like, a trade secret because it would allow one to...

[00:42:33] Igor: yeah 

[00:42:33] Jacob Haimes: Oh, they, they've tried, yeah. But, but then also, um In some cases, like you don't even need to. Like, it's just like they don't report it. A- and then like n- nothing happens because of that. Um, so

[00:42:47] Igor: there's a lot of shady shit going on. Like, like I found a bunch of links we're gonna put into the notes of like, um, Dalles, not Dallas, but like Dalles, Oregon. There was a case of Google trying to classify th-their water us-usage as a, as a trade secret, and Microsoft did a, a similar thing. other places, um, the, there were NDAs placed, I think in the Memphis case might have been it, but this is from m- from my me-memory now, so we need to check.

[00:43:13] Igor: But like where the city officials who are doing the negotiations have to sign NDAs, so they literally cannot discuss it with their constituents or like other politicians. So the usual, like, ac-accountability process is frozen out, and then the moment when the populace would, would get like a, a count of like how much water are we signing away to, to this or that company is after everything has been signed already and there's already like a legal obligation and maybe like contractual damages if you now like change the contract

[00:43:45] Jacob Haimes: And that's a lot of what some of these, like data center, um, sort of, um, commun- community grassroots, uh, you know, counter movements are about. And in some cases, they just don't do anything because those agreements have already been made. So even though the community is extremely vocally against this happening because, I mean There are, in general, very well-documented reasons why you-- why one should be concerned about a data center.

[00:44:19] Jacob Haimes: That doesn't mean that in every situation it's definitely worth saying, "Don't put a data center where you're planning on putting it." Um, but it does mean that, that, like, this should probably be a conversation. Um, and the fact that it's not and it's being intentionally subverted, um, so that, you know, they can pull one over on us essentially, uh, is, is really problematic.

[00:44:45] Igor: And I, I think, like, there's something that, like, you, you mentioned a few times, uh, wh- when we were preparing for this episode is like, in theory, it is possible to do, like, all of the build-out in a way that will be, you know, like, all of the externalities priced in, everything above board, you could, you, you could build, like, a version of Colossus that might be, like, a genuine, like, boon for the city of Memphis, um, as, like, a-- an economically neglected area. 

[00:45:15] Jacob Haimes: Yeah 

[00:45:16] Igor: you could say, like, you could, like, commit to using local co- local construction workers. You could commit to making sure that the infrastructure handle the build-out and making sure all of it is, is done before, and then, like, the company p- uh, pays for the improved water and power infrastructure.

[00:45:34] Igor: And that, that might, uh, have, like, downstream effects for, for the community that are positive, actually. You might have, like, some commitment to source the workers from the local thing. you might use technology that makes the whole thing, like, not zero impact, but, you know, like, for the local community at least kind of like overall, like, sustainable. None of this stuff is really happening. Like, they like to talk about it, but like in the Colossus case, it's a lot of like, "We will do this now," or, "We could be doing this actually now." Like, thing as, like, nuclear power. In theory, maybe that you could build, like, a very nice modular reactor and in theory have, you know, like, some good solution for the, nuclear waste storage afterwards. In practice, that's not happening, and it's using coal and dirtier technology. And this disconnect is also where, like, some of the bad critiques of the critique happens, you know, where, like, people raise, like, valid points of, hey, maybe all this water usage, all this power usage not worth it actually, and we might not want this in these communities for the price that the, companies are paying But then this is being lumped in with maybe some accidentally or through, like, mistreatment, like, worse reporting and just being dismissed out of band or being hand-waved away with these, like, promises of, "Oh, but we could do it cleanly, so your, 

[00:47:12] Jacob Haimes: Yeah.

[00:47:13] Igor: thing is i- invalid." 

[00:47:14] Jacob Haimes: And I, I think the part that really irritates me is that people on both sides take, uh, specific examples and then they say, uh, you know, "And this is evidence that, um, you know, it's all bad or it's all good." Um, or, you know, that it... And it, it, I- it's really not that you cannot extrapolate a specific case because there are so many different variables within this problem here.

[00:47:45] Jacob Haimes: You, you need to go in to each specific instance and understand these questions, uh, to even, like, begin to be able to have a meaningful representation of the situation that's going on


[00:47:58] Is This Too Complicated? Seven Questions Recapped
---

[00:47:58] Igor: I think Jacob is, uh, like wants to be like the voice of reason and nuance here, and like I want to like, uh, like, like applaud that. At the same time, do think you can simplify the checklist to at least get like a broad strokes of, you know, like if it's not using renewable energy, like the water usage is not accounted for fully, like you can make like, like at least like the bad cases are, are somewhat straightforward 

[00:48:24] Jacob Haimes: Sure. But even that isn't being done. So I, I, I'm not saying

[00:48:31] Igor: what, what, what, what, what I'm arguing is specifically the way you describe it, it sounds like it's very complicated and l- you know, y- y- like, like people need to be very careful around it and, like, uh, like, uh, leave it to experts and, like, I kinda want to push against that. It's like 

[00:48:43] Jacob Haimes: Oh, I don't think you should--

[00:48:45] Igor: Yeah. 

[00:48:46] Jacob Haimes: I don't think-- I, I by no means think you should leave it to the experts. I think that's, that is an important thing to say. I'm not saying leave it to the experts. I'm saying do the research yourself so that you, you have the information on it, and you can bring up a specific case.

[00:48:58] Jacob Haimes: And when someone says, "Oh, but what about this data center? What about this other data center?" Um, you say, "But, but we're not talking about that right now. We're talking about this specific case where I can tell you the actual situation, and I can say so with extreme confidence." And now we can have a real conversation instead of pointing at things that, you know, are essentially just smoke and mirrors.

[00:49:21] Jacob Haimes: Uh, I-- And, like, it's not-- That is not something that's gated behind, uh, expertise. That's something that is... It, it's gated behind expertise for someone with, uh, no, you know, prior knowledge of a specific instance to be given all of these pieces of information at, like, at a high level and then be able to make a snap decision.

[00:49:42] Jacob Haimes: That is gated by expertise. It is not gated by expertise to take a specific instance and go into it and figure out what is the situation here, what do I think makes sense based on these different factors, uh, and then have a conversation about it

[00:50:02] Igor: So le- let's do a loop and give people like a compact thing of like, of like what they need to do, like an overview again. Uh, because we did this now once and we can just like crisply summarize. And then I g- I, I want to throw in a bonus thing that we did not prepare for.

[00:50:17] Jacob Haimes: I do also wanna mention one thing that I think I missed on the first pass, um, but I can do that after we do the summary

[00:50:25] Igor: Okay. Um, so, uh, what type of data center is it? Because that determines, like, the load and connected to that you can ask, how are you gonna m- handle the grid impact and the noise i-i-impact and so on. That's like one concrete thing. Where is it? In what, like, neighborhood is it? And like how are you gonna like handle that responsibly the i-impact on, on that community, um

[00:50:55] Jacob Haimes: Oh, uh, also what are the, the legal regulations surrounding it? Um, and what is the climate there as well? Because both of those, uh, end up impacting things that are further down the list

[00:51:08] Igor: Yeah, the overall environmental si- situation, how are you gonna deal with that? And also, like, are you getting any exemptions from, like, protective, um, uh, li- like, like regulations 

[00:51:17] Jacob Haimes: Mm-hmm. that you, uh, might be asking about if you're, like, impacted. And how is it being cooled? How is it being powered? And like, where is it being sourced from?

[00:51:26] Igor: And like all, everything around that. Where specifically, like what I wanted to give people is like the terms of like water consumption, withdrawal are different things and should-- you should like critically look at what they're telling you when they say like ze- zero water and, and also like the, like power usage and like, like power impact, all of these things.

[00:51:48] Igor: Like you should look at it like specifically what they're using and like if something sus, like double-check that. then

[00:51:56] Jacob Haimes: Who pays and who owns?

[00:51:58] Igor: Who pays and who owns, m-mainly who pays, like how, what are like the incentives be-uh, being given? Are you actually getting like the tax benefits or are the tax rebates to the point where basically your county will not see a dime for, uh, like, for, for the whole lifetime of a data center?

[00:52:13] Igor: Because the content of a data center depreciate within three to five years, both because of like just being worn out, because like data c- uh, the GPUs will break if, uh, under constant load, and just the technology cycling through and it becoming e-waste. So that's something to g- to, uh, keep in mind. And, um, there's no guarantee that like if your rebates last for 10 years, that there will be a data c- data center in 10 years.

[00:52:37] Igor: Like this is not like a power center where it, it lasts for 35 plus years. Um, 

[00:52:43] Jacob Haimes: Yeah.

[00:52:43] Igor: and

[00:52:44] Jacob Haimes: the, the thing that I wanted to, to bring up as well was related to the, the water, um, and, and what you brought up with the consumption versus withdrawal. So withdrawal means that you are removing it and then putting it back, and, um, sometimes withdrawal isn't one-to-one, but usually it is because then that allows, um, you know, the data center to say that it isn't consuming any water, right?

[00:53:09] Jacob Haimes: So you can say, "Oh, well, we're only withdrawing. We're not consuming any water." What that means is that water is being used to extract heat and is then being returned to the source as heated water, and usually also with, um, some additional chemicals in it. Not a lot of additional chemicals, but, you know, for cleaning the pipes and, uh, just from the pipes themselves as well.

[00:53:35] Jacob Haimes: That has i- its own ecological and environmental consequences, um, because you're, you're now heating up the ecosystem that you are putting the water back into by a significant amount. So, uh, it, it actually really does still have an impact. It's called thermal pollution, um, and that is often completely neg- completely neglected when these numbers are talked about.

[00:54:01] Igor: Also like, m- the water quality that goes in again, like, uh, being, being tracked because like, it, it's, uh, if in goes drinking water from the aquifer and out goes like thermally and possibly like slightly chemically polluted like gray water, that is st- zero, uh, like water consumption but, uh, there has been a transformation and you can no longer use the water for the same things without, without processing it.

[00:54:26] Igor: Um,

[00:54:27] Jacob Haimes: And then the argument to that would be, well, processing it is relatively cheap, and then the argument to that would be, it depends on where you are. So again, all of these things really matter.


[00:54:35] Bonus: Data Centers in Space
---

[00:54:35] Igor: And, and do it then. Like, are you doing it though? Like, or do, are you talking about doing it? Yeah, so all of this, um, Bonus: data centers in space. 

[00:54:49] Jacob Haimes: Oh my God

[00:54:50] Igor: as of right now, complete bullshit for, again, three reasons. One, you need to cool that shit. 

[00:55:01] Jacob Haimes: Yeah, they don't have a cooling solution

[00:55:03] Igor: in space it is literally the same as your thermos, because the way your thermos bottle keeps it hot and cold, ins- there's a vacuum in the bottle, 

[00:55:14] Jacob Haimes: But Igor, space is cold

[00:55:17] Igor: Space is cold in, in, in the nerd sense, uh, that, like there's no molecules moving, it's act- like heat management is a big problem for existing like satellites like ISS and co. And that is why not having like a massive like power, like power consumption thing. You need to collect solar energy to power the compute, so you need to be in the sun, which heats your vessel, then you need to at the same time reject the heat, 

[00:55:58] Jacob Haimes: Oh, so

[00:55:58] Igor: have a 

[00:55:59] Jacob Haimes: the solution that I h- that is in the SpaceX like documents and stuff, uh, that I think they're going with right now is that there's actually just a continuous flow of coolant that's being ejected into space

[00:56:11] Igor: Okay, great. Like, let's talk about that. Y- you know what that means? 

[00:56:15] Jacob Haimes: You have to get it there? And they don't talk about that at all

[00:56:19] Igor: like, like, this is a low power thrust the whole time. You now need to l- uh, uh, uh, to, to like eject that coolant, like in a way that's gonna like be stable in, in, uh, in, in, in orbit. 

[00:56:35] Jacob Haimes: Eh, that's just a math problem. They can figure that one out

[00:56:38] Igor: So like if we try to do the steelman version of a counter here, you know, as-- because we're-- even if I'm s- not as reasonable as Jacob, I do try to like not be c-- uh, like a pure hater. The gamble that is being made is that, oh, but if we get St- uh, Starship, then like the tonnage goes down, and we can just, you know, like keep sending shit up, and it g- it all becomes amortized. Like, like the gamble is that like you need a massive satellite either because of the coolant or because of the heat rejection. So you 

[00:57:09] Jacob Haimes: Both

[00:57:10] Igor: like the ability to, uh, to bring massive amounts of tonnage up. So, so we just need to make, make the rocket bigger. That will sh- will still not make it economical b- uh, b- because you can just sink it on the ground of the sea or put it like in, I don't know, Arkansas. no offense to Arkansas, it's just like, like it sounds far away from everywhere.

[00:57:37] Igor: Like I, I'm sure, I'm sure it's a, it's, it's a nice place. I don't know your state that well. or Minnesota. I, I hear Minnesota is like large and cold. Like, And like the reason I'm saying is like for stuff to economically make sense, it needs to like out-compete. like, like, uh, for it to make sense to send a satellite to space, it would need to be so difficult to get permitting like everywhere, and we would still want to pay like then like the premium price for that like space compute. And this is before we talk about shit like radiation hardening, where like you literally cannot run the bleeding edge, um, chips in space because cosmic radiation will flip your computations, uh, uh, trying with all shielding, which then needs, it needs to have even more mass to shield it. Data centers in space, as of right now, are bullshit, and you should d-discount it, if they don't t- uh, present you like a fully solved like heating solution.

[00:58:42] Igor: And no, like ru-run coolant and inject it into space is not one. Holy shit, I didn't even know


[00:58:48] Outro
---

[00:58:48] Jacob Haimes: I'm pretty sure that's, I'm pretty sure that's the case, so yeah. Um, anyways, uh, that, that will cover it. That is all the muck that we have for this week. Um, if you like this one, if you think that other people should, uh, get maybe, I don't know, this, uh, tool for understanding data center and, and water usage, um, then please share.

[00:59:10] Jacob Haimes: Uh, you know, why we do this in addition to, you know, feeling good about ourselves and, uh, talking with each other 'cause we enjoy that, is because we think it is valuable to have this kind of, uh, reality check. So thank you, and we'll see you next time