Yet Another Value Podcast - SemiAnalysis' Jeremie Eliahou Ontiveros on the supply/demand dynamics of AI and data centers

Episode Date: February 11, 2025

Jeremie Eliahou Ontiveros, Technology Analyst at SemiAnalysis, joins the podcast to share his thoughts on the supply/demand dynamics of AI and data centers. For more information about SemiAnalysis, please visit:https://semianalysis.com/Chapters:[0:00] Introduction + Episode sponsor: Daloopa[1:22] Supply and demand dynamics of AI; perspective on data centers[9:03] Thoughts on DeepSeek; Jeremie's worry on the whole AI trade in general; AI training[13:58] Is there a rush to generate more data right now?[17:24] Jeremie's thoughts on NVIDIA[26:32] What breaks the up cycle for AI / is subscription ultimately the way to monetize AI[32:17] Interesting power plays (picks and shovels to support AI); bitcoin miners as potential AI plays[40:58] Bitcoin miner/AI play that Jeremie is interested in; deep dive on AIRian[50:27] Management activities from bitcoin miners; why are some of these companies not seeing the AI opportunity[57:26] Final thoughts on AI, bitcoin mining, power, data centers; why this all domestic activity focusedToday's sponsor: DaloopaEarnings season is hectic—there’s no way around it. But what if you could take back the time you spend on manual model updates? With Daloopa, you can.Daloopa automates your audit and update process, instantly pulling accurate, fundamental data from filings and reports directly into your models. That means no more wasting hours on repetitive tasks. Instead, you can focus on analyzing trends, refining strategies, and staying ahead of the competition.Stop letting manual work slow you down. Set up a free account today by visitingdaloopa.com/YAV and see how Daloopa can transform your workflow.

This podcast is sponsored by Delupa. Earning season is hectic. There's no way around it. But what if you could take back the time you spend on manual model updates? With Delupa, you can. Delupa automate your audit and update process, instantly pulling accurate, fundamental data from filings and reports directly into your models. That means no more wasting hours on repetitive tasks.

Instead, you can focus on analyzing trends, refining strategies, and staying ahead of the competition. Stop letting manual work slow you down. Set up a free account today by visiting Delupa.com. slash yavie and see how and see how dilupa can transform your workflow that's the lupa d a l oopa dot com slash y a v all right hello and welcome to yet another value podcast i'm your host angie walker if you like this podcast i mean a lot of you could rate subscribe review wherever you're watching or listening to it it really helps the podcast out with me today i'm happy to have on from semi-analysis jeremy jeremy how's it going hey good good man how is it

How's going here? Look, super excited to have you on today. I guess before we get started, a quick disclaimer, nothing on this podcast is investing advice. Please consult the financial advisor, do your own research. That's always true. But we were talking ahead of time, long list of companies we can talk about today.

So, you know, we might end up hitting 50 different companies today. So please remember, we're not recommending any of them, consults financial advisor, all that. Jeremy, super excited to have you on today. Come highly recommended. Obviously, you're over at semi-analysis. We started planning this podcast before the DeepSeek stuff even came out. And, you know, this has been AI, semi, data centers, all that has been the hottest place in the market to talk about debate for the past year.

Probably got even hotter and more interesting over the past two weeks post Deep Seek. So lots to talk up. I guess I'm going to start with this. The main, my main interests in AI and everything, aside from how I can use it as an investor to improve and everything, my main interest is in the data centers and electrification. and I know that's an area you specialize in. So we can go into different companies, different industries and everything. But I guess just broadly, if I said right now, hey, most of the investors I talk to are still on a thesis of AI equals unlimited demand for data centers by your nuclear energy, by anything you can because there's going to be unlimited demand.

Like, what would you think about that trend or how are you thinking about it kind of evolving as we sit here? We're taping like February Bith, 2025. off. Yeah, I mean, Unlimited is a pretty strong world, right? But like I'd say at some analysis, we do have a pretty comprehensive tracker of, you know, what's called supply and demand dynamics. Demand being basically, you know, yeah, it is every time someone buys a GPU or even a CPU, whatever IT equipment, you want to put it into a data center, and there's a necessity demand

for power and for data center space, right? And so that's something we track very closely. So since 2024 and going forward is very much all driven by, you know, Gen AI. So Nvidia hardware and obviously also all the custom ASICs, AMD a little bit, but pretty much only AI accelerators, that's driving the whole market, right? And just to give some perspective, you know, the global data center industry worldwide, let's say from 2019 to 2023 has grown at about 4 gigawatts per year, right? That's the IT capacity added per year.

in 2024 Nvidia alone adds 5 gigawatts of demand, right? AI is roughly 7 to 8 gigawatts including all the custom basics and all that. So it's just a complete change of trend.

It's all driven by Gen. AI. And so that's the demand side of things. And then supply, which we spend a lot of time tracking is data centers. So basically the question is are people building data centers fast enough to sort of power AI and have sort of space and power

where they can place their GPUs. And the answer that we forecast for 2024, 2025, and the few years beyond is no, they need to build data centers faster, right? Which means upside for the companies exposed in that supply chain. So let me start with a basic question. And I think DeepC kind of highlighted this, right? I worry as an investor,

and you know, you pick your power company, Talon, Vistra, Constellation Energy, all these guys, like, have become absolute investor darlings because of what seems to be unlimited demand, especially like a Talon with a great nuke and everything. I worry that right now, as you and I are sitting, there's an AI data rate, an AI race, and they don't have to optimize for anything because they're just trying to go, go, go, get the best model. But I worry, and Deep Seek makes me worry about this more, if we're running this out 12 months

from now, right? at some point you have to start you start optimizing right the gains get smaller and you start optimizing and it seems to be a natural first place to optimize would be hey you know we're literally consuming like manhattan level uh like the entire city of manhattan we're consuming that much power in one data center like how hard is it to start designing hey we we reduce the power a little bit and you could get into a really overbuilt scenario there right like we've seen it before with biver Hey, data demand went straight up for 20 years. I mean, I think it was cagering at like 30% from 2000 to 2020.

But if you built fiber in 2020, like there was a huge oversupply. And I worry we built all these data centers to say nothing of all this new energy we're bringing online. And 12 months from now, we start optimizing and say, oh, we way overbuilt. Yeah, that's totally fair points. I would just start by highlighting that the sort of progress on the efficiency of inferencing AI models is tremendous, right? And everybody is freaking out about deep seek, but in reality, DeepSik is not

changing the trend. If anything, deepsick is actually below trend, right? The curve is not shifting. So what's the curve? The curve is, you know, in two years, a model of the quality of, say, GPT-free, the cost to inference such a model has fallen by about 1,200, right?

So, you know, it's the amount of improvement in the AI space is gigantic, right? And something we also like to highlight is everyone is freaking out about deep seek. But actually, if you look at other sort of competing models, Gemini 2.0 Flash, you know, is way cheaper to inference than deep seek. And he's actually a better model, right? So people are freaking out about deep seek, but it's actually, you know, it's nothing sort of out of the ordinary. Like they're very good, to be clear, but they're not better fundamentally than the best models of Google and our open AI and so forth.

right so that was always the case this trend of massive improvements obviously the so you know people talk about the javens paradox whatever the the trend matters here is sort of the overall scaling loss which is that you want to have the best the best model right like and you've seen tremendous improvement in the capability of those models and you know maybe you've tried out deep research recently it's getting pretty insane right the stuff it can do And you've seen surely the benchmarks of like 03, you know, like the archa AGI1 benchmarks pretty insane as well. Anyway, the point here is that we're going to keep building better and better models, which one means more demand for training. But typically a way, way more demand than for training.

I mean, we can talk about that in more detail. So that's one aspect. And the other aspect is way more demand for inference, right? And like, you know, just a funny thing to note is. deep seek, right, despite all their efficiency, they actually have zero capacity to serve inference, right? There's like so constraint, right? So like two things funny to notice. One is on the training side, the deep six CEO met like the number two of the CCP. The next day, China announced $140 billion of state subsidies, right? I mean, it's pretty clear that deep sick told the guy,

hey, I, you know, I need subsidies to build more compute, right? That's for training. And then for inference, they can't serve new users, right? Like they have to turn. down user requests. They're, you know, in a configuration with like max batch size and like very low interactivity. Like using Dipsic, the user interface is just not good because they have to maximize the capacity of their GPUs because they can serve users, right? So the trend is, you know, even if you're building the most efficient model, like people think Dipsic is doing, they still want way, way more computers, way more data centers, you know, way more capacity overall. Yeah, no, that all makes sense. I guess I

I mean, we could just say Jarvin's paradox, you know, every response to everything we say. I guess what I worry about is, well, let me ask you this. You, I mean, I'm just a journalist, right? Every time I email you or something like you say, hey, what do you think of this? Like, you know, the deep seek news. Everybody freaks out about it. It was like two weeks ago. Every generalist freaks out two weeks ago.

And every person, I'm sure, when did you first hear about like the deep seek and everything? Yeah, I mean, deep seek actually starting popping out about. a month ago. Yeah. So every, every specialist, I would say, like, knows the news two months before it filters through the general. So, like, if I were to talk to you and be like, hey, are you at all worried about scaling laws at this point? Or like, what is on your mind as a worry for the whole AI trade at this point? Yeah, I just think, so like the way we think about that, we kind of split it in two phases. One is training and the other one is inference, right? Yep. So if you think about trading alone. We still see that over the, let's say, next two years, all the big AI

labs have massive plants. I think a nice way to visualize that is, you know, today the largest clusters, GPU clusters are about 100,000 hopper GPUs, which in terms of power is roughly 130 megawatts of IT power. And they all want to scale that massively, and they have plans for gigawatt scale data centers, even two gigawatt scale for a few specific sites, you know, towards sort of, let's say, 2027, right? So in two years, maybe two and a half years, they want to scale their biggest cluster single site from 105,000 GPUs, 130 megawatts to, you know, the gigawatts scale. So there's massive scaling on, that's training alone, right?

However, you know, we're talking about tens of billions of dollars in CAPEX, which, you know, they partially have the funds to finance at this point. but what's going to happen as well is at some point they're going to have to face some adoption of the underlying technology, some inference route use, yeah. Can I pause you right there and ask a really stupid question?

Really stupid generalist question. Training, you mentioned, we're scaling from people, right now if it's a thousand unit clusters, they're going to 100,000 unit clusters or whatever it is, right? I've always wondered, like, you also hear, hey, we're kind of at the end

of the amount of data that we have to give these models to start training on, right? So if I told you that, like, why do we need to go from 10,000 to a million or 100, or whatever it is when, like, yes, there's always more data like the world generates more data. Yeah, we can go get some books from the 1,500s, but like most of the data has already been input into it. Like, most of the internet is in these training models. Why do we need to go from 10,000 to 100,000 when it seems like the amount of more data we can put into these models is actually, like, getting kind of smaller? Does that question make sense? Yeah, and the answer is exactly because there's not enough real data.

So we actually have to use a ton of compute to generate synthetic data, right? So when I go from 10,000 to 100,000, like the answer is, hey, we're generating these synthetic data and using that to train it up. It's the extra 90,000 or whatever is just for extra synthetic data. Yes and no. Synthetic data is a good portion. techniques to sort of add more reasoning capabilities are also in other areas which is like reinforcement learning

which we choose a lot of synthetic data and another one it's pretty interesting but like the pre-training scaling laws which is you know building a model with more parameters and more data at some one of the issues with the pre-training scaling laws is

at some point the models are too expensive to inference but you can actually use those huge models to fine tune the smaller models and make the smaller models much better, right? Which is a technique that all the AI labs have used. That's how they came up with like GPT40 or Cloud 3.5 sonnets by using sort of those techniques. So yeah. Super stupid generalist question, but this is why I have experts on so I can ask super super super generous questions. And like, look, I read about this stuff, but you read about it all day

and I read about a lot of other stuff. So I won't feel bad being done. If you start having models that are running more and more on synthetic data, right, instead of real world data. Like, don't you have a risk where the synthetic data just starts, and I understand they have checks and everything, but you can't check everything in a model. And don't you have a risk that somehow you've got like, if it's 90% synthetic data and 10% real data, all of a sudden you've got models that believe the sky is red or the, you know, the ocean is green or something. Like, why isn't that a risk?

And those would be obvious things to correct. But we're all familiar with what is the phantom answer where I ask Google like, hey, what was Apple's big product launch in 1942? And it says, oh, Apple's big product launch. They were really excited to launch the catapult that year. And you're like, where did that come from? And no one can point to it. So don't you have an increasing risk of like hidden, hidden phantom answers if it's all

running off synthetic data? Hey, Andrew, the simple answer is kidding loss again. You just need to spend more compute to have better checks, right? To have the synthetic data, I'll be of, you know, simply a better quality. Okay, let me answer another question on data. So most of the data right now is stuff that that is online, right? Do you have a rush for, I think about Tesla's having the cameras all around it. And them always saying, hey, our key self-driving is we've got at this point 10 years of every car is quit with this.

We've got 10 years of data. And obviously Waymo is going to get that very quickly and stuff. Do you have a rush to generate more data tokens where you're just like, hey, I'm Facebook. I'm going to, you know, Google's got the cars driving around everywhere. F that. I'm going to put a camera on every single corner and in every single building that will let me just so I have video of everything that's happening in the world. And that is one, video, audio, all that.

And that is one way to generate, like, enormous amounts of extra data, right? Would that be a thing? I mean, you could argue that using platforms like chat GPT or cloud or whatever is, when you do that, you kind of contribute to, you know, providing more data to those companies, right? Because every time you have an interaction, you start on their servers. So, yeah, having, like, that's one way to think about that, is having a platform with many, many users is actually very valuable because it does generate much important.

No, I get that. Like, one of the arguments for Facebook is they've got all this private data with the interactions, the DMs, what you like on Instagram. They've got more data on you than anyone. That's great for their advertising. But I guess it was more talking about bringing it into the real world, and you have, like, I could imagine a world where, Chat GPT says we really need a lot of like real time physical data.

So we're going to go out and go to every building owner we can say and say, hey, we'll pay you $1,000 per month if you will let us record everything and get that audio and video. And then we can use that as data, right, in some way, shape, or form. So there's one step before that, which is actually starting to use all the video data that's out there. Yeah, which we don't really currently, right? like the YouTube data and stuff yeah that's right all the movies all the YouTube we actually don't use

the data we use like transcripts and stuff like that but we don't actually use the video data itself which you know even like just using all of YouTube would be orders of magnitude more data than what people use using internet data right like common crawl stuff like that so that would be the first step right

first we have to use the existing video data which is you know massively I mean you think about those YouTube stores and one thing I love about YouTube is you and I have such different YouTube and you'd never even realize it, but I mean, there's, it's what, like, I don't even know, there's millions

and millions of hours of video uploaded to YouTube every day, and as you're saying, it's not recorded, though, I'd also argue, hey, if you started analyzing like a thousand Mr. Beast videos for data, yes, that's interesting, but it's very different than like me sitting in my closet shoebox interview at you or like me, you know, penning my dog

on the street or something. It's a very skewed view of the world. Well, yeah, I mean, imagine all those great podcasts that YouTube could train on, that the model could train on, right? Like, it would learn a lot of stuff, right? Definitely on, like, specifically on investing or whatever, right? But there's just, like, there's, you know,

just a ton of data out there on the video, again, orders on magnitude more than what existing, you know, on the text format. So, you know, before your scenario of paying people to generate data, there's a multi-years of video data that we are yet to uncover. I want to ask about some more data center questions, but I guess let me just completely switch years and ask your question. There was a very popular, Matt Levine called it the most market moving

short report ever, the short report on Nvidia that came out like the weekend of Deep Seek, and people think it's partly the reason Nvidia stopped, opened down, what was it, like 10 or 15% or whatever, lost 600 billion of market cap. There was a very popular thing. And the basics of it was, look, Nvidia has had great success, but success attracts competitors. And for X, Y, and Z reasons, NVIDIA is, their future is not as bright as the stock market's giving it. And like one reason was, hey, all these tech giant hyperscalers that they sell onto, Amazon, Apple, Google, all of them are trying to build NVIDIA competitors. We're not sure if, as we go from training to inference, we're not sure if NVIDIA has that

big of an advantage over their competitors. And we can go, you can list X, Y, Z reason. I'd love to just ask you because it's buzzy. Everyone wonders, what are your thoughts on kind of Nvidia at? we sit here. Okay. Hey, so let me just say that, to answer that question, you know, I'm in a, I want to say privileged position because I'm part of the semi-analysis team where, you know, we're 23 people,

a lot of sort of technical backgrounds, very different backgrounds as well. And we have like, you know, an AI engineer on board who knows everything about model architectures, we have networking experts, and so on and so forth, right? So we actually answered those questions with sort of a very technological, you know, a very inform point of view, right? And basically what we see, sort of analyzing the engineering of all those systems, and comparing to what the market,

the common market narratives, I would argue that Nvidia's engineering talent tends to be underappreciated, right? So people know that Nvidia is a great hardware designer, right? That's not a secret for anyone. People, I think, increasingly understand that they have great software modes. although there's this discussion of whether the software mode applies to inference or not.

There's the third one, which I think is less often discussed, is the networking aspect, right? And so that's a position that semi-nesses has taken, which is that we think Nvidia's mode might actually be stronger on inference compared to training, because even if the software aspect, we agree with the sort of market position, the software aspect is less important than it is for training. What is more important is everything related to networking, especially the scale-up network, right? And you have to think about the engineering prowess that is building the Nvidia, GV-200, NVL-72, you know, where you have those 72 GPUs interconnected in an all-to-all configurations with NVLink, you know, an incredible bandwidth, which is unmatched by any hyperskator. So incredible bandwidth, and, you know, you have like, I don't know, like three or four thousand,

copper cables if you're at the rack scale solution like it's pretty insane what they're doing and this is state of the art right no one else has a as good of a solution and that's a huge modes for inference right so which means that as we go into more and more reasoning models especially as reasoning models have sort of a longer context window they need more memory to scale memory you need that scale up network and that scale up network is one of avidia's provinces by Delupa. Earning season is hectic. There's no way around it.

But what if you could take back the time you spend on manual model updates? With Delupa, you can. Delupa automate your audit and update process, instantly pulling accurate, fundamental data from filings, and reports directly into your models. That means no more wasting hours on repetitive tasks. Instead, you can focus on analyzing trends, refining strategies, and staying ahead of the competition. Stop letting manual work slow you down. Set up a free account today by visiting Delupa.com slash YAV and see how

and see how the lupa can transform your workflow that's the lupa d a l oopa dot com slash y a v let me just pause on it because in video you know it is i have no view on the stock okay so i will be honest there it seems like great they're growing rapidly but you do run into like hey it is priced pretty richly and semiconductors overall are very cyclical and i guess one of the things that the bear case was pointed is like hey nobody's doubting Nvidia is a great company. It's just like it is priced so richly for what ultimately every semi every semiconductor has a cycle at some point, right? They will have a cycle at some point. And I guess the second thing is A, the cycle, but B, look, Microsoft, Amazon, Apple, these guys are

not stupid. Invita is getting huge margins. And why couldn't they recreate the Nvidia, you know, like 90% of the performance? Why do they need 100% of state of the art performance all the time for all the AI models? Why can't it be 90%? Or, hey, if I gave Microsoft engineers, you know, the Nvidia stuff from a year ago, like, couldn't they recreate it within six months? I don't know. Like, why isn't Microsoft creating their own stuff 18 months behind, 18 months behind Nvidia for 50% of their capbacks? And then the other 50% goes to Nvidia say that, like, why isn't that a reasonable thing? So I threw a lot out there. I'm talking hypothetical worlds and weird stuff. But I guess the underlying crux is,

NVSC, nobody's done in the moat, but it just, it's quite, price, white. And you say that, I'm like, why can't there be worlds where it's yes, but? Okay. Okay. I mean, it's always yes, but I'd start with. That's why investing's hard. And like, I do think, I think like two years ago, my friend, Byr and Hobart from the diff said, the way you'll know AI is here is when Nvidia stock just starts going up and all your best friends who, all your smartest friends who are into AI or buying calls. And I think of you guys all the time because I think it was Doug who was like, yeah, unlimited power demand, just buy them, buy and buy them.

But at some point, where does it end? You know, like, where's the limit? Yeah, I mean, first of all, I mean, I'll just say, like, you know, I used to be a by side guy. I mean, you know, you can invest however you want, but I'd say generally you want to be with a cycle, right? And if it's an up cycle, stuff's going to go up and if it's a done cycle, it's going to go down.

Why? Words have never been spoken on this podcast. It's kind of as simple as that. And generally, even if it's a strong up cycle, even if it's richly valued, right, nobody cares. It's just goes up because, you know, the earnings keep getting revised up, right? And so, you know, that's why we have a generally a positive stance on Nvidia and typically a lot of other AI-related stocks, including power stuff. It's just because, you know, who cares about valuation?

Because actually the earnings are going to rise so much, right? And much more than, you know, what consensus expects, right? And I would briefly talk about, like, hyper-scaler Catex, right? Like, you've seen Google's Cappex, you've seen Metas Cappex. Microsoft, they haven't given a full-year guidance. But I think at the end of Canada, our 2025, if we look back at, you know, what people project right now and what they actually did, I think it's the same answer.

Cappex is going up massively. It keeps beating estimates, you know, not by 5%, by 20%, right, or 30%. And, you know, that's how we see the trend playing. out in this up cycle is a very, very strong upwards revisions, right, in what people are spending. So, you know, that's the first thing. Up cycle, the up cycle is strong. There will be a down cycle, you know. I was like, nobody's got a crisp wall. But it was like, when does the down cycle start? Because, you know, I do think there were some people who thought, like, Project Stargate comes out. That's the 500 billion Masa, Oracle, Trump. And I remember a lot of people

were like, oh, it is on. Like, everything's going straight, straight up, demand, electric, like, buy anything. This is your license by anything because it was just a huge check and a silly number. And then Deep Sea came out and a lot of people got worried, like, it's, you know, it's never been moreover. I guess my two questions for you, let's start with the first one. When Deep Sea came out, a lot of people were saying it's over. And one thing I noted was like, Deep Sea came out maybe to the public 24 or 48 hours

after the Satya, hey, I'm good for my $80 billion check and Mark Zuckerberg kind of talking about 50 billion plus or whatever in CapEx. I would imagine they knew DeepSeek and they were committing to this amount of CapEx before that because they continued to see like upside. Would you agree or disagree with that? I fully agree because, you know, as I just said, models like Gemini for 2.0 Flash are actually better and cheaper than Deepseek. So for them it's not a change of trend.

It might be a margin impact because, you know, they want to charge their API prices for the state of the art models, like, very, very high at high margins. But, you know, overall, it's not a change of trend, right? You said when we were talking just a few seconds ago, you said, we're in an up cycle. Numbers going up. You can buy them all while the numbers are going up, right? Very pod cycle. I love the mentality. But what breaks that?

Like, what would break that? Yeah, I know. I mean, great, great question, right? I think a couple of things could break that. First of all, I think it's very important to carefully track the adoption of AI, right? You can do see this stuff like looking at website visits of chat GPT and Gemini and, you know, cloud, whatever. You can also use whatever tool you can to sort of track open AI revenue, you know, and maybe clause revenue. I think the overall adoption obviously matters.

I also think the reason this has taken such proportions is because we've actually seen somehow the ROI, right? Because the way I think about it is hyper-scaters, they build their businesses on users first, right? Not revenue first. You've had acquisitions like, you know, I think in 2007 YouTube or maybe 2008, whatever. Acquisitions like YouTube where people were like, hey, this is way overpriced. YouTube has no revenue. But in Google's mind, it's all about this is a great platform. They have great users.

I want to have the users, and I will build a revenue model after I have the user. I'll figure it out, right? WhatsApp, same story. You could argue it's maybe less successful, but Instagram is like extremely successful. Obviously, anyway, that's the idea. Hyper-scaters are user first and revenue second, right? It turned out to work pretty well for them, right? did it. So now, if you apply that logic to Gen. AI, it doesn't perfectly because

Gen.I is sort of more capital-intensive, but in terms of users, what Chad GPT demonstrated is, you know, 200 million users in, what was that, like, two months, something like that. So I think it's pretty clear that GenAI has the capability to be sort of the next billion user plus platform. That's why everybody's in such a rush, because a billion user plus platform, means tens of billions of dollars in potential revenue, maybe even hundreds of billions of dollars. So, yeah, that's one thing, is the consumer adoption is extremely strong for these tools.

So I think that's just having the users, the hundreds of millions of users, was enough to get to the level of spending that we're at right now. However, what we plan, what we forecast is going to happen in the next two years. And to be clear, there's meaningful physical evidence of that, especially on the data center stuff.

Again, they are building huge gigawatt-scale data centers, many scale data centers for AI only. At some point, to justify those investments, you're going to need more than just the users. You're going to need, obviously, revenue. So again, I think tracking whatever indicator you can to understand the adoption of these tools is crucial. And, you know, if people stop using chat GPT

or if actually they struggle to sell those, $20 a month or $200 a month subscriptions. I think, you know, at some point that's just going to be like too much for the market, right? You know, I do, you just mentioned this. At the end, it is funny. So I just paid for my first chat GPT subscription because I was, I was like, A, I'm starting to use it more. And B, I just need to commit. Like, I think if you are a professional in just about any intellectual field, which is investing in

intellectual field, sometimes I say, yes, sometimes. I say no, I don't know, but I think if you're not using it constantly, then you're probably falling behind pretty rapidly and thinking of new ways. But I do wonder, like, is subscription the way that you ultimately pay for these things just because Bloomberg, yes, but that is more like networking tool. Most of these tools, Google search, most of these tools you end up monetizing on advertisements, like I'd be kind of surprised if subscriptions is the way they all ultimately monetize, especially if, as we were talking about earlier, data is the

way that you get better at these, like one way to generate proprietary data is have more searches, more uses going through you, more API requests, and how do you get the most searches? Make your product free and then make your product free and monetize it over ads. I don't know if you want to say anything there. I've got some more questions together. Yeah, I would just say that I think the most, the business models that people are exploring right now is basically having a free model with decent. capabilities, which is, you know, GPD 4-0 Mini, which is, you know, Gemini 2.0 Flash,

and those sort of free models, they're going to see their capability increase, but obviously you're going to have just that free version as a sort of advertisement tool, as a massive user tool, right? Again, hundreds of millions, not billions of users. And then on top of that, like you see the capability, and then you can sell all kinds of extra stuff. Let's see how. how they figure it up advertising could definitely be part of it as someone who's got your finger more on the pulse than this than probably 99.9% of the population my listener based everything which what a i tool do you use the most i like jemini you like jemini okay let me switch okay have you have you tried deep research it's actually it's actually pretty

incredible like if you want to do a uh quick stock initiation report uh it's actually fairly decent there are like some pretty good prompts I haven't but I will tell you one of the reasons I paid for chat GPT is over the past couple weeks I was researching stuff and I started playing around with the prompts I put in in chat GPT and using the reasoning on it and some of the outputs

I got were like it would have taken me a month to put it together and it was really incredible so I was like all right this is just I've got to use it as an investor this is me personally obviously there are people who are making their fortunes in semis or, you know, buy Microsoft meta and just go to a beach or something.

For me personally, the two areas that I think are most interesting are power. And you could divert that into, you know, the talent, the actual power producers. You could divert that into, hey, are we going to have, as this demand goes up, we have to somehow nuclear, nat gas, even coal, all of those are interesting on one side, the power. And other places, the data centers. So, like, this rush for data centers, and we're going to come. to the Bitcoin miners in particular, what in my mind? But let's start power. As we talk about it, we talked earlier about this demand for power, all this getting built. I guess when you think about

that power, like, what do you think are kind of, I know you guys focus on semis, but you guys also know the power supply, the power demand. What do you think are the most interesting kind of power plays? Is it what everybody likes to do and just go buy the power producers? Or do you think, hey, we've got this, it's such a crutch, we need so much more power, like actually the underlying commodities, net gas, we need more of this coal retirements, forget it, we can't retire these coal plants. Like, how do you guys kind of think about playing that? Yeah, I mean, something that I personally look at a lot is the data center infrastructure

stuff, right? Which is, I think it's a pretty good proxy for, again, like GPU demand, especially because it's a Cappex-driven type of business model. So everything relates to the electrical equipment, calling equipment. There are a lot of companies in there. There's a few big ones like Schneider, Eton, and Vertive. Vertive is what I was going to say, yep. If you want to be creative, you can look at the Taiwan stock market, right?

There are some liquid cooling stocks, pretty high exposure to liquid cooling, so interesting plays over there. And there are some smaller, you know, small to midcaps exposed in maybe MEP engineering for data centers, maybe cooling towers for data centers. There's a whole pretty big supply chain. And the thing is, you know, those pieces of equipment, they're overall commodities, right? They don't have the level of product differentiation that Nvidia has with its GPUs, for example. So even if, like, Schneider dominates the electrical and cooling market and virtue even eaten,

there's actually room for smaller folks to play in sort of more local ecosystems. So, you know, everyone can sort of benefit in that area. I mean, that's one, you know, because I personally look a lot at equipment, but you're right by saying also the power producers and overall net gas. Look, that's perfect. Let me ask the one I've thought about the most. I think from my small cat value special situation friends, I mean, yes, all of them are trying to find weird power plays and everything. But I think the one that it's most popular because the stocks are so volatile and they're so interesting is the Bitcoin miners. And, you know, basically for those who don't know,

the Bitcoin miners, mining Bitcoin is similar to AI and that you take some GPUs or Bitcoin miners, you throw a lot of power through them and you get Bitcoin. And that is a very commodity, very difficult business. But a lot of these guys woke up and they said, hey, you've got all these AI guys who are just desperate for places with cooling, with tons of power, huge places with GPUs. We're Bitcoin miners. We strip those Bitcoin out. We put AI in, GPUs in. And boom, We've got an AI data center and, you know, those things are getting valued at huge multiples. And I'll disclose, Core Scientific, actually, they did this, right? Last summer and people can go look it up.

They strike a deal with CoreWeave. The stock is a 4x in a year because they go from, hey, we're a crappy Bitcoin miner to, hey, we are AI data infrastructure. So I want to ask you, like, that was Core Scientific. We can talk about their assets you want, but there are tons of Bitcoin miners out there. And I know you know several of them well and several of them may be less well, but I'd love to talk. I can give you specific prompts or we can go through specific ones that you're familiar with. I'd love to just talk about the Bitcoin miners as potential AI plays. I would just like to provide a little bit of context before that.

And also add another compliment to the previous question. I think an area that you want to look into is sort of time to power solutions, right? Yes, yes, yes. Because we said in the introduction that so in our forecast, like we have an extremely granular forecast. of the supply, demand dynamics of the global AI data center market, we forecast a deficit that is going to increase in the coming years, despite the data center boom, which means, you know, if you want to solve the gap, you'll need time to power solutions, right?

For a while, people thought the IPPs, you know, behind the meteor nuclear with, like, the talent-type deals would be the solution. It's actually way trickier than expected. So you have to look at other solutions. And what I would add is this sort of supply demand mismatch is creating sort of value which without that mismatch wouldn't be economically doable, right? So that obviously impacts the miner because, hey, they're crappy miners, right? They don't have the data center expertise. They might have a few folks from like digital realty and equinics, but they don't have the full experience of building those solutions.

So in a normal environment, you know, people would never go to those miners, especially for, you know, 500 megawatts scale, right? Because it's a huge amount of capital, right? We're talking about a $5 billion investment. But in this environment of supply demand mismatch, the miners, which have historically been power first, right? Because it's all about finding sort of stranded power available in large quantities, extremely cheap, which is why they go to West Texas, they go to, you know, Wyoming, they go to North Dakota. And so, yeah, the miners definitely, you know, have the power. They can provide time to power solutions. So suddenly, you know, they're actually a very valuable asset to hypers

and other folks deploying GPUs, right? Yeah, no, I've said, they're clown cars that fell into a gold mine, right? Like, they went and they built out these huge things because they all thought Bitcoin was going to a million, and they did it almost without regard for costs or economics or competitive analysis. And then they built them like, oh, no, we're having. economics have come down and then all of a sudden there's people who are saying hey you've got 300 megs of power like if we want to go build that ourselves we're not looking at it getting until 2,029.

we could buy your facility throw the bitcoin miners away and be online in six months. F yeah let's go yeah and hey like if you if you go back you know roughly a year or maybe a bit less remember that core we've wanted to buy a core scientific right because they're all like you know i buy this guy I get sort of a gigawatt of available power and then I develop it myself. It's like so cheap. They offered to, hey, wait, man, actually. They were- They offered to buy them and then if you did the math on the contract they struck,

the contract they struck was worth 2x what they offered to buy core scientific for. So I think there was a deal there and I always looked at it and I was like, I don't understand what I'm missing. Like they offered to buy core scientific for, I'm making up numbers, a market cap of a billion and then like we'll pay you a billion or we'll sign this long-term deal that's worth you know 1.8 billion in NPV what and you keep like bitcoin miner upside and everything on your other assets which would you prefer and it was like if you had just offered 1.8 like you might have had a deal it was very strange that's exactly what happened pretty much

but so yeah the miners there are a few others like one we've looked into our fuel cells for example which we can talk about later if you want but yeah what was that fuel sales yeah that Yeah, that's another interesting one, right? Because it's typically these kind of solutions that in the old world would be, you know, non-economical. I mean, typically, there are some deployments. The EET, like, especially if you think about the leader, which is Bloom Energy, it is a real business. It's not a profitable business. You know, it's a business that tends to, they have overpressed a lot, right?

A lot. But today, actually, suddenly their solution makes sense, kind of for the same reasons than crypto miners, which is that, hey, if you can provide time to power, selling electricity at, you know, maybe 15 cents instead of the typical 8 cents or 7 cents, whatever, it actually makes sense now if you can, you know, shorten the timeline by one year or two years for access to power, right? So that's just another idea, which I mean, we can talk about later, but, hey, if you want to focus on the miners, let's do it. No, I'd actually love to start with the miners. Like, look, we, I've looked at most of them, some of them more depth than others. I'm sure

you are the same way. You tell me one, a miner that you are kind of most bullish on as an AI play. If you were like as a Bitcoin miner play, we're like, Bitcoin mine, it's not a very good business, whatever. But as an AI play, one that you think will successfully kind of make the leap, I'd love to just discuss them. Yeah, yeah. So like, I would argue it's all about how much power do you have available. And another aspect is also, you know, how much sort of uncovered upside. Because the easy answer would be, hey, core, scientific, but, you know, it's a... Yeah, I was going to say, put them aside because they've made the leap. And we can argue if they're undervalue or not, but they're an AI play at this point, right?

So put them aside. But I don't believe any of the other ones have real AI contracts. Some have starter contracts. Some have said, we've got an L.I, but I don't believe a single other Bitcoin miner has a firm, like 100 meg AI play. So if you choose any other one, you are kind of making somewhat of a bet that they can make that leap. I'd say the you could argue the craziest one but I think makes sense is iron

right just because they have you know 1.4 gigawat secured in West Texas now which remains to be seen exactly how they manage that but just by itself having that kind of power valuable right now the sort of downside that I see

and I think that's a trend that also value to other folks is I feel like you know just looking at earnings call management team is kind of hesitant into like how much they want to go into AI versus Bitcoin mining. It feels to me that they're still like very dedicated in developing their Bitcoin mining business.

I think that's the case for other folks like a riot or marathon. They're like very much into Bitcoin mining. In my personal opinion, they should go all in towards AI because, you know, it's valuations of a collocation business is way higher, right? You're talking about 15 to 20 times EBITDA for folks like digital realities beyond 20 times a beta. We're talking about, you know, 15 years of revenue visibility, the same amount of revenue, but again, like value at 20 times a bit. I'm going to earn my bones as a podcast host right here because I'm going to prove I've done

the work over months. Iron, you mentioned them. There was a corporate short report that I thought was at least very good on the economics of Bitcoin mining from, I'm looking at my notes now, from July, July 2024. And I think one of the issues with Iran is, if I remember correctly, you said 1.4 gigawatts, which is a huge amount of power. You know, the Amazon talent deal, if I remember correctly, it was for several hundred megs of power. So you're talking, you could build a Manhattan-sized AI center here. If that 1.4 gigs could get converted, it would be great.

But I think one of the issues is, and you run into this a lot with the West Texas miners, all of them in order to be allowed to build, had to sign curtailment agreements with the government. So, you know, if the retail load is too high, then they have to shut down and kind of like support the grid that way. And then B, I think a lot of the power is intermittent and, you know, it's coming from wind and solar. And I think when you roll all those together, and this is a question I have with a lot of the Bitcoin miners who say they're going to become AI plays. I think a lot of the power is intermittent. And that is you can't use intermittent power, whether you're getting curtailed, solar, wind, whatever it is.

You can't use that for AI because the AI, it needs to be on 100% of the time. And if you kind of think about it, hey, you're going to spend $100 million a year on power, whatever it is, you're going to put $5 billion of GPUs in there. If you're offline for five minutes, the GPU cost overwhelms your power costs completely, right? So I threw all that, but I guess for Iran, do you think they actually can get an AI given that power issue I talked about? Or do you think they've got certain assets that can kind of overcome it? I think generally the issue is not that hard to overcome. I mean, you probably know that most data centers in the world have backup generators, you know, which are meant to cover episodes where you don't have grid power.

So if there's a grid power failure, you know, you have your backup generators to maintain power. So that's one thing. You can build a bit more sophisticated systems. Some people have started deploying, for example, on-site battery systems to also manage those peaks. And other things also to keep in mind is when you build a data center, so it's not just about the diesel generators. There's also batteries inside data centers. In all data centers, it's part of the UPS system.

So you could increase, for example, the amount of storage you have inside a data center. Typically, people have like five to ten minutes of batteries, just enough time for the generators to turn on. But you could imagine going maybe a bit further if it makes sense. You know, it might or might not make sense. If it does, you could go beyond 20 minutes, 30 minutes, whatever. Again, it's just if it makes sense. But I think overall the issue can be overcome, like, you know, somewhat easily.

Let me push you back in a different way. So the core scientific core we've deal is announced, I believe it's June of 2024, which is not that long ago, right? Seven months, eight months. The Talon Amazon deal is announced, I think it was March, 2024. At this point, if you're a Bitcoin miner, you have had somewhere between six to nine months to try to make the AI play. And I mean, these AI guys, you're seeing it.

They are desperate for power. They're desperate for these places. And six to nine months in the grand scheme of things when you're making a company betting transition is not huge. But at the same time, it is a long time, especially when there's kind of a gold rush. And we still haven't seen anyone but core scientific, to my knowledge, sign a definitive contract. Now, I think IRN has experienced with buying GPUs.

doing like a AI as a service, I think Applied has talked about doing that as well. But I guess my pushback here would be, hey, if somebody, and I actually think the iron assets are, you know, on the scale of one being the worst assets out there and 10 being, I think they're probably a 6 to 6.5. And I'm not an expert. So somebody could come and be like, you're too high or too too low, whatever. But, you know, for Iran or any of these others, like Tara Wolf's another one. I had the CFO on the podcast. I think their assets are somewhere between an eight and a nine. We haven't seen them land a contract yet.

And again, I know these things take time, but why haven't we seen someone land a contract if any of these were going to land a contract? So F.W.A. just, Terilov, did sign a contract. I thought it was pretty small, though. I mean, I'm talking like signing a, you know, 100 meg plus contract, not the starter. Like, hey, we're getting 10 megs of like start training.

And you tell me if I'm misremembering what Terowulf signed. I think it's 70, so it's not that small. Oh, yeah, okay. That's bigger than I thought. Yeah. I would say the one that comes closest is applied digital, because they actually have a real data center. It's already built.

North Dakota, 600 megs, absolutely, yeah. So they have 100 megawatt data center built. It's already built, right? It's not a mining data center. I would say real data center, which they paid for, you know, like it's roughly $10 million per megawat. So they paid a billion dollars for this data center.

It's pretty much ready. They're going to need more capital to build the other data center. centers, which supposedly are covered by their non-binding LOI, but they recently got the capital. They got the backing by Macquarie. So I'd say applied digital is a credible number two behind core scientific, and I think they're likely to get their 400 megawatt deal coming through, right? So, okay, so now just taking a step back. So why those core scientific signed a deal? Why is applied digital like a decent number two? And other folks struggle, I would argued about technical capabilities within the team.

So both applied digital and core scientific already had experience in doing a color business, right? So before the flagship deal, Core Scientific actually signed a 16 megawatt deal with Corrieve in Austin, Texas. Was that 203 or early 24? I don't remember. It was early 24, but you are absolutely correct, yeah. so you know the point here and the point here is those guys have sort of seen this trend earlier than others they have hired people accordingly and they already had somewhat decent industry relationships and when I say hire they hired you know pretty senior folks at the color companies so you know they just built they just built good teams which is you know in the end what matters but like

it's a it's a human business now if you think about folks like iron, right? They just, again, they just need to hire the people. They just need more technical expertise. They just need to sort of may probably be more focused or streamline on what exactly they want to do.

Because, you know, at least in their last earnings call, they said, hey, we're not sure if you want to do Bitcoin mining or if you want to do AI data center. You know, at some point you get to choose. Can I pause you there? Because this is the other thing that makes me like smash my head against the wall. And Doug, who's

over to me now, so him and I talked about this when some of these companies were like last summer. And it makes me want to smash my head against the ball. And I'm not blaming you. I'm ranting and I love your response because I think there's an interesting question here. Any of these guys, right? They say, hey, we're not sure. Bitcoin mining or AI. We're not sure. And I'm like, look, if you're a Bitcoin miner, you're going to get valued at 200 megs, $200 meg, $100 a meg, whatever. If you're an AI, go look at the deals that people are signing. a $1,000 a meg would be low, right? And you've got this gold rush.

And then I would also add on top of that, I'd say, hey, all these guys, I've never seen share issuance like what they're doing. And, you know, the stocks trade at 200 meg, 300 to Meg, whatever. And they are diluting shareholders like crazy. Now, for many of them, it's because apply for one. They are building out this data center. But I guess I look at the two and I say they are insiders and they are not gung-ho about the switch, nor are they protecting their equity. And then I am an outsider saying, hey,

it seems like if you made the switch, you know, your stock would go from being worth 200 to 2,000 overnight. And instead, they're diluting like crazy. And I look at all that and I say, maybe I think there's a pot of gold, but the people who should know at best don't. And maybe this is because for many of them, not all of them, but for many of them, to be fair, like a lot of the Bitcoin miners came about in 2018, 2019, and they have more promotional backgrounds, to put it politely. But I look at that combo of facts and I say, Jeremy and I are out here talking about if you make that switch to AI, these stocks are growing up, 3x, 5x, 10x, not financial advice, just chatting numbers and stuff. And they're out here saying, 2004, some of these names we

mentioned, they probably increase their diluted chair count by 4x in 2024. Like, I don't think that's an exaggeration. I threw a ton at you, but I think you can see where I'm going. They don't see the gold mine. Why are we seeing it and they don't like? Are we wrong? Or we do it until... Because I'm sure they've had conversations with AI plays. Corwee, if any of these were great, Corwee would probably have bought them on the spot.

So let me throw that to you. Yeah, that's a fair point. I think there are, like, I obviously there are many things to discuss. I would start by something, you know, maybe controversial, whatever, but I'm sure you know that many folks involved in Bitcoin mining. In Bitcoin generally,

you know, I'm going to say the world, religious, they have a religious way to think about that. I don't know if that's the right world, but, you know, the laser eyes, whatever. Do their credit a few of them, and I won't say the specific names, but as recently as six months ago, they would have said, we're never going to do AI. We're Bitcoin miners to the core. And then I think that AI numbers went so high. They said, maybe we will consider some of our assets to AI.

So they are coming around to their credit. But I definitely agree with what you're saying there. Yeah. And, you know, I personally like Bitcoin. So I think they're right about several things. in their overall pitch, but you know, like if you're in their mind, you know, they see

Bitcoin going to a million, as we discussed previously. So Bitcoin mining is the business to be in, especially where in, you know, where you're in a cycle. Even if Bitcoin went from $100,000 to a million overnight, these guys would be value neutral between Bitcoin mining

and switching to AI at these current prices. So it's like, hey, if we assume Bitcoin goes up 10x, why don't you just make the switch to AI and then go buy Bitcoin in your person? account, man. That would be better for everybody. Well, yeah, I agree. I agree. But yeah, so anyway, that's one thing. The other one is, you know, people are generally

bullish on AI. That's not a secret. But, you know, there's still some skepticism. I think there's a lot of fear. Because to be clear, one thing you have to keep in mind is the capital required to build an AI data center is way, way higher than what they're used to building. So the good way to think about it is KAPEX per megawatts, which when you build a Bitcoin data center, I'm just talking about the physical infrastructure, not the hardware. Not, yes, absolutely. Those Bitcoin data centers, they typically cost $0.5 million per megawatts, right? Now, when you do an AI data center, especially when you don't have a lot of internal capabilities, so you have

to outsource a lot of stuff, we're talking about $10 million plus dollars. per middle one. So it's a 20x factor, right? So, you know, that's what you see. But you don't have to do it on spec, right? Like core we've paid for most of the catbacks for core scientific. So if these assets were so good, I definitely hear you. When you're saying, hey, you used to spend 500 and now you need to spend 10,000, that you got to raise a lot of capital. That's a huge switch. But if these assets were so attractive, all they have to do is go to, and NVIDIA wrote a check to apply, right? NVIDIA would write a check.

You could find somebody to write a check and cover that KAPX, and I just keep looking and be like, I know it's not been forever, but why haven't we seen one other proof point yet? But again, I would go back to sort of the industry relationships where I think Core Scientific and Koref have had a partnership for several years. Core Scientific actually hosted Kore with Ethereum mining a few years ago. They know each other pretty well.

And I'm sure, you know, they've had these internal discussions before doing that deal where surely, of course-scientific, you know, try to negotiate with them, how much amount of capital they can contribute, right? Like, and some folks, like, you know, Iran or probably, you know, marathon, whoever, I don't think they're really, they're really into the data center market, or maybe they're just hiring right now. They don't have the industry relationships. You know, I personally go to a lot of industry conferences around data centers.

was at PTC, for example, a few weeks ago, which is a pretty big one. I met folks like Applied Digital, Core Scientific was there. Hurlade is also one with a credible project. They were there too. But, you know, a lot of the newer guys, I didn't see them, right? They're pretty new to those events. So they have to, you know, just get familiarized with the overall industry, which is the work that some guys like Core Scientific have done way sooner, right?

you have to you know you need some credibility you need some industry relationships this podcast is sponsored by delupa earning season is hectic there's no way around it but what if you could take back the time you spend on manual model updates with delupa you can dilupa automate your audit and update process instantly pulling accurate fundamental data from filings and reports directly into your models that means no more wasting hours on repetitive tasks instead you can focus on analyzing trends refining strategies and staying ahead of the competition. Stop letting manual work slow you down. Set up a free account today by visiting dilupa.com slash yavie and see how and see how dilupa can transform your workflow. That's the lupa d a L-O-O-O-O-Pa.com

slash y-A-V. This was a super interesting conversation. Just if I was wrapping up the Bitcoin miner, it feels to me and I know apply digital. I've actually really, I've done a decent bit of work there. I think I really like the North Dakota asset, but I'm not a super expert, but I It sounds like you think, and you're hearing for people in the industry, you think that's a pretty solid asset. Yeah. Yeah, I think it's a pretty solid asset. Which, I mean, look, they've done a lot of dilution, but it's pretty crazy. When you start talking about a 600 meg asset, that could be what it would be worth if it was 600 megs applied to AI, it's a big number.

It's a big number. And it sounds like you like the IRA asset. Look, let's wrap up on Bitcoin real quick. Any other miners you want to touch on or anything? We're also coming up over an hour, so I don't want to drag this on for everybody. I mean, just very shortly on iron. It's just a more speculative one, but, you know, they just have a lot of power. Again, I'm not sure about the sort of human aspects.

They have the human capital and the hires. But anyway, I think HAD8 has an interesting project in Louisiana. Which one? On eight? Yeah, yeah, yep. They have an interesting data center. It's 100 megawatts in 2025 in Louisiana. expandable to, I think, 200 by N26, something like that, which is real stuff, right?

Like, they already have a solid site plan. They already have the power secured. So I think that's a pretty solid one as well. Let me ask you completely out of left field question. And we're coming up on time. We'll have to have you back on because this has been super interesting. I'm sure there are 100,000 questions. Generalist investors, everyone would love to ask you.

Most of the data center activity is happening domestically, I would say, in the United States. Forget Europe, but I'm always a little. surprise we're not seeing like a big day well they've got power issues I understand I understand why and I understand some of the reasons I've heard for domestically but I guess I'm a little surprised like

I could be wrong I could be missing it but I haven't seen like a meta saying hey cutter natural gas is basically free there we're going to build a you know a cajillion dollar a kid trillion megawatt plant over there and have basically free natural gas over there. I'm hitting the Middle East, but you could say the same thing.

Brazil, right? Brazil's got huge energy resources. I am surprised how much domestically that it's been focused on the AI data train. I'm sure I'm missing a few, but you tell me if I'm completely off there or why you think it's been so domestically focused. You're 100% right. And I would just say time to market the U.S. So, okay, I think an interesting way to visualize that

if you look at where are the large-scale data centers today in the world, and let's go back in time and let's say we're in 2022-23, where are the large-scale data center in the world? Most of them are self-built by hyperscaders, U.S. hypers, and self-built hyperscale is, you know, 75% to 80% USA, right? So large-scale data centers in the U.S. is nothing new, right? the U.S. has the experience in building 100 megawattscale data centers.

They have the supply chains. They have a decent amount of labor. They have decent amount of grid power. The U.S. is just actually a good place to build data centers. For now, very secure property rights. And one issue I have heard is, hey, you go build a data center, forget about the fact

you're going to spend a billion dollar on a hyperscalar data center. You're going to put $4 billion of GPUs in there. You spend a billion dollars. and then you put the $4 billion of GPUs, and then the government shows up with some machine guns and says, thanks for the $5 billion investment. We'll be taking that now. I've certainly heard that, but...

Yeah, that's a big risk. But just even without thinking about those, stuff like labor, for example, like, hey, you want to be in the Middle East, but actually getting specialized labor is not that easy, right? Like low-level labor, sure, but more specialized electrical, mechanical plumbing. Even lower-level label,

You hear those horror stories about Cutter when they hosted the World Cup about what they had to do with labor to build out all those stadiums. Yeah, exactly. So that just shows you that it's not as easy as it seems to go overseas. Again, the data center infrastructure has some specific requirements and no other country in the world knows better how to build data centers than the U.S., right? Especially large-scale data centers. So that's a simple way.

Plus, obviously, all the companies are American, right, the big AI labs, the hypers, the ecosystems is American, so it's all happening in the U.S. Perfect. Jeremy, this was so much fun. How could people find you if they want to get in touch, learn more? Yeah, I mean, I have a very underfollowed Twitter account that has almost no toasting. Very underfollowed Twitter account. I don't know, link it in if you guys want. Or you could just, I think, you know, the best way to sort of be introduced to Twitter to check out some analysis. The website has incredibly the, deep content. Other people in the team are also amazing. So yeah, I think I understand the best way

to reach out would be either through some analyses or just, you know, whatever, Twitter, link it in, however you guys want. Perfect. Well, Jeremy, this was a ton of fun. Honestly, the industry is developing so fast. We could do probably one of these a week and find a way to talk, but we'll probably avoid one a week, but we're going to have to have you on for a follow-up conversation, and this was awesome. Thanks so much for coming on and we'll chat soon. Yeah, sounds great. Thanks for having the endroop. A quick disclaimer. Nothing on this podcast should be considered an investment advice. Guests or the hosts may have positions in any of the stocks mentioned during this podcast. Please do your own work and consult a financial advisor. Thanks.