FULL TRANSCRIPT
Slava Rubin (00:00)
Alright, I hope everybody's been enjoying the World Cup. since we last spoke, there's been many games and a lot of drama. But here we are with the final four teams, the top four teams before the World Cup even started. This is gonna be our final session before the World Cup semis and finals. So I just wanted to give a shout out for all those that are celebrating World Cup. So
let's dive in. My name is Slava Rubin. I'm one of the founders here at Vincent. We're very excited to be bringing you the next in the pre-IPO series. Feel free to check out our website, our as well as our newsletter, the alternative investment report. and of course this podcast. We have the Smart Humans podcast. With me is Jan-Erik from Sacra. Jan-Erik, say hi.
Jan-Erik Asplund (00:44)
Always great to be here.
Slava Rubin (00:47)
And today we're gonna be talking about a very exciting topic, but not the most easily understood. We're gonna be talking about AI Infra, which is the artificial intelligence infrastructure build-out. We're gonna be discussing several companies that maybe you've heard of and others that you have not. And really there's a lot of energy going into this space, literally and figuratively. So we'll be covering Fluidstack and a few other companies. And today's session is brought to us by Templum.
Capital. They're our partner for today. And Templum Capital is an asset manager providing eligible investors with access to select private companies for investment opportunities. Feel free to check them out.
So, as always, we like to say who's in the room. It's pretty consistent, which is about three-quarters accredited and about three-quarters intermediate or advanced. So, you all are pretty serious in the room about what you're here to do. And over 50% of you have mentioned that you're looking to invest into pre-IPO in the next 12 months. It's been a pretty hot space. a word from our compliance department. Nothing in this presentation should be construed.
As an offer to sell securities or solicitation of an offer to buy securities. All investments involve risk and the possibility of loss, including loss of principal. Neither past performance nor forward-looking information isn't guarantee of future results. let's start with an overview first, and then we'll ask our survey question. Go for it, Jan-Erik.
Jan-Erik Asplund (02:14)
yeah, everyone has sort of a vague awareness of the fact that, you know, AI is a large consumer of energy. There's been a lot of demand for power and data centers and GPUs. but to sort of situate it, you know, where we are right now, we we sort of had this very big era of training, training models where you had OpenAI and Anthropic hooking up thousands of
GPUs feeding a you know huge amount of data into these models and then running this kind of cluster of of GPUs, which are chips for weeks, for months, to get these models started. where we are at today is more that you know, most of the demand is coming from this inference, right? Which is basically when I ask ChatGPT a question or you know, an agent
an AI agent operates and performs some task on your behalf. every single time an action like that is taken, it requires more of that compute, the same exact kind of compute that was used for training. So what was initially a sort of large batch, you know, of training compute that was needed has become a, you know, steady state and and you know growing dramatically each year amount of compute. So this is, you know, something that requires
tons of GPUs and over time, you know, more and more. So the main way these are measured is gigawatts. You'll see this everywhere. GW. for context, like GPT 3 was trained on less than one megawatt, GPT four about 25 megawatts. we, you know, we we are up at GPT five now and you know getting to six. And so we're at the hundred, you know, 500 megawatt range, right?
Gigawatts are sort of what's necessary now that we are not just training models, but delivering them on a, you know, global scale. and AI is in every product. and all the labs are competing to sort of have the biggest model, you know, have the most available inference. You know, you don't want to have downtime constantly. So we have moved from, you know, kind of obviously in 2022, very small scale relatively, to now we're talking about.
The kind of compute needed, kind of power needed is on the level of power plants, you know, industrial gigantic campuses, public infrastructure, that kind of thing. So that's we're talking about the, you know, industry of supplying all that compute, all that power to the AI labs, you know, to Microsoft, to Google, to every company that needs to either train AI models or deliver them at a high scale to a lot of people.
Slava Rubin (04:51)
You have this in the slide, I just and you mentioned it. I just don't want it to be lost, so I'm just gonna double click on this, which is one gigawatt is a thousand megawatts, which is a lot more. And GBT three train in left for less than one megawatt, right? GBT four around twenty-five megawatts, GPT five, which is pretty advanced, a hundred megawatts.
Or a little bit more. We're now talking in gigawatts for what they're committing to. And then you have here on the right even a terawatt. and is that a thousand gigawatts?
Jan-Erik Asplund (05:31)
yeah.
Slava Rubin (05:32)
I mean, these are massive, massive numbers. just want to make sure people are seeing that in regards to the potential growth rate. before I follow up with some more questions, let's do our survey here, which is how many of you are thinking to be investing into this space in the next 12 months? So just give me a quick yes, no, unsure. it's always interesting. This one's a little bit more confusing.
than the average space. So I'd love to see how many of you are looking to invest into this space.
Okay, let's see what we got.
so about a third unsure, which is a little higher, makes sense, and only eleven percent no. Interesting. let's let's keep the momentum going here. So I just want to separate something really quickly. So these, let's call it open AI, they don't have their own GPUs, is that right? So they're renting from somebody else and they're paying somebody else, is that right?
Jan-Erik Asplund (06:30)
Yeah, exactly. So they have been from the start exactly working with Microsoft being the first sort of big partner. but more recently with Oracle, which has emerged as one of the big sort of backend, you know, GPU provider, data center providers for open AI. and so yeah, they're kind of working with, you know, whoever has access to GPUs who can who can supply them.
Slava Rubin (06:53)
So when people talk about this space, they talk about the electricity needed, the power needed, and then they talk about like the GPUs, meaning the compute needed. Is it one and the same or or can they be separated?
Jan-Erik Asplund (07:08)
Yeah, they can definitely be separated. So for OpenAI, for example, you know, they like every other AI company, has been working tightly with NVIDIA. and so NVIDIA has sort of relationships all over the space. you know, famously they sort of have been very tight with Core Weave, who, you know, is public, but also Lambda and and a few of these other ones are very tight partners with NVIDIA. So, you know, it's kind of a situation where
A Coreweave or a Lambda might be sort of renting out and managing a data center full of GPUs, but those GPUs come from NVIDIA. so there's a bit of a you know it's a it's a multi-partner ecosystem. And there's also competitive dynamics you know that are interesting there, where you know Google, Microsoft are big customers, but also developing their own chips. and even OpenAI is now developing its own ships.
you know, potentially reducing its its reliance on NVIDIA in the future. but you know, obviously won't reduce their reliance on, you know, having data centers and having computes. So
Slava Rubin (08:09)
And I just want to use the synonym. So data centers, compute, and power, are we saying they're all the same?
Jan-Erik Asplund (08:16)
they're all pretty, you know, related in that you know, you're gonna need data centers to sort of to sort of hold all the GPUs and power to to fuel them. so yeah, they're all essentially I don't think we'll go too deep into the distinctions that matter there as far as like who'd supplies the power and whatnot. But yeah.
Slava Rubin (08:35)
Great, let's go to the next slide.
I think this is really valuable for people to understand in terms of I even heard the other day somebody talking about the business model of potentially potentially you know buying their own GPUs for the sake of them renting them out and almost getting like a yield asset as like a business. So just talk us about what's the GPU math here.
Jan-Erik Asplund (08:53)
Yeah, I'll I'll so I'll present kind of the you know the basic version and maybe the theoretical version, right? Which is basically that exactly you buy a GPU. let's say you spend 30k to buy a NVIDIA H100 and then you start renting it out on one of these you know, you're a Fluidstack or a lambda and you make it so any developer, any startup can come and use your GPUs.
By paying a certain amount of money per hour. that starts to pay back if it's utilized about 80% of the time by someone, you get 17, 18 grand a year in in revenue. And in theory, that pays you back before year two is up. so that is the sort of cost.
Slava Rubin (09:39)
About it costs about thirty thousand dollars to buy one, twenty five to thirty.
Jan-Erik Asplund (09:42)
Exactly. Exactly. So by year two you made thirty five K in theory.
Slava Rubin (09:47)
And for for the audience, really a GPU is like almost like a fancy computer, right?
Jan-Erik Asplund (09:53)
Yeah, a a computer chip that will perform the calculations that the AI needs.
so obviously this looks like an amazing business because after year, you know, year two, you're just pure profit, just keep making infinite cash. of course, you know, this is oversimplified. So any GPU requires maintenance. You're paying for the rent, you're paying for the electricity, the cooling, the the networking. also you're assuming 80% utilization, which is n by no means guaranteed. So often.
You won't have that. and then there's also one of the biggest questions that people had about these companies early on was is a GPU in year three, four still worth anything? Is it still valuable given both like deterioration over time and also new GPUs coming out? So if the you know the Blackwell chip comes out from Nvidia, is it no longer, you know, the old chips are are suddenly not
state of the art, you know, you're you're not going to get the same kind of performance. So there are some complications there. The sort of big, you know, way that companies like CoreWeb are sort of getting around some of the getting around some of the questions around utilization is they sell these all on multi year contracts. you're not just sort of charging per hour, right? Like a like a motel. you're pre selling your entire fleet of GPUs to
a Microsoft, an NVIDIA, an open AI that can that can, you know, that has no problem paying for you know, a billion dollars a year in compute. so that's sort of how the big ones have sort of gotten around that question and what's created the room for companies like Fluidstack to do more like charging by the hour. because if you're a startup you obviously can't, you know, pay for five years of compute up front. so yeah, I'll stop I'll stop there.
Slava Rubin (11:43)
And a couple of variables that people think about is what is it going to be worth, like you said, in a year or two or three, very similar to a car. Some cars, I mean, all cars lose their value when you take it off the lot, which is why some people are buying rentals first. but some cars will retain more value than others. and that has to deal with a lot of what's the demand out there and how well it can hold up. Very similar about the GPU, right? So it was uncertain how well these would hold up, but so far they're holding up very well. And that's because demand has been going through the roof.
Then the question is, you know, are we overbuilt? Is there a ton of demand? Is all of a sudden the market gonna drop six months from now, eighteen months from now? And we used to be able to retain full value and sell it for ninety percent of value within a year or two, or even eighty percent, does it all of a sudden become a forty percent value after a year, which totally changes the math, right?
Jan-Erik Asplund (12:30)
Yeah, exactly. And there are ways around. So this is, I think, one of the key things to if you're going to invest to understand. and you won't get it just from this presentation. But on the on the other side, you know, these GPUs can be sort of repurposed. you know, training requires like the best GPUs, and then you can repurpose them for inference down the line. and then there's also kind of the factor that like a lot of the business model is based on getting the the GPU paid off such that you can then go and
Upgrade to to newer ones, you know, as quickly as possible. so you turn over and quickly. And that's sort of the to stay competitive, that's really key.
Slava Rubin (13:05)
I know we're gonna be talking about specific companies here in just a second, but let's go to the next slide.
Jan-Erik Asplund (13:09)
Yeah. So just to break it up a little bit, you know, we sort of had a a very early era, very kind of even pre-LLMs, you know, when when this was being sold into universities and and research labs and places like Meta that were doing AI pre-LLMs, a lot of a lot of what we saw with the sort of launch of Chat GPT early on was that there was this shortage. and so just having access to GPUs and being able to offer that over the cloud.
was absolutely huge. and so you had companies like Coreweave and Lambda come up in that world. And you know, Coreweave famously was a Bitcoin miner before being in AI compute, which is why they had so many GPUs just hanging around in their in their office. what we sort of came up to then, you know, as you know, Chat GPT got better, Anthropic got better, this stuff started to disperse more.
We started to really see, you know, this kind of need for unique financing, ways to really scale up extremely quickly. and you saw a lot more players kind of come up into this space. You saw sort of the invent of the the neo cloud. and you started to see, you know, most famously you know, companies like companies like Lambda, Crusoe using their GPU supply as, you know, collateral with with banks to
take out huge loans, which they then use to buy more GPUs, creating this, you know, kind of circular engine of growth. Today we are in kind of the acceleration of that era, right? The infrastructure's the scale is so much bigger of what labs need, what what the hyperscalers need, and the, you know, the financing deals are so much bigger than they were before. And
A lot of these questions about yeah, long term demand, is it overbuilt are now starting to come more into play. but obviously demand could so far continues to keep growing. So yeah, that's kind of a overview.
Slava Rubin (15:06)
Great. who do we include in the neo cloud category?
Jan-Erik Asplund (15:13)
Yeah, so for NeoCloud, you know, I think of companies like Fluidstack, companies like Crusoe, companies like Together AI. it's a bit of a spectrum. So at the smaller end, you do have companies like you know, Fireworks, you have companies like Run Pod, basically just kind of you know, companies that create an alternative to going to Amazon, you know, Google for your compute.
Microsoft for your compute and maybe Coreweave as the kind of first wave of NeoClouds now kind of part of the, you know, part of the establishment. But yeah, that's the that's the idea.
Slava Rubin (15:49)
Cool. Let's go to the next slide.
So before we cover all these really quickly, a couple of quick questions from from the audience. The the GPU math that didn't include power costs, is that right?
Jan-Erik Asplund (16:03)
No, yeah, that that didn't include any of the cost. That was purely kind of the upside equation.
Slava Rubin (16:08)
Exactly. And then in regards to all of this shift, I know it's a broader macro question, but as everybody moves into AI infra, is this just all net negative for crypto mining?
Jan-Erik Asplund (16:23)
That's a good question. Actually, I hadn't thought too much about it because a lot of these companies came from crypto mining, obviously. So there was a little bit Yeah, exactly. yeah. But I would say as you know, the economics of I I think it's great for crypto miners, you know, who can convert their sites to AI, which is so much more in demand right now. so I think there's that. But I think yeah, it's it's not as profitable to to mine most crypto.
I would I would guess. so yeah.
Slava Rubin (16:52)
Cool. Let's talk about the competitive map. And and as you speak about this, if there's anything that you find to be defensible about a specific company, please highlight that, as opposed to them just being let's call a commodities across.
Jan-Erik Asplund (17:05)
Yeah. Perfect. Yeah. We'll definitely dive deeper on that once we go into each company. But yeah, just to sort of start off, you know, these are the main companies we'll look at: Lambda, Crusoe, FluidSec, Core Weave is is already public. these are kind of the earliest and the, you know, some of the biggest examples of of you know what we consider neo clouds in in the AI infrared. and yeah, some very interesting differences between them. I think on the second layer.
Or some other companies that folks might have heard of, you know, RunPod, Modal, Base 10, Fall AI. I would also throw in Together AI fireworks, which are, you know, very much more startup oriented, although they've grown with some of their customers as they've scaled. but the point is, you know, instead of having to sign a multi-year upfront contract, you can rent GPUs hourly. so that's one big thing. The other big thing is a lot of them have more.
a lot of them have more features built in for let's say you want to fine-tune a model on your data set. you know, you can go to to Fireworks or together AI and that's available through the API, right? Whereas like Crusoe and and Lambda, Core Weave are more oriented around, you know, large scale, need a ton of compute. We can handle everything else ourselves. these are a little bit more like developer tools on the bottom, run pod, modal, base 10, et cetera.
Or they'll help you create video, you know, create some kind of workflow pipeline with with video models, right? Which is a big one for for Foul AI. it's all about media generation and media inference. So you know those ones differentiate a little bit more on the developer experience, developer tooling, whereas the tier one companies here, as we're putting them, they are a little bit more large scale. although I think Fluidstack and Lambda are more developer centric versus yeah, CoreWave and Crusoe.
Slava Rubin (18:56)
what makes them more engineering focused?
Jan-Erik Asplund (18:59)
Yeah, so it's basically a little bit more tooling that's built in for doing things like let's say fine tuning, you know, doing evals on your on your models. It's built a little bit more with a lot of that stuff, you know, kind of built in to the to the experience versus a little bit more of the yeah.
Slava Rubin (19:20)
Yeah, I know that Stripe really ran away with the market because they were so engineering focused and had those types of features. Is that something you would kind of say Fluidstack is getting, or is that too strong of a statement?
Jan-Erik Asplund (19:34)
It's probably slightly too strong because Core Weave and Core Weave and Crusoe do as well. I mean, they did do they do as well, have a lot of this stuff. It's more about the the go-to-market. you know, they have a Fluidstack, Lambda have more of a self-serve experience. Actually, Lambda might not anymore, but the others do much more self-serve so developers can get started right away. which I think is a really, you know, it's really great for them because there obviously are a lot of startups trying to do this.
but yeah, they're all, you know, awake to the need for developer tooling to some to some extent.
Slava Rubin (20:09)
Not on your list here, but you know, SpaceX is part of their IPO just announced that they're now renting out their computer they built up as well, having some monster clients, including Anthropic and then also Google. And then separately about a week ago, Meta came out with a similar announcement. the stock popped, thinking that this is a whole new pretty large potential revenue line. are these direct competitors to this slide?
Jan-Erik Asplund (20:35)
Yeah, I think to some extent there is, especially with let's say Core Weave and Crusoe, there's some direct competition, especially at the high the higher end, the larger scaled companies. I think at the at the sort of on the other hand, right, all these companies are kind of competitive with each other. And yet, you know, they're all in the in the hundreds of millions, you know, billions in revenue.
the demand has just been so insane that you know it hasn't really slowed anyone down. But there definitely is overlap between what SpaceX and Meta are doing, which is essentially repurposing all these GPUs that they bought in order to start being a NeoCloud themselves. and then there's a few other players too, people have probably heard like Nebius, which was part of Yandex, the Russian Google, which has been a big player.
doing deals with Microsoft and Meta and then NScale as well, which is another really big one in the private market. trying to think of other ones, but yeah, I think that's those are basically the key ones.
Slava Rubin (21:33)
What's the I know we could talk about this at the end, but w before we talk about each company now next, what's the concern about there just being an overglut of build out in this space? And you kind of have the same thing happen from twenty-five years ago with fiber optics in the dot com era. You know, because we have Coreweave worth fifty, sixty billion, you have Nebius, you know, fifty sixty billion, you have Lambda right now.
Looking into IPO, Crusoe talking about thirty billion, Fluidstack just raised or is raising it at eighteen billion. N scale talking about IPO at thirty billion, let's call it. I mean these are pretty big numbers and there's a lot of them, and we haven't even talked about the impact of SpaceX and Meta. So can this all be bullish or is there a concern that maybe this is an overbuild?
Jan-Erik Asplund (22:23)
Yeah, I think, you know, with the sort of the AI infrastructure space in general, a lot of these, you know, that pull slide with how many gigawatts are being brought online. all this stuff is made years ago and it decisions were made years ago in advance that we would be where we are now. and so the you know, one terawatt, that's like for the future, but all the sort of financing and and
All the decisions are being made now. And so if there are changes to demand, then it would be yeah, obvious it would be very you know disruptive to the projections and the growth and the revenue of of the neo clouds. This, you know, obviously means that if if yeah, exactly, if there's any kind of slowdown, you could end up with a lot of
overbuilt capacity. and you might still have a giant, you know, AI might still be the future. You might just have a lot of companies that that don't survive kind of that shock and that transition. So I think that's probably the major risk in all these companies. Yeah.
Slava Rubin (23:30)
Okay, and then anchor here for a question, that's just a a big client, is that right?
Jan-Erik Asplund (23:33)
yeah, exactly. I who have been the key partners. Yeah, exactly. And for Core Weave, it was my you know, do huge deals with Microsoft. most of these have one or two hyperscalers or labs that are like really driving a ton of the revenue. So so customer concentration is also know, something to look out for.
Slava Rubin (23:51)
Agree. I think we've done a good job of the overview, talking quite a bit about it, answered several questions. Let's dive into the specific companies.
Jan-Erik Asplund (24:00)
So Fluidstack, yeah, is interesting company that was originally a Airbnb for bandwidth. Right. So you could sign up at home, let your computer get used by someone else for its compute. so you can imagine this was sort of the the early days of of AI, more so than where we are today. that was the original, you know, sort of idea was was renting GPUs over the over the internet.
We have sort of, you know, we we had the AI boom happen and FluidStack started getting into connecting, you know, you with data centers, right? So instead of your your neighbor's GPU, you'd be using a data center GPU that was underutilized. you know, it was some provider that maybe was only getting 50, 60% utilization. And so they could bump it up by working with FluidStack. and then FluidStack, what they focused on was the customer onboarding.
And they focused on the software configuration of the clusters so that they could be used for AI training and AI inference. So when we're talking about developer experience, you know, a lot of that was in the early days stuff like stuff like this, where you're trying to help AI startups just get up to speed and start using these GPIs and make it easy for them. So that was sort of where they got their initial product market fit. They had this very highly startup customer base and
Of those customers started getting really big. Character AI was one of the first big kind of AI success story companies. You've had pool side and AI coding, you've had Mistral in Europe all using Fluidstack, you know, getting really big on Fluidstack, graduating into their sort of enterprise offering, which they began having, where you would rent out more GPUs long term, similar to what the biggest neo clouds now are doing.
And this has, you know, helped them become yeah, one of the bigger players in the space and one of Anthropic's partners on data centers. And yeah, we don't have 2025 revenue for sure yet, but this is sort of our our estimate right now is that they got up to 900 million annualized 400 million sort of trailing 12 months revenue for a 45x multiple.
Slava Rubin (26:13)
We're in July of sixteen twenty six. So do you have any sense of what twenty six is looking like, or we in the dark there?
Jan-Erik Asplund (26:22)
a little in the dark, but definitely the Anthropic deal has has been has been huge. So I think that was that was at the end of twenty twenty five and I would expect that they've had a very good twenty twenty six.
Slava Rubin (26:35)
What's the size of the Anthropic deal again?
Jan-Erik Asplund (26:39)
it's hard to say there's a fifty, you know, fifty billion was the top line number. That's not just going to Fluidstack. You know, that also means sort of building a lot of the a lot of the data centers. so Fluidstack is one of the key partners. It's unclear. So it's unclear how much of that flows straight to Fluidstack, but yeah.
Slava Rubin (26:56)
So I remember when we were talking about Coreweave before it went public, we were saying Coreweave is basically super connected to NVIDIA. So where NVIDIA goes, Coreweave will go and probably have more volatility. Is that similar here with Fluidstack and Anthropic? Where Anthropic goes, Fluidstack goes, or is that a bit too tightly connected?
Jan-Erik Asplund (27:14)
I think it's fair. I mean, it seems like they're definitely, you know, have a lot of belief in FluidStack to come out and do this deal and make them sort of a named, you know, big named partner. and for Anthropic, it has a great there's a great reason to do it, which is you you sort of diversify your computing sources and you get to, you know, encourage another player to exist, but be sort of you know, keep them very tight as Anthropic gets to with Fluidstack.
You know, it's maybe a little different from Coreweave because there's just so many more options now. so it's not as tightly wound, but is definitely very promising.
And then just as a as a sort of aside about FluidStack, but also about other neo clouds, you know, the sort of arrangements here that are being done are are kind of complex and worth understanding, like when it comes to talking about Anthropic, right? Anthropic is, you know, in need of tons of computing capacity. Fluidstack doesn't have, you know, these data center sites that are all plugged into power. those are owned by, you know.
For example, in this case, a bunch of Bitcoin miners who have been around for a while. Fluidstack is sort of in the middle. They're the, you know, the operator of those sites. They turn them, you know, they turn these GPUs into AI infrastructure, which again is not just sort of a plug-in, plug-and play. It requires a bunch of a bunch of work. And then Google is also involved. Google, you know, acting as a guarantor on the lease, making it sort of credible.
for for the for the site and for the lenders to actually finance all of this construction. And then Google getting you know equity exposure to the underlying operators. They get distribution for their chips that they are building. And so it's it's it's kind of a whole you know complicated complicated web, you know, where you have Google using Fluidstack to put Google design chips into facilities so Anthropic can have compute.
and you know, Bitcoin miners are sort of the data center, you know, the real estate play behind behind it. so yeah, that's something worth understanding for all the sort of AI neo clouds that are that are not also sort of owning data centers like CoreWeb is.
Slava Rubin (29:21)
Sorry, so Google owns like fourteen percent of Terra Wolf and five percent of Cypher, is that right?
Jan-Erik Asplund (29:26)
Yeah, yeah.
Slava Rubin (29:27)
What what does Google own of Fluidstack?
Jan-Erik Asplund (29:29)
with Fluidstack, I think that I don't think that Google is an equity investor, although they they might be actually. It's mainly that they're you know backing these these giant leases that Fluidstack has signed in. And it's all a part of trying to sort of flood the market with Google's chip, which is competitive with NVIDIA's GPUs.
Slava Rubin (29:49)
Interesting. I mean, Hut 8 was actually one of my picks a while back on the podcast. I mean Hut 8's been on quite the ride. okay, great. And Fluidstack right now we think is at about an eighteen billion dollar valuation in private markets. We were saying that's about a 20x from last year's revenues. we're not exactly sure what this year's are looking like, but it's been growing pretty rapidly. and it actually did raise pretty sequentially pretty rapidly as well. So
It is a sort of company which has high risk, high return potential. Let's go to the next slide, cover the next company.
Jan-Erik Asplund (30:22)
Yeah, Crusoe is really interesting one that we've covered for a while at Sacra. And they are essentially using the sort of byproducts of, you know, natural gas facilities. Most notably, they did this in Texas. You know, you end up when you process natural gas, you have all this stranded, stranded gas, which gets burned off and just wasted into the atmosphere. And so what they Crusoe did is they put these data centers right on top of that gas and then used it.
Converted that sort of wasted energy into into, well, originally Bitcoin mining again is a very common theme here. So they were using it to mine Bitcoin. but then 2022, 2023 came around and they started to use it to power compute, most notably so far for OpenAI and Oracle, which, you know, all that news about Stargate and building a huge data center campus.
in Albiline, Texas, you know, Crusoe was sort of the unnamed, unmentioned partner that was that was building out or or had those sites and was sort of a key player there so that was that was big big for them and that's sort of what got them to to 500 million this year there have been some sort of recently they've been in the news for delays
with some of these later phases of the Stargate project not going quite as planned, just having delays and some of the financing. So yeah, it's it's a it's a very interesting company, different approach, you know, using sort of stranded energy. So they've had a sort of environmentally friendly, you know, positioning that they've been able to they've been able to use. and yeah, another converted Bitcoin miner.
Slava Rubin (32:00)
So this one on price, just so we jump in here, it's it's a little bit tricky. So they did raise in twenty-five on the five hundred million projected, and they raised like about a ten billion dollar valuation, which would be like twenty X of revenue from twenty-five. We're not exactly sure what their twenty-six revenue is, but the rumor is that they're raising it at a thirty billion. If you use the twenty-five revenues compared to fluid stacks, similarly, then it's a sixty X of twenty-five revenue. I'm
guessing that the revenues in twenty six are higher than five hundred million, so that multiple won't be as high. It's hard to predict what it would be. I mean, even if it was a double projected a billion, which is pretty dramatic. I'm not saying I know that it is or it isn't, but that would still be a thirty X, which in comparison is more rich than Fluidstack, but Fluidstack had that massive jump to nine hundred million last year and we don't know what their twenty six is. Is that right?
Jan-Erik Asplund (32:50)
Yeah, exactly. It's all a little slippy. So just to be aware of that I think is important.
Slava Rubin (32:56)
But what is nice on the Crusoe chart is it seems like it's stepping stones as opposed to the Fluidstack one, which both positive and negative, it's a massive jump. Right. So that could be a game changer for the whole company or it could be volatile.
Jan-Erik Asplund (33:09)
Yeah, I definitely agree.
Slava Rubin (33:11)
Let's go to next slide. Next, company
Jan-Erik Asplund (33:13)
Sweet. And then Lambda, you know, has been around for a while, longer than Crusoe and Falluestac. And they sort of started out again, like selling into academia, into research labs, universities, and they really early on were really known as this kind of developer-friendly alternative to Core Weave, you know, where you could you could rent on a much smaller scale and there was more kind of
Polish on the software side and the configuration side. And they, you know, grew from there. And over time, they've sort of, they sort of moved a little bit away from that initial positioning. So they got rid of their you know, on prem hardware business where you could you could have a workstation physically in your university lab. and they're a little bit more of just a you know, straight up neo cloud competitor to Core Weave now. they have a tight relationship with
With NVIDIA, you know, which we we haven't talked a lot about, but NVIDIA kind of using companies like Lambda as a strategic partner, giving them GPUs, you know, and then also, you know, leasing capacity back from Lambda, kind of guaranteeing utilization for the company, validating the platform, helping them raise more money, helping them get more customers in a way that locks in NVIDIA GPUs as the dominant form of compute.
In the ecosystem and moves demand away from your Googles who have their own chips, your Amazons who are developing their own chips. This is a big dynamic with Orweave, but it's probably the next biggest company with which it's a dynamic is Lambda, where you're benefiting from Nvidia's attempts to drive competition in the ecosystem through companies like this. So what we have with Lambda is basically taking a little bit of this developer-centric
approach this developer DNA and now scaling it up into the sort of gigawatt scale that companies need, you know, these days. and yeah, a little bit less control of the physical sites. you know, Crusoe, for example, is a little bit more of a had a little bit more of a real estate play or data center play. Same with Core Weave. Lambda is a little bit more based on
you know, co-located GPUs and and rentals of of of spaces in data centers. So something to to understand there.
Slava Rubin (35:30)
I know this sounds awkward to say, but this one seems cheap compared to the other ones. Is it just because it's more mature and it's a little closer to IPO? So the the price is becoming a little bit more realistic, or what's your point of view on that? I mean, 'cause it's kind of pl you know, playing like seven and a half to ten X of of revenue, right? Which is not cheap in the normal world, but comparatively.
Jan-Erik Asplund (35:52)
Yeah, I I agree. I I think it's a little bit cheaper. It's not growing quite as fast. and it also is, I I think probably part of it is the fact that it's it is close to going public. I I believe originally they had talked about it going public by now. so it's a slightly delayed, but this is probably the next one to go public of the ones on the list.
Slava Rubin (36:12)
So is it worth trying to buy now, you think, or wait till it goes public?
Jan-Erik Asplund (36:18)
It's a question. I think it's at five, well, at seven or eight billion. yeah, I don't think it's a bad idea. I I think I haven't checked Coreweave stock price. That's what I would probably look at to see to feel comfortable with the sort of trend in the market. I mean, I'd say it's worth it to try.
Slava Rubin (36:36)
It's a great moment to think about that for a second. If you go to the Coreweave chart, so their one month, they're down almost twenty percent. but then their one year for the last twelve months has seen quite a bit of volatility, but it is down. it was pretty high a year ago, but you're down thirty-four percent. we have seen as low as, you know, seventy ish, it's right now at eighty-six.
You know, it's it's multi year from its IPO. i is not the prettiest of charts, but but we have to understand that an IPO is at like 40 and it's now double the IPO. So you can't get upset. it was as high as like, you know, 160 or something. So I do think Coreweave had the benefit of kind of being alone, and it was kind of like the only thing you could invest in, like it. I I do think this is becoming more competitive.
I do think it is IPO season, which is a compliment for these companies. I do think there'll be IPO pops. I just don't know what the price will be six months after IPO, after the lockup. but I guess that's why we're all talking here. let's go to the next slide.
Yeah, so talk to us about the market.
Jan-Erik Asplund (37:45)
Yeah. So I think there's kind of the the bear case or the the bull case is basically that you have, you know, this demand for AI compute that is just rising continuously. seemingly without real end. I mean, even even even as the models maybe the jump from iteration to iteration is not mind blowing anymore. We're still seeing that.
the demand explode because it's just it's still disseminating across the sort of economy. enterprise is still very far behind on that adoption. So you know, models are getting bigger and companies are using more and more of it. it hasn't really, you know, in a in a massive way broken through with consumer. so there's still a lot of sort of upside room to grow. And you have
sort of the biggest, you know, players in the space of compute, of cloud compute, right? Hyperscalers, the AWS, Microsoft Azure, Google, Google Cloud. they cannot bring they can't bring capacity online quickly enough to just to to satisfy it, right? Even as big as they are, you can't acquire the land and build the data centers and get power hooked up and and
Get tens of thousands of GPUs when NVIDIA doesn't really want to give them to you as much as they want to give them to Coreweave and Lambda and others. and so it creates a market for neo clouds for all these companies that we've talked about, whether serving, you know, the biggest comp like, I mean, Microsoft is one of the biggest customers of these kinds of of CoreWeave, Crusoe and others, and then all the way down to startups that are, you know, building AI models and training them. So
That's really the core the core bull case and what has propelled these companies so far, you know, from roughly 2022 to 2026. and neocloids also have the advantage of of this kind of fewer you know internal constraints. They can move faster, they can partner with NVIDIA. and so there's a lot of sort of degrees of freedom for them there. The risks, however, you know, are are there. one of the risks we talked about is
question of if demand slows down, right? If demand slows down, models, you know, stop getting better, let's say, you will have obviously disruptions to a lot of the the revenue projections and and growth and build up plans that these companies have. You'll also have the risk of OpenAI wanting to control their own infrastructure because these chips are so important, you know, that now open AI is working with Broadcom.
To build out their own their own chips. you have OpenAI doing a deal with Cerebrus, which we've talked about before, which is a different kind of chip fundamentally. So we're talking about frontier labs taking more of this stuff internal and moving spend away from neo clouds. So that's a big risk as well. but yeah, I I yeah, so I'll I'll stop there.
But those are kind of the two big downsides, you know, or two of the biggest downsides for these companies in the future, potentially.
Slava Rubin (40:51)
Yeah, I I think it's a very dynamic market and we're all on this converse all at this conversation trying to figure out how do we make money. I think a hard way to make money, unless you have inside information or super connected to the space or the company, is knowing which exact company to pick. I think that is tricky based on knowing the fundamentals and why that leadership team is better.
And why they have better contracts, et cetera. The more macro question, is this a good market to invest into? And I would say that's a big strong yes. I would compare it to like cloud compute. When things went to the cloud, there was a question, was Amazon gonna win that? Was Microsoft gonna win that? All of a is Oracle in that? Next thing you know, they're all doing directionally well.
Because that market was a strong market. And I think similarly, the need for AI infrastructure is only going to increase. And I do think that one is the one of the existential threats is a massive simplification of the compute needed for the models. If you go from these gargantuan compute spend models to all of a sudden one one thousandth of the spend on compute because things became more efficient, I think that could be super powerful.
But even in that situation, I still think the net usage of AI compute will go up because of just the amount of compute that will be done, even though the compute per query might go down significantly. Comparatively, even though our TVs are a hundred times more efficient now than they were a decade ago, we're still using more electricity across the board. We're using less electricity per TV, but we're using a lot more electricity across the board for all the various things we need.
And I think that it's the same situation. I'm quite bullish on this market. I just don't have a strong point of view on which horse to pick. So in that situation, I would think to diversify across the various players. Also very concerning to me from an opportunity and a and a risk is look at SpaceX, look at Meta. Who else is going to get into this market? Is it is, you know, are we going to be hearing new announcements of other players sharing their c overbuild capacity?
into the market. And then eventually, who pays for this? So I mean, that's that is b very bullish because you have strong, huge companies moving into the space, they wouldn't be wasting their investment unless there's a good opportunity. but obviously the the music can stop and you have to be making sure you have a chair to sit on. So you have to know where you want to invest. So let's just talk with a final slide about these companies and the bull case. What's possible?
Jan-Erik Asplund (43:36)
Yeah, I think, you know, one other thing I'd add to what you said, which I you know, I thought that was great. I agree with everything you said, is when it comes to thinking about the future and where compute demand is coming from, I think it is worth considering also that, you know, the typical large language model that we chat with is not the only form of AI, right? And we are in the very, very early, early, early stages of robotics, which we've had, you know, webinars on.
And if you think about LLMs, it's like we're at 2022 era of LLMs when it comes to robotics today. And so that is gonna improve over time and that requires compute all the same and also requires you know more compute because you're not just processing text, but processing real-time video data and also processing not in you know text dimension only, but in in 3D space.
and incorporating, you know, force. so things like that are going to drive even more demand for compute. So I yeah, personally, I'm not too worried on the bear case level about, you know, we suddenly don't need the amount of data centers we have today. just because of that, sort of the the growth and demand that will come from a bunch of different angles, not just sort of typical LLM training. I think it's possible we do reach that point at some point, but
I think the the wider bear case, you know, might have to do with yeah, the open source being one being one component, labs bringing more of this stuff in-house, making it not as appealing to to you know be a neo cloud company. in this case, you end up with these companies, the the multiple, you know, craters, right? Because you're not just dealing with you're not dealing with any kind of real specialized scarce inventory at that point. You're
basically looking at this company as a you boil it back down to the GPU math, you know, with all the realism of of paying for rent, paying for power, depreciation over time. And that starts to look like more of a sort of infrastructure business, you know, maybe valued at one or even below one two X revenue, which for all of these companies at this point would be would be, you know,
Down from where they are today. The sort of bull case, however, is that you know compute continues to be demanded and hyperscalers continue to fail to supply it in the amounts and in the sort of with the urgency of timelines that it's needed. and so these players, especially the ones that we've pointed out here in the private markets, FluidStack, Brusso, and Lambda, you know, by virtue of being the biggest neo clouds already.
they're gonna to disproportionately benefit because they have the ability to get the financing needed to go, you know, to get the next data center spin spun up and to acquire the customers, you know, and sort of own the demand and, you know, sort of ride the wave. so I think between these three, you know, you're looking at three different businesses sort of predicated on three different advantages, right? For Crusoe, it's like about
power generation and getting power much more cheaply. for Fluidstack, it's kind of like being embedded in with Anthropic and with Google in that ecosystem. And then for Lambda, it's sort of a tighter relationship with Nvidia. You know, this have this cloud product, more established developer relationships with customers. So it's kind of like pick your poison between those three. in the bull case, you know, I think
Although I think in the in the bull bull case, they probably all, you know, have a measure of success. But I think it's like picking what you maybe are comfortable with and maybe understand better than than the others, is how I would think about it.
Slava Rubin (47:19)
Yeah, and in general they're moving pretty consistently from your predictions, is that right?
Jan-Erik Asplund (47:22)
Yeah, I I think a lot of it has to do with the macro with with respect to these companies. Yeah.
Slava Rubin (47:28)
And if just to put you on the spot, if you had to pick only one right now to put a thousand dollars into in the private markets, which one do you pick?
Jan-Erik Asplund (47:34)
I think I kind of like Fluidstack. I mean, I might just be kind of recency bias. They're sort of they've had a bunch of wins, you know, in a row doing the Anthropic deal and whatnot. But I like, you know, they're they're less capital intensive. They're sort of a middleman, you know, software software centric, more of a software centric player, so higher gross margin theoretically. and
Less of a risk around the infra build out like you would for a Crusoe who own these sites. So that's how I think about it. But
Slava Rubin (48:03)
I totally agree with that. and just to hedge, I do think the right move in this space is to think about it as a basket and a movement across the entire macro space. is there anything else you want to cover, Jan-Erik
Jan-Erik Asplund (48:16)
No, I think that's that's that's great and great call on Hut 8
Slava Rubin (48:21)
Nice, nice. well we covered a lot. We had a whole ton of questions. Tried to cover what is the AI infrastructure space, which is confusing because there's a lot to cover here. the market is starting to go public in the last few years, and now you're gonna see more coming in the next few years. you know, be careful out there as to what you invest in, make sure you're doing your own research. but there are a lot of opportunities as this could be a bullish market moving forward. So thank you, Jan-Erik, and thank you everybody else for joining.
And enjoy your final week at the World Cup for all those that are celebrating. Have a good one
