Welcome And Europe’s GPU Moment

SPEAKER_01 0:01

Welcome to Uplink, where we explore the world of digital infrastructure, uncovering the technology fueling AI and cloud innovation with the leaders making it happen. We're here in London, uh, of course, and I had to catch up with Ben because Ben's been doing a whole heap of cool stuff all across the GPU space. I think I've been following you, I don't know how long you've been pumping all this interesting content out.

SPEAKER_00 0:25

It's been probably a couple of years at this point.

SPEAKER_01 0:27

A couple of years, yeah. Yeah. You sort of sort of followed this journey of well, the explosion of, I guess, training and GPUs and usage of NVIDIA chips. And then also I think you've got a much more um focus around Europe and a much better I mean, we we spend our lives everywhere, but your insights into Europe is really fascinating. So I'd love to unpack just what you've seen change, where we are, any insights, any predictions what's happening. Absolutely. And I know you've got a very big um play on the finance side of things as well. So I'd be curious to just sort of unpack some numbers in your mind around what you think makes sense or how how things are getting financed and contracted, and if that's changed at all, or yeah, no, absolutely.

SPEAKER_00 1:09

So and yeah, Europe's a really, really interesting market at the moment. Um, obviously, like for the past couple of years or so, so much of the focus has been on the US. So there's the the multi-gigawatt sites here, or like there's the meta Hyperion deal, which is the same size as Manhattan, or there's the one Project Stratos recently, which is double the size of Manhattan, because apparently now Manhattan is a standard index unit of measure for data centers, right? Yeah, it used to be football fields, now it's Manhattan.

SPEAKER_01 1:42

A friend of mine worked on that actually. Uh yeah, from from Meta. Yeah, yeah, that's really cool. He was the sort of the that came up with the concept.

SPEAKER_00 1:49

Yeah, yeah, yeah. I I'm here for the name, right? Like Hyperion and Prometheus, Prometheus being the the titan that stole fire from the gods and gave it to humanity, right? Like there's and then Hyperion's like the light bringer, which in the context of AI is quite an interesting one, right? But whoever um whoever came up with that name, I think they they deserve it. So what is that?

SPEAKER_01 2:08

So what is what as in Prometheus is what the name of the data center?

SPEAKER_00 2:11

Yeah, so they said um Prometheus was the one gigawatt site, and Hyperion's the five gigawatt site, which is which again when you think of search, right? Which is like so that's meta, is it? Exactly.

SPEAKER_01 2:20

So that's meta. Yeah, yeah. What would be is that the largest that were like five or something? I don't know. What you can't go much bigger than that, can you?

SPEAKER_00 2:27

I mean, I think N scale's recent land acquisition, uh tower or something. Um that's eight, and then there was Fermi America, which was meant to be 13 gigawatts, I think it was, but that's now there's a whole lot going on with that. Um from exits from CEOs and litigation and premise, all kinds of stuff that's going on there. Um yeah, I mean, so that's the American market, right? It's these these massive deals that Europe just cannot do in the same way because we don't have a spare five gigawatts or eight gigawatts or 13 gigawatts lying around, right? Which means that the smaller pockets of land and the smaller pockets of power that have typically been overlooked up until this point, right? Because it's been very much hyperscale driven. Yes.

SPEAKER_01 3:17

Whereby you've got the AWS A GCP buying.

SPEAKER_00 3:20

Yeah, for their own or going via the the bigger neo clouds to assist in that procurement because ultimately they can only deploy so fast and demand is so high, they need those deployments like the core weaves, like the Nebbiases, yeah, the N-scales, right? And from an economies of scale standpoint, it doesn't make sense for the hyperscalers to go below a certain level, and then it also doesn't make sense for the bigger neo clouds to go below a certain level because there's that the same dynamic that you have with the hyperscalers procuring through the neo clouds for certain capacity, you end up with the same situation, right? We're below like I don't know, 40 megawatts or something. It doesn't make sense to allocate the time. So there's all of these.

SPEAKER_01 3:59

So that's what you're saying. Yeah, I think that's the feedback we get is hyperscalers don't want to look at anything below 40, 50 megawatts. Yes. Um, which was which is kind of like how we we sort of fit into the puzzle. We sort of fit this, we could do sort of two megawatt and below. Um it fit in the cracks, as they say. Yeah. Yeah.

SPEAKER_00 4:16

Which is a really interesting size in the UK, in Europe as well, because there's distressed assets, commercial assets, old industrial assets. With access to power, with access to power, with the grid

Financing Small Sites And Thin Margins

SPEAKER_00 4:30

connections that potentially become interesting at that scale if you're able to monetize it. And that is ultimately then dictated by like, can you finance it? Right? Because two megawatts is it's a large amount of capacity, but it while the returns on paper can look interesting from like an IRR standpoint or from an MPV standpoint, right? The the absolute return typically isn't high enough for the investment. For the investment, right? Or for the investment. To retrocit it, you mean, or to build it out? Generally, unless you've got some sort of anchor contract, which means that the economics are viable from like a cost of capital standpoint.

SPEAKER_01 5:06

Yeah, correct. Um Which is the game that we're sort of talking about is really financy. Like it's all about like how you finance these things. Exactly. Yeah. Exactly. Those deals are difficult to construct, right?

SPEAKER_00 5:19

Um it's very, very large amounts of money, like 50, 60, 70 millimegawatt potentially, um, depending on the hardware generation. And then there's all the supporting costs, and it'll be like typically 15 to 25% of um gross revenue for non-power OPEX, and then power cost ends up feeding into I mean, making the margins potentially very thin if you're only offering bare metal, and then if you're offering like a two-megawatt site with all of those costs factored in plus 10, 12, 15, 17% cost of capital on a three-year loan whereby there's some very aggressive amortization taking place. That does not leave an awful lot of room to then actually be able to pay for the business to make things work right.

SPEAKER_01 6:03

Which I think is what you've originally you were sharing a lot of numbers, and it's like this doesn't even make sense. I think is some of the pieces of the puzzle that was really interesting. Um so that I guess it's have you seen a big change recently? Like what we've seen a big change since sort of Opus or the launch of chat um of uh Anthropic's Clawed Coding in sort of late December, and then in January it all sort of took off, and these agents took off. And so we've got all these customers from an inference standpoint that have just gone through the roof um in terms of revenue and and and we just saw basically only a few winners. It was like Anthropic and OpenAI were the ones actually getting income or revenue. Um NVIDIA was winning everywhere, and then you had the hyperscalers and so forth providing in effect GPUs for those. So they were all getting revenue is sort of staying in that space. But what's happened since, say, January, February, March is it's sort of come down the next level, and there's a whole range of companies that have just exploded from an inference standpoint. Um, so that is kind of very different for the last three years that we've seen at least, which is why we've sort of landed in that inference cloud space. We see some tremendous value that we can bring by being in all these countries and being distributed and all those factors, and not having to have training um platform, i.e., you know, a gigawatt campus with every GPU running exactly the same moment. And then so it's yeah, super hard to um uh build the build that, which is why I think a lot of these interesting companies, particularly in Europe most recently, have all popped out of nowhere, really. Um, yeah, I mean maybe are you what give us your perspective on how that's all playing out.

SPEAKER_00 7:40

Yeah, so these companies are it's been such the way that the market has developed for the past couple of years is really interesting, right? Because we we've gone from being basically a supply-led market, right? Where by 2024 you've got a few players that are deploying massive amounts of capacity across Europe. Um, Northern Data being one of them, 22,000 H100s online, massive capacity, massive fleet across Europe. Um who who is using them then? Who are they? Typically training at that point, right? So big training customers who need those big clusters. And they're US companies or these some European companies, some Mistral, for example, right? Yeah. Well, Mistral seems to be the Now they're an LCP. Yeah, right. Yeah, I sort of they're now they're they're procuring hardware, they're providing their own, supplying their own compute. They're probably I'm sure there's something interesting that's gone on there in terms of like how the NVD relationship works. Yeah, yes.

SPEAKER_01 8:37

Um, they have to shore up their own supply. And that's exactly it. Uh I wonder whether they needed to get into this space, it was or or as in they wanted to get into this space, or they actually just hadn't a choice. Maybe, maybe that was what it was. Um not sure.

SPEAKER_00 8:52

I mean, the the way that this market is developed, right? It's vertical integration seems to be the way to go. Because if you're again back in 2024, right, you're trying to get hardware that hasn't been broadly financed online with exotic financing structures. So like the the first GPU-backed deals that Coreweave did, for example, right? Interesting way of doing things, hadn't been done before, and these companies were also new at that point. What are they doing? They're having

Inference Boom And The New Winners

SPEAKER_00 9:20

to deploy into co-location data centers, right? And that is a massive cost center, right? Um, at the time, demand wasn't anywhere near as high. GPU prices fell from what was it, I think eight dollars per H100 per hour down to a dollar seventy over the course of a year or so. And when you've brought stuff online at 15 to 17% cost of capital, and you've built your projections around eight dollars, and then all of a sudden the bottom has fallen out of the market. That doesn't make things easy. What do you think that was?

SPEAKER_01 9:51

Because now all of a sudden it's reversed as well. So what use cases, right? And I again I think So was there an overbuild of H100s or something? Even maybe the use case that was it there.

SPEAKER_00 10:02

Potentially, um maybe there was some overestimation of how capable the models were going to get as quickly as they did, right? Because like back in 24 it's what GPT three, GPT-4, relative to the capability of the stuff that we've got now, it's like it's it's it's not the same, right?

SPEAKER_01 10:18

So it's almost the whole market's matured in what is a couple of years, which is kind of funny. Yeah, it's insane when you think about it.

SPEAKER_00 10:25

Um but then I think to Mistral's point, right? Like why have they done it? It's the same thing that a lot of the neo clouds are doing. That they're no longer building or going into co-location data centers unless they have like massive demand and they need to fill it right now. A lot of them are typically building their own facilities. Why are they doing that? Well, you've got more control over your supply chain, you've got more control over your costs, and then if you're deploying modular, for example, right, your build cost per megawatt is going to be significantly lower. Modular being like a uh it's like containerized containers, yeah.

SPEAKER_01 11:00

So they are significantly better priced.

SPEAKER_00 11:02

Yeah, yeah, yeah. So eight to ten millimegawatt, including civils versus sixteen to eighteen megawatt if we're going for the conventional plus much faster route to market.

SPEAKER_01 11:12

Yeah. Um it does make sense. I think in Europe that would make a lot of sense because there's so much like I'd say power that's all you know, there's wind farms everywhere, there's all these um uh like renewables all over the place. Getting behind the meter, I suppose, and getting access to that. Do they still get grid access as well? Or does it like when it's not windy there? So I mean that's part of the bottleneck.

SPEAKER_00 11:32

You know, like European grids are slow for from opening. Slow everywhere, but yeah, I think especially in the UK.

SPEAKER_01 11:40

Yeah, yeah. Europe and the UK. I think a lot of regulation, a lot of rules and restrictions, a lot of um eco and uh you know sustainability access have put in place, which you know made it may have made sense, but it's making it very hard to compete, I think, on that sort of thing.

SPEAKER_00 11:57

But I think that brings us quite nicely back to like why is Europe interesting at the moment, right? Because a couple of years ago, yes, Europe has been really slow to get going. But if you look at the the kind of pushback that is now happening in the US, right, around like this DC is going to use however much water per year, the impact that it's going to have on utility bills for the residents of the area. And I think a lot of people are suddenly warming up to the idea that no, like we probably do need some regulation, right? And that's the way that Europe typically works, right? Stability is generally put over speed or um pace of innovation, for better or worse, right? Risk management generally tends to be a bit higher from an environmental or from a social standpoint. And now that there is that pushback happening in the US, the systems to navigate that kind of pushback in the US don't really exist in the same way, right? You don't have that kind of, I don't know, class action litigation, for example, or pushback against data centers, the the the way for corporates to engage meaningfully in that way, um, the systems and processes that you might find in Europe possibly don't exist in the same way. So Europe is an interesting market at the moment because while there is regulation, the regulations exist.

SPEAKER_01 13:14

What they're a burden, but when they have a way when you tick the box.

SPEAKER_00 13:30

In the same um, in a way that you wouldn't necessarily get in the US, which means that potentially, in this instance, maybe things in Europe can move a little bit faster. Now, that being said, it's obviously still going to be American money that ends up paying for the majority of this because the US markets are much deeper, there's a lot more liquidity, but and the European market is very fragmented. Um, but things are moving in a in an interesting direction at the moment.

SPEAKER_01 13:52

Yeah, I've been on this world tour. Um I've been to Milan, Paris, Bologna. One of our customers is Lamborghini, that was pretty cool. Uh then we went to uh back to Milan, then did Zurich, um, Frankfurt, Paris, and then London. And everyone's sort of as it's interesting just to get different perspectives. We've got a financial services customer, one of the big ones in Zurich, um, actually sorry, in Par in Paris that is is using a whole heap of GPUs for trading purposes. So they're using they actually got their own models that they're using for trading, which I found really fascinating. It's the first time I've started to hear a lot of enterprises actually leveraging GPUs as opposed to just US-based sort of um AI firms in effect. So I think that's the first example I've started to see where it's prolificating out to traditional enterprise cycle finance companies using their own GPUs and their own models and what have you for their own purposes. Um and then the other piece is a huge number of these companies that are then helping code as code or you you know, um particularly from an inference standpoint, the importance of becoming closer to their users is happening. And so it's the rollout is happening. We're in a good spot at the moment. Yeah. Yeah. The rollout's definitely occurring right now in the US because the adoption is so strong there. But as soon as that adoption pushes through, I guess it's like um, you know, Facebook users and so forth, you obviously need all the inference here. So as this progresses down at ChatGPT, Anthropic, all those, we're going to see huge builds that have to be here as opposed to are here because it was easy. Now, the other thing I I get, and this is what I want your perspective on. So many random people have told me, oh, there's all this um power in Glasgow, there's all this or something because of the, I don't know, maybe there's um wind farms. I don't know what's up there. Um and Sweden comes up a lot. Uh so I don't, there's lots of random locations that were sort of brought up as where these data centres are building. Is it because they've got I mean Norway maybe?

SPEAKER_00 15:55

I don't know if Norway, huge amounts of um power in Norway at the moment, and ultimately it's because power is cheap. Yes, really, really cheap. Um Norway's an interesting So literally it's just following the cheap power in Europe. Um sort of thing. Norway's interesting because it's the grid is very hydroelectric. Yes, um heavy. So it's all hot, okay. So renewable as well. Um, and there are lots of interesting sites that are miles away from anywhere, which from a large data center build out sampoint, you're not necessarily going to get the same kinds of pushback. Now, whether or not they're too far away, so to your point, around customers now needing the inference endpoints to be closer to the customers if you've got multiple gigawatts in northern Norway, depending on what the networking and the latency is looking like, maybe it's suitable for some use cases.

SPEAKER_01 16:49

Yeah, not for probably not voice uh maybe, but uh yeah. I think the answer is though, if you can't guess it's here, yeah, then the second best is there as opposed to trying to push it back to the US or something. So where was um was it the Nebius and NSCale? Yep. They had some massive announcements.

SPEAKER_00 17:06

Yeah, so NSCale have got a couple of facilities in Norway. Um Nebius have got what was that eight billion dollar thing?

SPEAKER_01 17:13

Which one? Oh, I don't know. There's too many of them.

SPEAKER_00 17:17

Um yeah, it's become Nscale have also got something in Portugal, I think, at the start campus in Sinas. Um Nebias have got something in Finland. Yeah. Quite a large facility. What is Finland? Cheap power again. Well again, is it um?

SPEAKER_01 17:34

The order, um, just from the way that they've got toaps of oil, I know, uh or at least Norway's got plenty of that. Yeah, it's cheap, plentiful power.

SPEAKER_00 17:44

Um I mean, again, like margins have been historically so thin because it's GPU per hour and demand has been limited that the places where you can maximize your margin, ultimately power is a huge part of OPEX, right? Given how power hungry these large-scale deployments are. So it it makes sense that that's where exactly, right? But then I think as we move

Tokens Change The Business Model

SPEAKER_00 18:10

from GPU per hour rental to token-based revenues, right, which ends up becoming a little bit more like a a utility. Um, in that scenario, then sensitivity to power becomes less because revenue potential ends up being so, so much higher on a token consumption basis. It ends up looking a bit more like the um the Xerox copier model, right? You're paying per copy as opposed to paying for the machine. At the moment, people are paying for the machine, then having to pay up for the copy on top, right, or for the deployment on top. But then if you're able to just consume tokens and point intelligence at things, then it's a very different operator, very different business model.

SPEAKER_01 18:49

And are you seeing many folks go to that model yet? Or is it still at the moment they're really contracting the whole machine out? And that's probably a function of the finance, I assume. Like when you look at Coreweave, I mean their CEO will constantly talk about wrapping, you know, a triple A credit-rated company up with a five billion dollar deal, put all the components into it, lock that, and that's one thing, and then moves on to the next. And so I guess that's harder if you're saying I'm a token generator, unless you are contracting the number of tokens out, but then you're really just contracting the compute anyway. So it's just a different way.

SPEAKER_00 19:21

I think that that at the moment, the for these large-scale, like really large-scale build-outs or like massively distributed clusters, whereby you can have like a server at the end of someone's road, for example, that then serves that local area. You need to be able to finance that on a we don't know. On 100% exactly. So that has to be on a 100% debt basis, and you need to know that there is going to be enough diversity of customers there that the fact that one of them by themselves isn't necessarily credit worthy isn't a problem because in aggregate you have basically in credit worthiness. Like internet connection or whatever. Yeah, yeah, not really that worried. Yeah. But we're not there yet. No. Because the the consumption's not yet high enough and potentially even like the use cases aren't yet compelling enough.

SPEAKER_01 20:09

So the other piece is because I was just thinking that through, like, let's say it is tokens. If you're ChatGPT or if you're anthropic, are you I mean, they're I presume they're really creating their own tokens today. Do they have an ability to just take a token from some random location, or are they really needing to bring their model and run it back on a bunch of GPUs? You know what I mean?

SPEAKER_00 20:30

It's a is token really that so it's so the you've got the input versus the output tokens, right? So you deploy a model and then anthropic's pricing, I think, is a really good topic at the moment, right? Because you've had a number of different customers who've spent hundreds of millions of dollars because someone gave the entire organization access to AI with absolutely no limits and they've spent $500 million in a month. That sounds difficult to deal with, right? But like I a lot of the the models have now got to a level whereby there isn't much that they can't do, right? But then in if you're every time you ask a question, that's going to be consuming input tokens. Every time there is an outputted answer, that is going to be the output tokens. And there's a price point for the input, and there's a price point for the output. Sorry, explain that. Input being You type something out, you fire it in, um, and then when it ingests your question, that's processing the input tokens. When it's then reasoning and then outputting the answer.

SPEAKER_01 21:29

Oh, I say you break them into the two.

SPEAKER_00 21:31

And they're priced at different points. So fables just been released, it's ten dollars per million input, fifty dollars per million output, right? Wow. Which is potentially pretty pricey pretty quickly, right? Especially if you've got it on the code.

SPEAKER_01 21:45

So it's like you can feed as much as you like in at slower cost. But I see what you're saying. So if you feed it your own written dossier or build me a business plan based on this or it ingests it and it figures out what that is, there's going to be a cost associated with that.

SPEAKER_00 21:58

And then when it's about the

SPEAKER_01 22:00

And then it outputs it. That it charges you a bit. Because it's smarter than you, so therefore.

SPEAKER_00 22:04

Exactly. So there's that's where the value is, right?

SPEAKER_01 22:06

That's a charge more for its own thinking. It's actually valuing my thinking, exactly. And not highly. Yeah, exactly.

SPEAKER_00 22:13

I now know where I stand in this relationship. That's fascinating. But the that that's the token basis, right? So that's where the billing works there. And if you think about it, there is an economic value associated with one of those tokens, and it's a bit like a unit of thought or like a a way of attaching a dollar value to a unit of intelligence, right? Which you could probably then make as an argument, a pretty compelling argument, that it's a utility. Like in some way you can point intelligence at things and then there is a cost associated with that.

SPEAKER_01 22:46

And you're saying it's a lot more it's financially better to try and get into token economics versus the first time.

SPEAKER_00 22:52

Um I mean over time it's yes. Yeah, if you if you're margins higher. If you run the numbers.

SPEAKER_01 23:00

So what's the difference between between producing tokens versus just the the hardware element to it?

SPEAKER_00 23:06

Like a So the models are gonna be running on the hardware, right? And then you can run that model on the hardware, you can rent out the GPU, and that's gonna be I pay X amount per unit of time for this time-bound bucket of math that I'm able to do, right? Because this GPU can do X amount of computation, but like computation in and of itself doesn't necessarily have any economic utility, right? But then if you're running a model and that model is intelligent and there is economic value attached to the output of that model, right? The computation then also has value, and then the tokens are the result of that computation, and then there is value that's associated with that. So you're further up the stack, yeah. And because you're further, it's like um refining crude oil into uh more valuable products, right? It's a similar kind of idea.

SPEAKER_01 23:56

And so if you think, all right, I've got I've got my platform. Um we well, we actually had model as a service. We actually launched it, we ended up selling our GPUs, so we ended up not we ended up our customers ended up preventing us from launching our own product, um, which we which we will bring back online shortly. Um yeah, but it's I guess how does it if if you want to run the model, so let's say I don't know, we pick a model, it's I and I assume it's not open AIs and Anthropics model.

SPEAKER_00 24:23

Um so it could be, I mean, if you're if you're servicing Anthropic as a tenant, it's gonna be the same thing. It's just they're capturing the value from the tokens that are being generated, right? If you were to deploy Quen or something like that, which is more of the open source model, which is the open source model, then you're the one who's capturing the token economics instead.

SPEAKER_01 24:40

And then the then there's this fine-tuning I've heard. So where they optimize or they fine-tune it to make sure that the GPUs that you've got running beneath the model is, I guess, tuned to be high more performant and go in both directions, right?

SPEAKER_00 24:55

So you can you can optimize the the hardware environment so it's more efficient, your utilization is higher, which basically means that you're you're using it. What are you doing? You're actually optimizing the hardware for the particular model because it's uh so the the software and the orchestration. Okay. Right. So going down into basically the guts of the GPU and making sure that the workloads are running efficiently such that any power that you're spending running the GPU is then also running the model, right? So you're getting some economic value out of burning power effectively, right? And then the fine-tuning going the other direction, you can take an open source model off the shelf, um, and it will be a snapshot of whatever the training data was, right? So some of the older models, you'll ask it I don't know, a time-bound question, like who's the prime minister in the UK or who's the president in the US? And it'll like it will come up with like Biden or whoever or something, right? Yeah, when it was trained. Whoever it was at the time because of when it was training. Fine-tuning is a way that you can take potentially some of your company data or some data that is specific to you, and then refine what the model knows by adding more data to it. Rebeach it in effect to a certain extent, yeah. Or which means that it's then better at a given thing because it has more domain-specific knowledge and you can do that after the fact. Um, which then potentially increases the the utility of that model in that specific domain.

SPEAKER_01 26:14

Is that like a good example where a law firm might say, I want to now make train this specifically on legal blah blah blah? Yeah. Yes. Yeah, exactly.

SPEAKER_00 26:23

Um, or you can go one step further and you can start looking at like truly domain-specific models, right? And this this is one thing that's been quite interesting to watch over the course of the past really six months or so, because up until that point it's been text-based models or video

Domain Models For Science And Cyber

SPEAKER_00 26:40

models or image models, which have general utility, and these are general utility models, anyone can use them and they know a lot about everything. In the past six months, we've started seeing more Neolabs evolve or appear, right? Who are building more domain-specific models. So, Basecap Research, for example, they're a biotech company that's based in London, and they've got more genomic data than any other company on the planet because they've spent the past decade or so in this field gathering. Special data, yeah. So they have all of this data and they're training biology-specific models on genomic data. That's cool. So instead of it being just like text from the internet or whatever's been scraped from wherever, right? It's proprietary biological data on DNA, on genomic sequences, and that becomes really interesting.

SPEAKER_01 27:30

And that's uh they want that as a private model, obviously, I assume. And then and so that so who how do they figure out who to run as a private model?

SPEAKER_00 27:38

So, I mean, so they built that from the ground up. Really? Exactly. So biologically uh genomic-specific model with all of that data as the base parameters, and then train that model from the ground up to instead of it being a large language model per se, it ends up being like a large genomic model, right? Which is then trained just on biological data. Luminary have done similar with like physics-based models, um, other world models are starting to appear now as well, which are looking at physics too. So then you can get into things like material science and that sort of thing, um, which is a hell of a lot more interesting than a chatbot, right?

SPEAKER_01 28:11

That's what I guess the kind of funny thing is, all of what you just referred to as the promise of AIs, and that's what's coming in theory. Um, when are we gonna have these physics breakthroughs or these, you know, that's the everyone's like, when are we gonna solve cancer? What are we gonna do with these pieces? So um I think in in sequencing, we actually had quite a long time of okay AI use cases, like you know, the ChatGPT exam example. And it solves a problem. It's like helps me. I don't know, I've got this fridge, I need to fix something how but it steps you through it. It's probably more akin to like a Google in the past. You just took a lot of effort out of it because it would solve for it. We saw coding take off, I think, in December, and then it's just gone crazy through the roof. So the use case has actually got literally amazing, and they're solving, adding huge value to companies, it cost you a bit as well, but it's it's adding huge, tremendous value. We we see it in our business, just uh our small company, um, you've got something like 90 developers. I don't know what they've probably 10x what they can deliver, and that's going everywhere. What's the next use case that we expect to sort of see really take hold?

SPEAKER_00 29:21

You can make an argument that cybersecurity is it right with the mythos. Yeah. Um what's your view on that one? Uh I mean it it's it's good marketing. Yeah, this is free IPO, it's very good marketing, um, if I'm being cynical. Um, but I mean, other models like I say, it's all like it's it's genuine.

SPEAKER_01 29:38

Exactly.

SPEAKER_00 29:39

And again, never waste a crisis. No, of course, yeah. But I I think that you could probably make an argument that mythos is the the clawed code moment, but from a cybersecurity standpoint, right? Is that okay, these models are now good at this thing as well. Oh, they're also good at this thing, and it's gone from being just like chat to coding to now cybersecurity, and then that seems to be the cadence in that they're very good at most things, and then all of a sudden they become incredibly good at another thing um as the capability increases. And I think cyber is now. Um physics could be next, potentially. There's lots of interesting companies doing um interesting work in the the physics model space at the moment. Um, and then physics ultimately unlocks an awful lot of other stuff.

SPEAKER_01 30:22

Totally. Yeah. And then obviously, yeah, if their genomics example is surely you would think AI would be incredibly good at deciphering through thousands and thousands and thousands of pieces of data to then stitch link together the pieces of the puzzle that we're all missing to to give us, you know, the answer, or even or even produce a human-specific, like a a me specific drug that's related to all my details or whatever. Like personalized medicine.

SPEAKER_00 30:51

Personalized medicine. And drug discovery is one of the potential applications of these models. I think that's something that base camera research aren't looking at at the moment, is the applications into pharmaceuticals, drug discovery, that sort of thing, which is fascinating.

SPEAKER_01 31:03

So you started a um well, yeah, I saw your posts and then you started a a newsletter in a newsletter indeed. How often does that come out?

SPEAKER_00 31:11

Yeah, so I started the the GPU.ai um in December 24. And if it all falls apart, then that's a nice domain to be sat on to probably sell to someone at some point in the future. Um and I But you were just writing what you were saying. Exactly. So super simple format, which made it repeatable for me, but then also gave me an incentive to keep up with the market of a top story from an editorial standpoint, a market move, which could be a funding round or an acquisition, something near-cloud related, something model or software related, something hardware related, something hyperscaler related, and then something data center related. And I've done that every Saturday since the beginning of December 2024. I've not missed an issue yet. Um yeah. I think it's issue number 110 is gonna be out tomorrow. Um circa 4,100 subscribers, which is a lot in this space. Yeah, 52% open rate, six percent click-through rates, which for for niche B to B Yeah, it's amazing.

SPEAKER_01 32:07

It's pretty good.

SPEAKER_00 32:08

I mean 50% 50 people at 2%.

SPEAKER_01 32:11

I don't even know.

SPEAKER_00 32:12

I think my emails are probably at lower than two. But it's it's been really it's been an interesting way to to give myself an incentive to keep up with what's going on. Um, what motivates you to stay across it as exactly and then you fall.

SPEAKER_01 32:27

Then other people can digest what you've had, so it's helpful.

SPEAKER_00 32:29

Which form of perspective and like the way that the market's developing because I've kept up with it, and then with things like broad code, right? You build tools that allow you to do it more effectively. So I've got engines in the background that are pulling in every single press release from every single major company from across the space, running it through um some analysis and then turning it into a graph of the industry. So, like who's buying from who, who's selling to who, who's investing from who, mining filings from the SEC as well, um, or just sitting in this as right now, a very inefficient, inefficient super based database. Um but there's all kinds of interesting insights that you can you can pull up with this. Hiring signals are interesting, GPU pricing is interesting, and then ultimately it's a case of right, what does all of this mean and where is it going and why is it going in that direction? And then if you then take that data and you feed it into your basic, I say basic, um, Excel financial models, to then interrogate what's going on, right? Because every single one of those decisions is a business decision. What lever has been moved

GPU.ai Signals And Market Mapping

SPEAKER_00 33:40

in one of these financial models, and then what does that mean for the business going forward?

SPEAKER_01 33:44

And so that I think that us like business models were always great because you'd say, Oh, there's a five-year plan, we're gonna execute the three years, and then but it's almost like every month it's changing. So what um, yeah, so I think it's a super insightful for people to try and pivot. I I said to someone the other day, I said they were like, What who do you think is the winners? I think the winners are gonna be the folks who can pivot the fastest. Because even if you're a winner today, it can change. What you need to do is pivot very rapidly to wherever the puck's moving as opposed to like, you know, where it was. Um and it's that's why it lends itself well to a lot of faster moving companies as opposed to a monolithic sort of giant company that's just by the time they've thought to do something, it's already moved and the strategy, even they're chasing the wrong thing all of a sudden. We've seen it just change so rapidly, just in the compute space alone, um, just in the last three or four months.

SPEAKER_00 34:36

So I mean, yeah, because your business has changed quite a bit in the past three to four months from the the latitude acquisition to the storage side as well. Yeah, right. Like why? What what was the the impetus on your side to move more from the networking side to then to see the opportunity? Well, how did you see the opportunity?

SPEAKER_01 34:54

We had I like I had this, I mean, we've been following Jensen for a long time, I had this belief that inference was going to come at some point. Albeit it took quite a lot longer than I thought, weirdly. Things move fast, but just um it's training has been the use for so long. And then the inference, I suppose, was only going to a few companies, and I didn't think it would flow I thought it would flow through sooner. I actually thought Enterprise would have adopted it sooner. I thought a few more pieces of that puzzle would have happened quicker. Probably should have thought through that and um that's I think a lot of people thought I thought the same thought at the time. Because it's actually been a while. It's been I mean, it's fast and it's slow at the same time. It's like watching full self-driving on Tesla. When it first came out, I'm like, oh, it's I'm like, my kids are never going to have a driver's license. As we're approaching that, I'm like, well, maybe they're gonna have to get a driver's license. So I think I probably predicted one piece of it, that well, for us, what we do, we spend our lives connecting companies from data centers to clouds, between clouds and between data centers and all the other elements. So internet and IXs and Internet exchanges, and then we sort of followed the bouncy ball of products. So we keep adding and innovating products along the way to solve our customer problems. Everything that we did was built from an automation standpoint. So we automate physical infrastructure in data centers. So if you actually boil us back to what we really are, we physically we automate physical infrastructure in data centers. That's kind of what we do. So if you actually have that lens, from a network standpoint, it's Cisco switches and uh firewalls and a whole range of um network infrastructure, then there's lots of physical fiber, and then there's racks and cables and what have you in a data center. What we realized was what we were doing is connecting a customer from data center A to data center B, but really it was a server in data center A to a server in data center B, or a server in data center A to a server in the cloud. Um and so one of our customers was latitude.sh, and they were pumping all these connectivity connections with us, doing a huge amount of um um network capacity. And I was fascinated by what they were doing, ended up having this conversation, realized that what they'd built is an incredible automation platform for physical infrastructure and data centers. It just so happened that their physical infrastructure was compute uh and GPU. And then I I thought, well, if we brought the two companies together, we we have the trifecta, which is compute network and storage. Well, I needed to add the storage element. So we acquired the compute business, which has CPU and GPU. As part of that acquisition, I was like, you need to add storage, which we just launched last week. So now you have super high performance storage, low cost object, um, you've got object, file, and block, super um super high performance, and then you've got low cost, like $8.49 for a terabit, terabyte, and um and free free egress, no costs to get access to it, and a hundred gig connection across our platform. So you kind of stitch the the value that we have from the network to these pieces of the puzzle that these customers are trying to solve. So that it's like we had this vision around it, and then I thought, all right, well, we will build it over time. We even had sort of a a business plan that would say we even had earnouts for the um the acquisition, which was sort of over a three-year period, and all these factors built in. And we thought this aggressive growth if we landed here, and and in the last two months we six X'd the size of the revenue.

SPEAKER_00 38:09

We hit the three-year goal. Now what?

SPEAKER_01 38:11

We hit the three-year goal in like Yeah. Days, uh which then throws every- You're like, oh, now what we need to rethink um how big we want to go with that, which is where we've gotten a lot of partnerships with uh NVIDIA to say, all right, well, let's think more broadly around how can we scale NVIDIA chips globally. Inference cloud became that piece. So um yeah. We we we we're lucky also that we're a publicly traded company, and so you have instant access to capital through the public markets as opposed to a private company that probably needs to raise that in a different fashion.

SPEAKER_00 38:47

Like a project basis, and then there's the equity to do it all. And then how do you raise the equity? Is there a parent company? Yeah, it's a it's a whole thing.

SPEAKER_01 38:55

Yes. Well that uh and last piece probably before we go, is um yeah, I mean, so for us it's been awesome, but we we now need to execute. So we're we're hustling fast and hard. We need access to chips we which we which through our partnerships we can get there because we actually have we have thousands, like 11,000 CPUs under management as well. Um and we buy them from the the same companies that we buy that you buy NVIDIA from. Like NVIDIA is through the Dells, the Supermicros, and all these sorts of cats. Um anyway, so it's sort of aligned beautifully for the two companies coming together. Like it was like perfect place, perfect time, and we'll just call it incredible vision from my side. No, I think I think we had this vision, but it the it the wave came in faster and a lot bigger than we we thought. And the good news is we had a big enough surfboard to to catch it. Um yeah, but the last one is yeah, on the on the on the the finance side, because we've seen lots of companies appear. I caught the finance bros have entered the tech space, and you you know when celebrities are building data centers, that's the first part. And then it's a top signal.

SPEAKER_00 39:58

Hey, we've we've seen this one before, and it's a top signal.

SPEAKER_01 40:01

Yeah. And then the the the next one was like, I guess you and you said it before, it's like a lot of this crypto bros as well have are in this space. So they kind of ended up in the right place at the right time because they ended up with power.

SPEAKER_00 40:15

Yeah.

SPEAKER_01 40:16

But did they really know what they were doing from a product side or tech side? I don't know. Or if it's sort of or they need other people to manage it for them, or they don't it doesn't matter actually at the end of the day.

SPEAKER_00 40:25

It's just I I think it depends, right? Because you you've got tiers within crypto. So you've got the I mean Crusoe and Core Weaver's great examples, right? So the Crusoe was Bitcoin mining when we flared gas, core we from a background standpoint, um energy hedge funds, right? And then crypto then ultimately was an interesting way of engaging those various skill sets, and then they've done incredibly well off the back of it because they tend to be financial first technical set. Totally.

SPEAKER_01 40:54

And and and probably the biggest challenge is the financial company. Exactly.

SPEAKER_00 40:58

So if you're coming at things with the financial loads and then you learn the technical, it's okay. But then if you're if you're technical and then let's say you're from the crypto space and you're more of a a shit coin aficionado, for example, but you've been in the right place at the right time a couple of times, which means that you've got some liquidity that you probably have absolutely no business having, um, then things like a coin that way to the moon, yeah. Then you're probably going to find things a little bit more challenging than other. I guess you can pay your way through it, you can go find it. But that's exactly it, right? And then you can just hire your way into the game. Yes, yes, or you can acquire your way into the game. And there's there's nothing wrong with that, because at the end of the day, right, like it doesn't really matter how you get there, yeah, right. It's just are you able to manage the execution risk, the financial risk, and whatever other risks are invariably

Cooling Constraints And Final Takeaways

SPEAKER_00 41:50

going to present themselves when you're trying to deploy hundreds of millions of dollars worth of hardware in multiple places at the same time with a technology that ultimately no one really knows where it's going to end up. We just need some GPUs. Yes.

SPEAKER_01 42:05

That's the other piece to it. Yeah, yeah. It's harder. Well, for us, I mean, compute cyber is easy because we've it's been going for 40 years, you know, a CPU, and you could always say it's seven-year lifetime, probably run it for 15 years. In some instances, if you really wanted to burn the thing to death, put a VM across it. GPUs are different. The the weird part is the everyone's everyone's been wrong. And a great example is H-100s. We were like, oh, that, you know, down, no one needs them, five years they're dead, three years that we don't even know if there's any useful life beyond that. What we're finding is H-100s, they don't even make them any more and more expensive than they were when you had bought them in the first place. Um, as you said, prices have gone up in on them. Um so yeah, and and they're they're lasting. And what's happening is we're also seeing them last longer. The actual drives in them, like literally hard drives, are failing before, and so you can replace those. So I guess we that is different to what everyone was thinking as well. So it's it keeps changing. Yeah, it's like the business plan is so hard to have. So that's why I say you've just got to be able to pivot really quickly to whatever makes sense and make sure you cover off your financial commitments in a period of time either like in the life of the contract, whatever it may be. Um and hopefully your life of the contract's not a huge, long, hugely long contract. So yeah, interesting.

SPEAKER_00 43:24

And on the data sense side as well, right? Like the the days of the 15-year one capex cycle lease are dead. Yeah, right. Because it's what 150 kilowatt racks now, 250 kilowatt racks soon, 600 kilowatt racks, not long after that. Yeah. Megawatt, yeah, two, five, ten, thirty, whatever it's going to be, I would do that, right? And like so much. Your single capex cycle data center is a very good point.

SPEAKER_01 43:52

Um I mean we we we we find it, we have customers asking us for different pieces. The hardest thing I think to find is liquid cooled DC at the moment, because every we can find 40 to 30 kilowatt racks, air-cooled uh, you know, uh enough. Um but yeah, if you want uh just a couple of liquid-cooled racks, it's very difficult. And the cost to go and retrofit or put put some infrastructure in to deal with it is just astronomical. So you end up in this like actually doesn't make any economic sense. And so what you're seeing is all these companies that just built data centers for their whole life are actually like, you're right, it's like actually it's different now. It really is different, and the density is so different. We've got data centers that just have this huge land, a few that had a few megawatts, which used to fill fill the whole thing. Now you just got like three racks in the all this white space and then like three racks right in the middle of the room. Exactly. So, how do people follow you? How would they um subscribe or or or what have you? I follow you on LinkedIn, I get the the update every week. It actually is super helpful for me.

SPEAKER_00 44:55

LinkedIn is the uh is the best and only social media platform I'm using at the moment. Um because I haven't had time to learn how to use any of the other ones effectively, more than anything. Um how does the newsletter? The newsletters on LinkedIn. Yes, the newsletters on LinkedIn, it goes out there, and then you can also subscribe at thegpu.ai or lowercase or one word. What's the other one? Stack or something there. Substack. Substack. I use Substack for a bit, but they take too much of the revenue. Oh, okay. But don't take any of the revenue. Gotcha. And they've got a nice MCP connection, which makes my life a little bit easier as well. Perfect. Oh, you can punch straight out of your. You're gonna have to make sure they put my own Beehive affiliate link in the show description as well if you want to sign up.

SPEAKER_01 45:33

Perfect, yeah. Please, we'll put that in. All right, so subscribe, check it out. And what about your business? You also um what what what should people contact you for?

SPEAKER_00 45:42

So LinkedIn is the best place to do it. Um, or reply to the email, replies to the newsletter.

SPEAKER_01 45:47

And who would contact you? What are you solving for?

SPEAKER_00 45:50

So I'm doing a lot of advisory work with investors, with data centers that are looking at this space and thinking, hang on a moment, we've got a nice strong balance sheet. Yeah. Why are we not engaging with this in a more meaningful way? Um, investors who are looking at this space, potentially looking at funding GPU deals, but they want a sanity check on the numbers, they want some support with due diligence. Um, I've started working with some newer players who are looking at the market as well. So if you want some advice on how this market operates, what's going on in the market, why things are happening the way that they are, um, what the potential returns can look like, what a business model looks like, that's where I'm playing.

SPEAKER_01 46:30

Got it at the moment. And you're seeing it all. Oh, that's awesome. Well, I appreciate the time. Great to see you 3D as opposed to just assuming your content quietly.

SPEAKER_00 46:39

Yeah, playing a real person. I am not I am not AI. So yeah.

SPEAKER_01 46:43

That's right. That's awesome. Fuss, real pleasure. Thanks. Pleasure.

unknown 46:47

Thank you for that time.

SPEAKER_00 46:47

Good to be here. Thank you very much.

SPEAKER_01 46:50

Thanks for joining us on Uplink. See you next time.