Hello and welcome to Onward. My guest today is Curving Play, who was the CTO of Cisco's automation group. Curving has been building data centers for over 25 years on nearly every continent in the world. He has a rare combination of technical and operational knowledge. Before we get started, I want to remind you that this podcast is not investment advice that is intended for informational and entertainment purposes only. Thank you. Thanks for having me. So I'm excited to have you because data centers are a combination of the two things we focus on, which is tech and real estate, and here you are an expert on data centers. So let's just briefly, before we get into it, why are data centers a big deal? How much demand, how big could they get? If you look at just the rise of AI and where it's gone and the new use cases that are coming out for the use of AI, all of those depend on a data center somewhere. So it's almost impossible for us to get to the call it the promise of AI without having these data centers to be able to consume, produce, and modify all of the data that's coming out of AI so that it's consumable by end users, by businesses, by everybody that needs to get access to it. So what that means is like the internet, it started small, and it got to the point where it was ubiquitous. And ubiquity means that it has to be everywhere, it has to be accessible, and at some point it becomes almost like a utility. And that's where I think data centers are going. And if you consider that as the paradigm of where we're going to end up, then it's inevitable for data centers to be everywhere, and there to be many, many, many data centers, much more than we have today. So I guess that's the overview of why I think in whichever way that you look at this data centers are going to be a super critical part of the infrastructure going forward. Well, let's put some numbers on it. So I saw a number saying something like a trillion dollars of data centers per year started by 2028. Do you have any stats on this? Yeah. So what do we know? We know what's been, I guess, publicly announced, and what's been forecasted by all of the really big companies that are building. We know that, for example, the Stargate project is multiple gigawatts, probably 10 to 20 gigawatts worth of data centers. We know that in the Middle East, they are trying to build AI hubs to encourage more AI in the region. We know that there's going to be the rise of sovereign AI. Why? Because everybody is scared of what AI will do with their data. So sovereign AI is a way for governments to protect their data, but still have access to AI. So if you consider that 200 plus countries are building data centers of their own, and you're getting hyperscalers building data centers for everybody to use, a trillion dollars is a small amount of where this thing could be. So a trillion dollars sounds like a lot. The biggest impediment is going to be the time to get to that outcome. Depending on how you build data centers, it could take anything from six months to five years. So saying that there's going to be a trillion dollars worth of data centers in 2028 means that maybe the capital is committed, but not all of that will be built and not all of that will be ready to be consumed in 2028. I did a little bit of prep, and I found that the interstate highway system built under Eisenhower cost in today's dollars a half a trillion. So you're talking about building two to five interstate highway systems just to build the underlying highway system for AI. Yeah. I think you've got to look beyond. If you look at the effect of the highways, it actually gave rise to automobiles and motels, which didn't exist before. And that's not counting in the half a trillion dollars that you're talking about. So same with data centers, there's going to be a demand for power, which is going to be probably another trillion dollars by itself. All of the equipment that goes in the data center itself, which is going to be refreshed. And if you look at the current pace of things somewhere between three to five years, that explodes as well where you're getting the infrastructure piece and you're getting all of these ancillary investments that need to happen in order to make this data center useful. So imagine that you build all of these data centers, and there's hundreds of gigawatts worth of data centers, and there's no connectivity to the people that are using it, completely useless. So fiber rollout, broadband rollout, faster Wi-Fi, where it chips on end devices. All of those things are going to require investment in order to actually get the promise of what we think this is going to be. So lots of questions now before I ask you about your background, so people understand why you're a great person to speak to. Can you just say what a gigawatt is? Like some measure, because people say gigawatt, I don't think most people know what a gigawatt really means in terms of how much power is that. You could say that a gigawatt would power California for less than a year. Okay, so that's one data center. That's one data center. I actually started, you said it was 10 to 20 of this. 10 to 20. So if you think about the average consumption using your TV, doing stuff at home, that doesn't equate to anywhere near what these data centers consume. When you look at where these data centers are built, they built close to power stations and in rural areas, typically somewhere between 10 and a thousand times more power than that community consumes or that state even consumes. So if you look at the really big data centers, they are the biggest consumers of power in that state or even in that country. If people had to picture what you would do with a gigawatt, we're talking hundreds of gigawatts in the next couple of years, it could power the earth for the next 50 to 100 years. Okay, well, we'll get to the double edge sort of that because I think there's some questions about the environment that I want to follow up on. But let's go back to, so why are you an expert on this? Have you ever built data center, Kervin? Yeah. So I've been building data centers for about 25 years now before data centers were actually cool to build. And the very first data center I built was in a container without access to power. And it had to be run by a diesel generator 24 by 7 in order to keep it running. There was in the 90s, I then worked for a company that had the largest data center footprint by square foot through the 2010s. That's not a good measure, by the way, anymore of who has the biggest data center, not by square foot because that's not how people consume it anymore. It's by actual usable IT load, which is another dynamic that we should talk about because there's a lot of wastage when we take the power from the power station and consume it in the data center, there's a bunch of wastage that happens in order to power the servers that are inside the data centers. So I've been building these data centers from before they were called data centers. I used to build them for telephony exchanges back in the 90s and 2000s before the rise of mobility and cloud and where data center overtook telephony. On one stage, data centers were built for telephony. So if you look at the largest data centers in the 90s, it's AT&T and Verizon and they had the biggest data centers because they had to support all of these voice calls, that's far overshadowed now by data usage and video usage. So I've been through the transition of telephony, mobile data, video, and now AI. And all of those things required a slightly different way of thinking about data centers, building data centers, financing data centers and imagining how these data centers are going to be in the future. In a nutshell, I've watched the rise of data centers over the past 25 years. And your last job was at Cisco, right? That's right. I was CTO at Cisco and I was responsible for automation, for service providers and subscribers exist in every country. But essentially bringing automation and AI into what was a legacy environment of service providers in the world that they used to live in, where now they've been relegated to providing bitmaps. So what does it take to build data center? Maybe you could take us through the process in the United States, if you were to say, you're going to build, I don't know, a figure watch, the right number, but you pick a large number, say, OK, you and I are going to fund it. We'll put aside the financing for a minute. You and I are going to build data center. How do we do it, Kervin? Yeah. So there's a couple of steps that you need to think about as you approaching this data center building. Probably the most important thing is you cannot have a data center without power, as you mentioned. And for simplicity, let's take 100 megawatt data center. I know you said financing aside, but that runs into a few hundreds of millions of dollars, just for perspective, to build a 100 megawatt data center. So what will you need? First you will need access to the power. Your first hurdle that you're going to come up with is when you go to a power producer and say, I want to consume 100 megawatts from your power grid, they're going to say, well, number one, we don't have that much spare capacity lying around. So we're going to have to build it for you and that may take a couple of years. Number two, depending on where your data center is, we're going to have to transmit that power through the grid and the grid cannot support that type of power. And as you know, in the United States, the grids have been aging for quite a while. And they've been multiple different, I guess, ambitions to try and modernize the grid. And you still see overland wires. You still see forest fires as a result of the aging infrastructure. So it's important that when you get this power, you also need to be able to get it to where you want it to be. That means that you're relying on the power producer to transmit that power to you via a grid of some sort to a location that you decide. Then you've got the next problem, which is, is the location that you've chosen suitable for a data center? Why? The environmental around a data center are unlike anything we've ever seen before. It generates an immense amount of heat. The floor loading is really, really high, which means that you need to worry about what's happening underground. In order to support this immense weight that you want to put on the ground at that place. And as data centers become more compact, which means that they consume more power in a smaller space, it actually makes the problem worse. So you can get smaller data centers consuming a high amount of power, but the land can't support it. So then you have to figure out how to essentially fortify the land in order to build this data center that you want to build. So now you're at the point where you've got the power. You've spent probably, I don't know, six months to a year getting the environmental approved by local authorities, or potentially national authorities in order to build this. And you're ready to put a spade into the ground, essentially. At that point, you have a couple of choices to make, which is what type of data center are you wanting to build? Now let's have a little bit of a history lesson here. When you were building a data center in the early 2000s, a 10 megawatt data center was huge. It was unheard of to have a 10 megawatt data center. Nobody thought that that would ever need more than a 10 megawatt data center in one location. They obviously didn't plan for the explosion of mobile, the explosion of cloud and explosion of AI. So we get to this point where you've got designs that were good for the 90s and good for the 2000s. And those companies are trying to essentially leverage their R&D that they had, which is 20 years old, in order to deliver something that you need today. What happens then is you end up with this huge sprawl of data center that takes up way more space than you think it does. And there's quite a few Google Maps pictures of data centers in the Midwest. And if you look at the size of those things, it's larger than farms. So now you've got essentially old technology trying to fit into a new paradigm. And then you run into the manufacturing concern, which is, how do I manufacture that amount of stuff, which requires steel, it requires aluminum, it requires cabling, it requires cooling, it requires plumbing of the scaler, which these manufacturers have never seen before. And you see that some of these manufacturers are at the point where they are trying to launch new ways of building in order to become more efficient. But essentially, if you walk into a data center today, that technology is 15 to 20 years old. Your challenge with this money that we have is trying to figure out how to make the most use of this money, how to be most efficient about it. And the second problem you're going to have is if you try to do many, many things with the data center, you end up with having to make really tough decisions of segmenting the data center to support these different workloads. So let me give you an example. A data center birth for mobile data looks very, very different in design, in format, in consumption, from something that's a hyperscale cloud data center. A mobile data center is not at peak utilization all the time. So you get peaks and troughs. And at some point, you can turn off parts of the data center, which is really, really efficient. But if you're running a hyperscale data center, typically, these things are running at 40 to 70% utilization all the time, which is a very different heating and cooling problem that you have. So you need to extract all the heat from it, which means that the design of the data center needs to change based on the fact that you're building a hyperscale data center. And then you transition to something like an AI data center. And when you're doing training, these are publicly available stats, where meta was trained in Lama for somewhere between three to six months, that's 100% utilization for six months. And that's a huge difference to a mobile data center or a mobile data data center that's using you know, 20 to 30% utilization. So these things end up with problems that you never had before. How do we extract the heat? How to build racks so that they can consume the power that you need? So to the cables overheat, what is the fire suppression that you need? It's very different to what you would need in a traditional data center. So these are design decisions that you need to make way before you even think about putting a spade in the ground. Okay. And we're not even 50% of the way there. So once you have the data center design, then you go to the power producer and say, this is how I'd like to consume the data. And typically what's going to happen is they're going to have to build an entire micro grid for you. Multiple new substations in order to deliver the power from where the power is produced to where you're going to consume the power, which means that you're running a completely separate grid or you're overloading an existing grid, but you building substations in order to get that power and consume that power. And at that point, you've got the manufacturing that starting, you've got the construction that's happening. And if you're talking about traditional data centers, it's brick and mortar. So this looks more like a real estate deal at this point where you're building buildings, it may be two, three, four stories high. Each of those floors don't need to be fortified because you've got tons and tons of equipment that need to be housed on these different floors, which means that it's not a typical, even a commercial data center. It looks more like a factory floor than a commercial real estate build. And as you can imagine from your background, it gets exponentially more expensive when you start to put more and more steel into the building in order to make sure that it can carry the weight that it needs to. So that's happening in parallel while the data center racks and the equipment are being built. And at some point, the shell is built, the power is available. And then you start to bring in all of the internal data center components that you need, typically the cooling and the power distribution first and then the racks and then the equipment. And lastly, you get the network equipment that comes in to be able to connect the data center into the internet. And that's where you end up with the final piece of this, which is how do I connect this data center to the internet and who's going to use it. And at some point, all of the data center providers are thinking about making this data center available to everybody on earth. In order for that to happen, the data center actually needs to be able to connect to the undersea cable network in order to be most effective. So somebody in Europe could use a data center in the US, for example, and vice versa. Typically, this is somewhere between a six-month to three-year project before you can even get any revenue in the door. And this is part of the problem of building data centers. In order to get revenue, you have to have a phenomenal amount of investment first before you can start to realize any revenue coming in the door. And at that point, negotiations for the off-take could take anywhere from six months to a year. And the integration of whoever's using your data center could take another three to six months. So we're talking about somewhere between four and five years from when you think about building this project to when you actually start to see revenue. And I know that one of your further questions was why aren't people making money now? And that gives some sort of clue as to why. How is an AI data center different? I mean, I know you're saying it's denser, more power, more expensive. I mean, it's a GPU versus a CPU. What does that mean in practice? Let's talk about the, call it the latest generation. And for simplicity, we'll talk about the NVDRX because it's something that everybody knows. So they produce something called an NVL72. And the NVL72 has two unique constraints. Number one, it requires somewhere around 150 to 200 kilowatts per rack, which is if you look at existing data centers, they were maximum 20 kilowatts per rack. So we're talking about five to eight times more power than they need. And secondly, they need to be water-cooled. Previously, these data center racks were air-cooled, which means that the air blew through the server. And you had things that look like household air conditioners extracting air from the data center and cooling it and pumping cool air back in. Now this has to be completely water-cooled, which means that you need to have essentially water plumbing to every single rack. Take that all the way back to the outside, heat exchange it, so cool the water back down and then pump the water back in through the server so that it can cool the server. So this is a fundamental shift in how data centers are built. So where previous data centers were built for air cooling only, a water-cooled data center is a fundamentally different design, and you see quite a few different approaches as to how to approach that problem of water cooling this AI data center. And once you have that, you've got the startup problem. These things consume call it 150 kilowatts per rack. If you try to start all of these things up at once, you literally will overload the power producer. So you cannot turn every single machine on at once, which means that if you have a brown out, you have to have a controlled shutdown and restart of these data centers. So as not to cause a brown out, essentially for the entire county, city, state, whatever may be affected by these things. That's where the problem starts, where the utilization of this, the power utilization, the cooling utilization, the designs that you need to now implement for liquid cooling looks completely different to what hyperscale has been building before. Hence, when you see all of these new data centers being built, they specifically say that they're building in AI days, and because it is a different design. So this must mean a lot more waste water. And you said wastage was a big challenge. Can you talk a little bit about not just the environmental impact, but just the practical constraints of needing that much water, having that much heat to dump into the area noise. I often describe these investments. We've invested in a few is like investing in an aircraft carrier. There are $2 billion, $3 billion, there as big as an aircraft carrier, there's complicated. And so their impact to the region is just enormous. Yeah. So to go back to a previous point, that's why environmental are so important. So what do we have to worry about? We have to worry about the amount of heat and cooling that you need out of this data center. And of course, the noise and the noises too fold. The first dimension of the noise is the machines actually running themselves creates a huge amount of noise. That means that they cannot be near residential areas, or at least they cannot be close to residential areas, because you'd hear the sum of the data center all the time. The second thing is data centers generally have a very, very high uptime. So we talk about uptime in terms of percentage uptime, and it's four nines, five nines. And if you talk about a five nines data center, it cannot be done for more than 36 minutes in a year. What that means is that you need to have backup power, and that backup power is supplied by generators. And if you're talking about a gigawatt data center, the noise that those diesel generators make in the amount of diesel generators that you'll need, it literally feels like a thousand aircrafts taking off at the center. That's the call at the air pollution part of it. Then it's access to water. You need relatively speaking clean water, which means it has to be filtered. Why do you need it filtered? Because it's going through all of these cooling components. And if the cooling components have impurities in it, it starts to build up, it stops cooling, you're losing millions of dollars, because service will blow up. Absolutely. Service will blow up. So you need to filter this water, you need to have access to clean water, which means that you can't just take any recycle water and pump it to the service in order to cool it. And then there's the air cooling piece. Typically you're running inside of a data center in degree Celsius somewhere between 18 and 22 degrees Celsius, but the equipment runs at 55 degrees Celsius, which means that the delta of 30 degrees Celsius needs to be extricated from the data center and released into the atmosphere essentially. So you are increasing the micro climate temperature around the data center, whether you like it or not, whether you're using liquid cooling or air cooling, you're increasing the temperature of the micro climate around that data center. And these have environmental concerns on wildlife, it has on flora and fauna, all of these things need to be considered when you thinking about building a data center. Yeah, I did read that some data centers are using that extra heat to heat households. I don't know how common that is. Yeah, it's not as common as you think, because realistically, the households can't consume all of the heat that comes out of the data center. So you're still going to have excess heat coming out of the data center. And also very realistically, you're not going to heat a home in spring and summer. So you've got maybe six months at best of being able to use that heat. So in most cases, given that from our previous constraint that you can't be close to a residential area, it makes it really difficult to extract that heat and pipe it to a residential area. There's one really cool project that I did work on with entity in Japan. So in Tokyo, obviously the population density of Tokyo is really, really high. And we were able to build a data center in a residential area, but we had to fortify it to reduce the noise and reduce the heat that came out of it. And we did some really innovative things there, like using plasma cooling instead of air conditioners and liquid cooling. Essentially, it was a free cooling solution, which means that it doesn't generate any noise from cooling, which significantly reduced the noise threshold that we had. How much more expensive did that make it? Probably twice as expensive. Twice as expensive. Okay. Yeah. But and you know this from your own personal experience, you want things to be really, really quick when you access services that are all should buy a data center, which means that the closer you can bring a data center to the user, the better experience they're going to have. And that's the dichotomy that we live in. We want the user to have a really quick experience, which means that we have to bring the data center as close as possible to them, but we need this immense power requirements and immense cooling requirements, which means that it cannot be close to the user. And you need to balance those things. Just to compare it to who I think of as the biggest builder of power in the world and just think of how they would be doing it, I think China builds the equivalent of the amount of power in all of United States every three years or something like that. They announce or build a cold fire plant every week, one to two times a week they build. So there's just they're the best at building at scale. And so I compare that power that you need. We need the United States needs versus China. Do you have a sense of how China is going to do this? How much more than might build? I know they're chip constrained currently because of the export controls, but how would you compare the scale that we need to build versus what China is likely to build? Most recently Elon Musk has been tweeting about solar and how much more solar China has than the United States. And I think that's the way that they meet the demand. There's a couple of dimensions to consider on this. Number one, they obviously have a lot of land that they can build solar plants on. Number two, solar technology is going to get better and better every year, which means that they don't really have to build new plants in order to get more power. You just need to upgrade the solar panels and the battery technology. So the advances in battery technology means that you can capture more of the power and store it for a longer time, which immensely improves their access to power. It also means that you can build it in any region that you want, in the most remote of places, as long as you've got access to solar. So as much as they are building coal stations, I think their competitive advantage is actually going to be solar. Again, paraphrasing Elon here. If we don't in the United States back solar as one of the sources that we use as power, we're going to be way behind China. And it's going to be extremely difficult for us to catch up. But I don't hear anybody building data centers with solar at scale. Stargates using natural gas. Yep. So the two options that we have now are natural gas and nuclear. We've heard some announcement last week about regulation around nuclear that allows smaller nuclear reactors to be built. And I think that's going to be huge in order for us to supply power in general, but supply power to data centers specifically in terms of this conversation. The advancements in nuclear are going to be big for us. And if we can leap frog, I think that's the way that we stay at least on par. But I don't think it's a one-size-fits-all. I think it's a combination of solar, natural gas, nuclear, and potentially other data sources or other power sources that are going to come online that help us to augment this power demand that we're going to have. Let me go back to the reads. You mentioned that, but there are two major public reads that are data center reads and they seem like they've underperformed Equinix last week announced that it was going to take 100% of its cash flow and spend it on catbacks, 4.5 billion. They don't seem to have much revenue growth, a few percent a year. Why are the public reads underperforming so much and at the same time data centers are the most exciting real estate investment? Great question. If you look at Equinix and Digital Reality and some of the data center focused companies out there or data center real estate companies out there, they were building the data centers of the past where the utilization was really low. And if you've got low utilization of your data center, you don't make money, essentially. So getting your utilization up from 20% to 40% to 60% is huge for those companies. Unfortunately, they have really old states that cannot be used for AI unless they completely revamp those data centers, which means they need to redesign it. And what does that mean? They need to take it offline and what does that mean? That means it doesn't produce revenue. So they are in a very difficult position. They need to essentially rebuild their existing estates to support the new AI designs. But if they choose to do that, they end up with a situation where they have to actually turn off data centers for any way up to three years before that can start to produce revenue. Contrast that with somebody building an AI data center today, where it's completely over-subscribed. And we know this from core waves public filings a couple of weeks ago. They've got I think $27 billion worth of backlog for data centers. And when those days come online, they are going to be 80 to 100% utilized, which means that it's far easier for them to recoup that investment versus a traditional data center that's running at 20, 30, 40, 50% utilization. So that's, in my opinion, the primary reason why this is hard. But you cannot do it as an independent operator. So trying to build one data center, it's going to be really difficult to recoup your money. It has to be in combination with a company that has a lot of other estates so that users can move those workloads around and use your data center for very specific workloads. If they need to, those are things that make the offering more compelling. Unfortunately, you need these older data centers that are able to support the old workloads, the hyperscale data centers that are supporting cloud workloads and the new AI data centers that are going to be supporting AI workloads so that you can essentially bring to market a complete offer. So you're saying one company needs all three of those things. And other than the hyperscalers, which is to be clear are Amazon, Google, I mean, arguably meta, Microsoft, and then Oracle. Those are the five I think of, does anybody else have that scale? In terms of footprint, certainly digital really equinix and entity have the scale to do that. They've got the footprint in multiple countries, in multiple regions in those countries to be able to do that. But again, they had to build data centers, and I say they are there at the time, we had to build data centers to support many, many, many different types of workloads versus a hyperscaler that supports a far fewer subset of workloads. So Google, when they initially built their data centers, they just had search. They could hyper optimize their data centers for search. If you look at digital really equinix entity, they had to build data centers to support very, very different types of workloads. A small enterprise that's hosting a single server to back up their user laptops versus running an entire SAP instance versus a company like Adobe running all of their cloud applications in some data center. So every one of those workloads are in an equinix data center, and they have to be able to support this. It's difficult to optimize for any one type of workload, which further complicates the ability to extract revenue. Do they have the power also at these states? You're talking about so much more power. Typically, they don't. When you want to consume 100 megawatts plus, these existing data centers were not built for that. So you're going to have to get more power, build more infrastructure, more grid infrastructure, more substation infrastructure, then build essentially all new wings of data centers in order to be able to support these new workloads while trying to not disrupt your existing revenue history. So I was with one of the big private equity shops right now, you know, Blackstone, Starwood, a few others are putting out, like when they're building tens of billions of dollars data centers, and he was worried that they were overbuilding. So the tech industry, because I've sit across both, feels like there's infinite demand, and the real estate industry is worried about it being overbilled. There's a new one gigawatt announced every week. And so what would you say to him? The underlying infrastructure that you're going to build is always going to be used. So what does that mean? It means the brick and mortar, the power, the cooling, there's always going to be a use for it. The problem is that the equipment inside the data center, at the rate that technology is evolving right now, it means that every three years or less, the equipment that you have in it becomes obsolete. So in videos, last evolution, the H100s, basically nobody wants to buy them anymore. They all want Grace Backwell 200 and the upcoming Grace Backwell 300, which means that all of the other stuff becomes obsolete. So if you're a full stack provider, in other words, you own the GPUs and the equipment inside the data center, you've got to get your utilization up really, really high to stay ahead of that curve. If you fall behind the curve of utilization, it's a very difficult business to make money. Because I was looking at that idea, let's say you were at $100 million of GPUs and infrastructure to support that with three year amortization and you wanting, at least to say a 20% annual return, you're talking about you spend $100 million needing $35, $40 million of current cash flow, just to cover amortization plus your return and investment. I can't believe you could put out $100 million and get a 35% return on cost, really can't exist, does it? Not in five years. So when we model data centers, you model it for 20 to 30 years. You don't model for five years. Nobody models a data center for five years. But you just said the equipment inside is absolutely within absolutely three years. So this is a great segue to talk about the different types of models that exist out there for data center providers. They are data center providers that essentially lease out power shelves. So the brick and mortar the power and the cooling and they say, bring whatever you want. You take the risk of the stuff inside it going obsolete. I will sell you essentially power on a per kilowatt hour basis. So you can do whatever you want. You can put whatever stuff you want in it. You are responsible for all the stuff inside. So if it goes obsolete, that's your problem, not mine. And that was the business of a digital reality and an Equinix, right? That's what they did. Then you get cloud providers that are moving up the stack to pass SAS and IAS. They're saying we will take responsibility for this infrastructure inside the data center. We also take responsibility for keeping it current. And the way they made money was to be able to charge a really high margin on the services and the software. So in other words, if you're not providing the software on top of the infrastructure, you are not making money. Hence, the hyperscale is make money because they sell software, just selling the server or renting the server out and trying to amortize that over the lifetime of it. Very difficult business, which is why you see very few providers actually doing that. A lot of the Neo data centers, which are the name given to all the AI data centers now, they are offering a full stack service, which means that you can rent a piece of a GPU for a fixed amount of time, whether that's minutes hours, days, months, whatever it is. And that's where the margin comes in. What does that mean though? It means that you need people to be able to manage that data center, manage the hardware, manage the software to the maintenance. It becomes more like translating this into real estate terms, more like a managed apartment where you are responsible for everything in that apartment and the tenant just moves in and uses the infrastructure and when they done, they move out. So it's more akin to that. So you don't make money off the brick and mortar itself, you're making money off the services that you're providing. We have an AI application we've been working on at our company and we're out there trying to get tokens we want. And so if we're working with a bunch of the big model providers working with Amazon bedrock, we're working with Google's Vertex. You have to shop across all of them. You get the tokens you need. But I never see core weave or anybody else who's providing GPUs. So are these guys rapping core weave? If you're selling software, because I'm an end consumer of tokens, which is how people meet their AI, where is the intermediary that I'm missing if I'm the end consumer too? So Google is probably a bad example because they own everything. Let's take a company like publicity or cursor AI. They provide you the service and the service is AI search or it's AI code assistance. Okay. Behind that, they are consuming core weave. So they are buying the tokens from core weave. No, they are buying the utilization per hour of a GPU from core weave. They're not buying tokens. They're buying the ability to use the GPU. You think perplexity is buying GPU time from core weave. They're not buying tokens from open AI or Anthropic? No. They do a combination of everything. So they need to run their own models. They need to consume Anthropic. They need to consume Open AI. So they do a hybrid approach of all of those. But there are not that many companies at scale that would need their own models plus Open AI plus Anthropic. I feel like maybe there's 10, you know, cursor, bold, maybe not even bold. Maybe they're just literally a wrap of Open AI. So how does core weave get to 35 billion or 27 billion of backlog when there's so few AI applications that are consuming the scale you're describing? A company like Open AI will consume from core weave. They will consume any data center that they can get their hands on that's not owned by the hyperscalus. Okay, that makes more sense to me. Anthropic 2 or is that most? Yes, Anthropic as well. Okay. Now what about this question of decentralized for centralized because so many of the problems you described is because the centralization makes the data centers massive, massively expensive, massively power-consuming heat, water, etc. Can you decentralize it or does that really break down the efficiencies? There's a couple of generations of technology and enhancement that we need to go through in order to get efficient decentralization. So let's describe the outcome first and then we'll come back to the problems. The outcome is that everybody on earth in the fullness of time has access to their own model, train on their own data that only they have access to and nobody else can use. That means that there's 8 billion models out there. And many of those for the most part are exact copies of other models. There's a few pieces of the model that are trained on your data that look different to my data. So they look like copies of each other. If you take this to the logical extreme, it makes sense that your model can be smaller and it can run closer to you on a smaller data center. But the problem with that is the way that the models are built today, it requires a huge amount of memory and it requires the ability to move that memory between GPUs in order to make the tokens come out quick enough. And if you don't have that and that's where at least one of Nvidia's modes, if you don't have that, it makes the user experience really bad. So it'll work, but it's really bad. And for the listeners out there who have tried to run an AR model on the laptop, it works. It's just really slow. And over time, it's not something that you want to do on a daily basis. You'll do it for a oneself, but you don't want to do it for every task because it takes so long for even your Mac M4 to produce these tokens on the newest models that it just doesn't make sense. It's cheaper to buy the tokens for a hundred bucks and use it for a month, two months, three months. The one dimension is that models need to change. And I think fundamentally where we are now is not the endpoint of where we need to be for AI to be non-illucinating and efficient. So that's number one. Number two, it needs to be able to run on less memory, which means smaller models that can retain more information, which necessarily requires a different architecture of the AI model. And number three, the GPUs today consume an immense amount of power. And there has to be a few more steps of how do we consume less power, but have the same outcome. If those things are realized, then you get to a point where you can efficiently decentralize these models so that everybody on Earth, eight billion of us can each have our own model running on our own data center as close as possible to ourselves. What does that mean? It doesn't necessarily mean that you're going to run it at your laptop, but it may mean that at the end of your street, there's a call at a mini data center that's hosting the models for everybody in your postcode, a user code. How many years are we are you talking about here? Like that seems like pretty radical re-architecture of how AI's works today. You would think that, and I think the future is here. It's just not evenly distributed. So there's a couple of companies that are developing ASICs that allow you to do that today. Low power utilization, run your own model in a relatively small form-factor device, but there's a couple of problems with them, which means that when you develop an ASIC, you freeze the technology in it for the fullness of time. And the way that technology is evolving right now, which means those things could become obsolete in six months, in a year, in two years. So the economics don't work out for that. When we get to a point where we are, for example, cloud computing, where everybody knows everything, almost for the most part, then it's easy to do that, and that's where cloud computing is going now with edge data centers. They're moving cloud workloads as close to the user as possible, building these edge data centers that consume 10 kilowatts to 20 kilowatts, but are able to support thousands, if not tens of thousands of people. So that's happened already with cloud data centers that will happen with AI data centers. My prediction is somewhere more than five years less than 10 years. Let me just go international for a minute. I'm going to come back to home. So the Middle East has been announcing huge data centers. We're going to ship the millions GPUs there. What's their strategy and what do you think about this strategy of trying to build data centers in the Middle East? Well, I think every country and every region should be thinking about building their own data centers for AI anyway. Why? Because southern AI is actually an opportunity and a way to control your future. Think about this and we've experienced this literally in the last year or so. Open AI comes out of the new model and it turns out the new model doesn't work as well as the old model. If you are tied into that ecosystem, you're going backwards relative to everybody else. But if you control your own destiny, which means that you have your own data in your own data center and you can choose what to run and when to run it, it gives you the control that you need. So I think that in order to foster innovation, you need to have these regional data centers that are specific to a region or a country or some geographically separated entity so that you can foster innovation in that location because the other problem is brain drain. Everybody wants to go with a GPU user and you need to keep them local. And then the second thing is from a government perspective, which is why southern AI is so important, there's regulations and laws that don't allow you to move your data or the citizens data out of your borders, which means that if you want to use AI, you have to build an AI data center. I can see how that plays out in a way you said 200 countries in the world, but it seems like the Middle East strategy goes beyond that. They're trying to become a hub for a world AI in terms of the scale of what they've been announcing. I've heard a lot of people think that's smart. The oil doesn't go forever or they can become the source of where people build AI because they have a lot of power and they don't have the same regulatory challenges the OS has. I think last time we spoke, you had some skepticism about it, wondering if that's something you could expand on. It's also to do a data sovereignty. You don't want your data as a individual user or as a country to be hosted in a region that you have no control over. So us in the US, 99.9% probability that we won't run any government workloads in the Middle East AI data center. Maybe even more than that. Why? We don't want a security issue to cause the leakage of private data to users that shouldn't have them. And if we're thinking about that, then every other country in the world is thinking about them as well. So what does that leave? It leaves public common open services that can be consumed from a essentially worldwide AI data center. And if you think about back to my previous comment on everybody having their own AI with their own data, you want to know where that is. So I'm not convinced that just building a 10 gigawatt data center means that the whole world is going to use it. It's not just about the power and the compute, it's about what the AI does and what data goes into the AI and who has access to that data. You could imagine any private company in the United States saying we better host it here. So our customers, our US customers, our US consumers, you know, don't have risk beyond US sovereignty. 100% and it may be even more local than that. And I'm in California. California may say, well, we want our own AI cluster because we don't want to share California data with any of the other states. Oh god. Sounds like an interstate commerce litigation to me. I've also heard you say some things about semi-analysis, which is a leading thinker in this area, but when it comes to hardware and so it's curious what you think they often get wrong because they're fabulous analysts, but you're an operator. So what does an operator know that an analyst doesn't? I think fundamentally the experience of operating a data center is very different to the analysis of why you should build a data center and how to make money off a data center. On the face of it, you could build a data center with AI servers that has exactly the same capacity as another AI data center. The analysis still holds true, but the type of equipment that you're using, the failure rates of the equipment, the combination of the failure rates of the individual components is not something you can model in an analysis. And that only comes from operational experience. So in other words, do the econ's fail more than once a month? What is the effect of that on your design of power? Therefore, what is the effect of the design on power on your cost? Therefore, what is the design on your backup and redundancy strategy? How many more generators will you need? If you have more generators and more air conditionings and more downtime, how many more people will you need? How many more people are you going to need to be able to operate a data center? How much more is that going to cost you? That all factors into the real operational cost of a data center. So the build of a data center, they get really, really good information and there's plenty of information out there. The operations of a data center, generally, the data center operator is not exposing those costs. How many times the servers are failing? How many times they have to replace equipment? How many times the equipment goes down? How many more shifts you need to have? And again, for 36 minutes of downtime, you're probably going to have three shifts of people. So you think you need 100 people, you actually need 300 people. When I'm invested in data centers, it's like going into the FBI, yeah, yeah, you sign at confidentiality, you're not supposed to disclose anything, it's high security. So I can imagine that there's not a lot of data out there. No, and that's part of your mode of if you're an efficient operator of data centers, you can make money. If you're an inefficient operator of data centers, you do not make money. So let's relate this to an investment strategy. I could say, well, I've got $100 million. Let's give it to someone to build a data center. Only to find out that it costs a billion dollars to operate that data center because they're being super inefficient about how they operate it. And the problem is once I've invested the 100 million and I have this asset, I have to operate it. So now I'm committed for the billion, even though I didn't know that that was going to be the case. And I have no option at this point because I've invested $100 million dollars, what do I do to a right of that $100 million dollars or I just keep operating this versus someone that has figured out operations of a data center has chosen the right equipment so that the meantime between the failures is minimized based on the equipment that they've chosen and maybe it's a $150 million dollars data center, not a $100 million data center. But the uptime is higher, the utilization is higher, the maintenance costs are lower and now you need $50 million to operate that data center. So these operational costs could really kill you. It really kills an investment. And sometimes you don't know what you don't know. You start and certainly you can count on probably one hand the number of people that have built more than a gigawatt worth of data center. If you think somebody that knows how to build a 10 megawatt data center and knows how to run a 1 gigawatt data center, fundamentally different skill sets required. Is that the biggest misconception in the data center business or is there something else? No, that's probably one of the biggest I guess unknowns of how much effort it actually takes to run a data center. It's not like a commercial building where you turn it on and the tenant does basically most of the work. It's a 24 by 7365 operation where three shifts of people all the time, people on call, nobody gets to go on holiday. It's a tough business to run a data center. And if you're not prepared for what it takes to run a data center, you come shot. And then you get consolidation, which is what happened. So the reason why digital reality and economics are the size that they are now is because they've been acquiring. They've been acquiring the smaller data centers that couldn't operate as efficiently at the small scale, where digital reality and economics could at a larger scale. Well, this was fabulous. I feel like I learned a lot. Do you have any parting predictions over the next few years in terms of how AI plays out? Right now, we are in the stage of everybody's training a model, at least everybody that has the capital to do it is training a model. We're getting to the point where it's going to be a flip toward more influencing than training. So now I think we're at the stage where there's a lot of training happening and probably less influencing, but we're going to get to a stage where the models are most stable. The technology has moved on and it changes to influencing. I mean, literally, I think it was three days ago, Google launched a project on Kaggle that gives you $150,000 if you could figure out a way to run an AI model on your phone for something useful for an end user. So if that happens and we get to the point where the AI model runs on your mobile phone, then it's all influencing from that point on. The influencing is going to dwarf training by orders of magnitude. That's both good and bad. It's a different way of consuming GPUs. It opens up the lane for people that are developing A6 to come in and develop these A6. Google specifically have been developing the small model to run on their devices and other devices that have the same hardware. That changes the dynamics for all of the GPU providers and all of the data center providers. So running it on your mobile phone means that you don't need this data center for, call it 80% of the things that you need to do on a daily basis. Like, what's my next meeting? Where did I have dinner three days ago? Those are things that can be answered on your mobile phone versus do me some deep research on all of the semi-providers out there and tell me the pros and cons of why they are leading the market. That's not something you do on your mobile phone. But that's still required. And there's going to be those use cases that require these really complex models that process for hours if not days to get your answer. So to summarise my predictions, far more inferencing than training and smaller models running closer to the user. Is that Apple's actual game plan here because they seem to have been out of the game so far? I think Apple at the stage where they can choose what they do and they can run a hybrid model with their secure AI cloud that they've got. So it makes it feel like it's running on your phone, but it's not. Which is great because they own the ecosystem. So when you type in something to your phone, you may think it's running on your phone. It could be running in their secure cloud, but it's not running in a public AI data center somewhere. You still get the speedy response. And then for things that need to be done in your phone, like email summarisation, that can be done on the phone. So I think it's a hybrid. Apple and Google are probably the best place to take advantage of those advances that are coming. But again, there's going to be so many more use cases. And the big thing is if you look at how quickly video has evolved AI video, it's clear to me that AI video is going to dominate the space in the next months, if not years. It's going to absolutely dominate the space. It's going to dominate the usage of models of GPUs, of devices, creating video. Yeah, my brother's an entertainment business and he just hates it. I think it's great, but it's not so good if you're a writer in Los Angeles to wrap that up. I think the way that we're going to see AI really being used is as a superpower to people that have the knowledge. There's a lot of rhetoric out there that says anybody can build anything with AI, even if you have no experience. But the reality is that those people are going to get a 20 to 30% improvement on what they were doing because they don't understand the intricate details of how to make a movie. But people like your brother that are in the entertainment industry, it's going to give them a 300% boost. I keep telling him that, so we'll see. Yeah, so I think that's how we see this thing play out. The people that have the knowledge and know how to direct AI are really going to make a huge difference. That's already how technology has been asymmetric. It makes the top 10% 100 times more powerful and that has its own issues. But, Kevin, this has been fabulous. I really appreciate you taking the time. Thank you for having me and I hope we're able to cover enough ground so that we have people thinking about how to invest, what to look for in the market and a way to not go in the market, I guess. I certainly learned a lot, so onward. Great. Thanks, Ben. You've been listening to Onward featuring Kervin Palais, former CTO of Cisco's Automation Group. My name is Ben Miller, CEO of Funrise. When you invite you again to send your comments and questions to
[email protected]. And if you like what you've heard, rate and review us an Apple podcast and be sure to follow us wherever you listen to podcasts. For more information on Funrise sponsored investment products, including relevant legal disclaimers, check out our show notes. Thanks so much for joining me.
PodcastTranscript AI