← The Infrastructure Investor Podcast
The Infrastructure Investor Podcast
KKR: Parts of AI market running ‘a bit hot’ in one of tech’s fastest capex cycles
00:32:42
Loading summary
Transcript31 lines
- [00:03]
A
Hi, I'm Bruno Alvis, editor in chief of Infrastructure Investor, and welcome to the Infrastructure Investor podcast. In today's episode, I sit down with Valdemar Slazak, KKR's global head of digital infrastructure, to talk about AI infrastructure. We begin by considering whether there's a bubble forming around AI infrastructure and why that might not be the most useful question to ask. From there, we look at why parts of the market are overheating during one of the fastest capex cycles in tech history. We then explore the lifespan of data center infrastructure compared to the rapid obsolescence of the GPUs that power it. Along the way, we also touch on the industry's coordination tax problem, examine KKR's molecule to the rack strategy, and much more.
- [00:54]
B
Foreign.
- [01:01]
C
Welcome to the Infrastructure Investor Podcast.
- [01:04]
B
Thanks for having me. It's a real pleasure.
- [01:07]
C
We're recording this podcast. I think it's roughly a week after Nvidia announced their latest earnings. They were upbeat, but they only eased fears of an AI bubble for about five seconds, I feel. And we're really at this point where talk of an AI bubble is a bit incessant. I feel this is the wrong question to focus on. I think you've alluded to this also in some of your writings and so have others, but basically, you know, the introduction of any general purpose technology, whether it's railroad or the Internet, usually comes with bubbles attached. It comes with exuberance, overbuild winners and losers. So it feels like we would all just be better off by accepting that with AI infrastructure, it's not going to be different. So I have two questions to start us off. One, do you agree with this general assessment? And two, to the extent that there will be losers and things will go wrong, do you see people in our industry, in infrastructure, those involved in funding this AI capex getting caught up in it, and what could that look like?
- [02:17]
B
That is such a great question. And I don't really think that AI itself is in a bubble. I mean, we certainly believe that this is a long term, you know, demand for compute power, connectivity is real as anything we've seen, you know, in decades. But I think one of the points, Bruno, that you alluded to, which is this is one of the fastest CAPEX cycles in tech history and whenever things move this quickly, there clearly will be parts of the market that will run a bit hot. And so the right question, as you pointed out, isn't if AI is in a bubble, but it's where is there enthusiasm that's sort of getting ahead of itself and where do people need to stay disciplined? So we all have seen these massive data points. You know, big tech is spending hundreds of billions of dollars a quarter. AI infrastructure spend has grown 70, 80% year over year. We're seeing a tremendous amount of investment going into GPUs and at the same time the real questions that many in the industry have rightfully so is generative AI revenue is still in the tens of billions of dollars. So is that a sign that AI is overhyped a little bit? Potentially. But there's a question of a disconnect in terms of potential timing that I think I certainly worry about the investment relative to the revenue and modernization of this technology.
- [03:42]
C
Yeah. Is at the heart of it. I'm going to throw some some numbers out just to frame that part of the question just for our listeners benefit. But for example, a recent J.P. morgan report estimated that to get a 10% return on what was their modeled AI investments of $5 trillion through to 2030, that would require about $650 billion of annual revenue into perpetuity from the hyperscalers. So that's something like 58 basis points of global GDP. And this is my question to you, Valdemar. I know the hyperscalers are very profitable. They literally print money, some of them with their other activities. But if you're somebody sitting where you're sitting on the infrastructure side of things, how concerned should someone funding this AI infra capex be? A about the fact that the AI services are not generating meaningful revenue just yet, but B that they seem to have to generate an awful lot of it into the future to justify this massive capex.
- [04:42]
B
One of the things really people talk about is generative AI is a first step towards ultimate AGI. And I think the question that we ask ourselves a little bit, whether in infrastructure or through our broader franchise at kkr is whether ultimately today's frontier large language models are going to be able to evolve quickly enough to become the true agents. Which is the ultimate piece that everyone keeps talking about that you can have a digital intern or a junior analyst and associate that is effectively AI based that you can really leverage and therefore I think ultimately see that tremendous level of productivity enhancements that would justify that. The question really is how this matters for infrastructure. Is it? If you believe that we are on the verge of having fully capable automated agents two years away, I would argue that we are under building infrastructure. If you think that is two to five to 10 years away, then I think flexibility really matters. The shell power and systems that ultimately will be able to adapt across this entire evolution of GPU generations is important. How you go design and operate the cooling systems and the model architecture is important, but we are probably spending a little bit too much ahead of that. If you believe that this is really a five to ten year cycle and not a zero to three year cycle. And so I think this is the key question in the industry. So then you look at and say, okay, where do you see potential overheating in the market? The areas in infrastructure that, that I think we certainly look at and worry about is, as I mentioned at the preamble, whenever you have this level of capex increase in such a short period of time, you have to say that some of this may not be invested prudently. And so we do see parts of the system getting a bit overheated. But ultimately I don't believe that the industry as a whole is on the verge of a massive overbuild.
- [06:41]
C
Yeah, but you introduced a few elements there that I think are very interesting, especially when you start thinking about longevity of some of these assets. And if I'm summarizing you correctly, you're introducing the idea that we are on this track to AGI. There's a question of whether even large language models are capable of making that jump, that it's even bit of a wider question. But there's also the question of how close and how far away we are to it and that has an impact on the velocity of this capex spend and the fact that some of these investments might turn out wrongly.
- [07:15]
A
Right.
- [07:15]
C
Because depending on how much you believe how close we are or we aren't, then people might have a problem. And that includes people in our industry. Correct. Some of the people building the infrastructure layer.
- [07:27]
B
You're right. I mean, I think when I pull back, you know, if you think about the core markets, North Virginia, Dallas, London, Frankfurt capacity. Short. Right. I mean those are markets that ultimately are serving the core cloud capacity. You have queues that are three, six years, very low vacancies, very high rates. I don't think that there is an overheating in that market. I think we're seeing a lot of large giga scale builds which are focusing on training these large language models in locations that are incredibly remote. And even though you may be doing this for high quality customers, and you may be able to say to yourself that during the initial term of the contract, because it is a take or pay, you just alluded these are very profitable companies can, you know, effectively pay their obligations, but you are making a bet on the residual value of that asset. And the question is, if for whatever reason AI demand doesn't materialize as expected, or there is re pivoting or there is new power generation that comes in and there is a greater importance of latency, which that means location close to ultimately users, what is the renewal of those contracts at the end of those contracts? And are you able to secure enough of a return during the initial term of the contract, or are you making a bet on the residual value and the releasing of that? And I do think that in parts of that, the assumption that some are making, not all, but some are making, is that just like every other data center industry data center in the past 10 or 15 years releasing is pretty robust because there is scarcity power indefinitely and so I'll be able to release it. I think we question that premise a bit more.
- [09:13]
C
Yeah, that makes sense. And actually I want your help to try and understand something about the longevity of AI data centers in particular. Perhaps, you know, tell me if what I'm about to say makes sense. But if I understand it correctly, the GPUs make up a significant portion of the cost of a data center. Yet the shelf life of Some of these GPUs is quite short, something like five, six years max. Then you have Nvidia pumping out new and better chips every year, which is kind of faster than Moore's Law.
- [09:44]
A
So even let's say you have a
- [09:45]
C
long term contract for your data center, let's say at the end of your 10 year contract you have a box which is full of half obsolete chips, which you know, in theory should score a much lower rental value at the end of your contract compared to the start of your contract. So at worst there's a new generation of chips that might require you to refurbish your data center so you can accommodate them. What's your perspective on this? Particularly from the point of view of somebody who has to exit these assets?
- [10:16]
B
I am so glad you asked this question because I was recently listening to another podcast not as well informed. This is a more general purpose podcast and oftentimes when people refer to data centers, I think they conflate two big pieces. So maybe let's just rewind back and say what does the actually data center provide if you think about it, is ultimately connectivity to power, right? So building a substation, operating that, building the infrastructure itself, I call it the building shell and the site, and ultimately providing connectivity with many sort of diverse routes where typically data centers, I think this is pretty much universal these days. You know, end is the servers themselves, right? Which is the chips, the equipment that goes inside of it. So that's typically provided by the customers themselves. With a refresh cycle, as you alluded to, we can argue whether it's three to six years. There's been a lot of discussions around this, you know, recently with some of the accounting changes by hyperscalers to a longer life. And then there's a question unknown around the GPU useful life. It's not as much about the pure useful life, depreciable useful life because the chips burn out. It's more about the obsolescence element because the new chips that are rolled out are so much more efficient that customers would want to go and deploy them because they're getting so much more compute per unit of power use. Right. Like if we just delineate those two things, you're saying, okay, data center operator is building something that has a 20 to 40 year life horizon. Every three to five years, typically there's a refresh cycle by the customer to put in a new server equipment. Oftentimes because of the technology evolution, what ends up happening is you do have to retool some of that box and shell because your power densities are rising. You mentioned as well, power density has now gone through the roof. You know, in the last 10 to 15 years, people were building things at 8 to 10 kilowatts per rack, maybe even lower than that today. Nvidia with its latest, whether it's the H1 hundreds, H2 hundreds, or Blackwells, you're talking about 80 to maybe hundreds of kilowatts per rack. So you need to have a different power delivery system as well as a cooling system to be able to accommodate that. And, and so I think the complexity of building these assets today is much greater than in the past because you do have to think about how you future proof the design for this potential evolution of the use cases. And the biggest one is clearly moving towards liquid cooling. Second one is the power density element. And how do you think about your engineering of your electrical and mechanical equipment to be able to accommodate that? So I think those are the two different kind of pieces of that.
- [12:59]
C
Right. And just to make it clear also for the audience, and again, thinking about a typical infrastructure investor here, if the need arises to do some retooling, some refurb, et cetera, what is the division, the cost division here? Who needs to do this? And I'm asking you, by the way, with an eye towards so called stabilized data centers. And this that we hear more and more. And I just want to try and establish who has to do this if the need arises.
- [13:27]
B
Yeah, it's a really interesting question. And I think typically the way that it's being provided for is through additional, effectively non recurring charges that are being made because you're retooling the data center to provide additional capacity to the hyperscalers. And the hyperscaler signs a contract, let's just pick random numbers, right. Signs a 15 year contract, which is the take or pay contract for a certain amount of power capacity that they're paying that's on a take or pay agreement with the utility themselves. If they are coming in with a new design within the initial term of the contract, then what ultimately has to happen is there is an ongoing negotiation of again, what would be the non recurring charges that I think that are being passed through? Is there an additional expense that is being effectively provided for by the data center operator? And how is that reflective in the adjustment to the lease rate of the data center with the hyperscaler? Oftentimes, whenever that happens, there is also an extension of the term of the contract. So there's sort of a very dynamic discussion. The question that maybe you're asking is like what happens when you run to the end of the term of the contract? Right. And there is that sort of a stare down. And this touches a little bit on the question of the training model. Training data centers versus Cloud. We are very much of a high conviction in this theme. And all of our data center platforms, which includes five of them globally today, have really been operationalized around proximity to users and fungibility of use cases and operationally operating in markets where there is scarcity of power and land. So we have had a history of a track record of recontracting those assets, oftentimes at higher rates because of again there is a supply and demand imbalance. And because once you actually deploy these servers in a certain place, especially for cloud, there's an ecosystem that builds around it which makes it very difficult in terms of kind of a rip and replacement. And so if there is an end of the contract, let's just say a 10 year contract or 15 year contract, there may be a refresh cycle for data centers. And we've seen some of those happen to be able to accommodate higher power densities, et cetera, it becomes a negotiation right at that time. What also is interesting, Bruno, is that if you think about the lease rates that were signed 10 years ago and those lease rates inflating at typically somewhere between, somewhere inflation. Right. So let's just say 2%. Oftentimes there's a, there's a fixed element to that and when you look at it where they're priced relative to the market rates today, they often are very much below the market rates. You can look at that and you can say that okay, well then it becomes a stare down of is the hyperscale wanting to leave and how, how scarce is the space that you have and you're making a bet and saying well maybe I can improve the quality, the power density of the data center and others and get a 50% bump up in lease rates. Right, because the market rate is so much higher than what it had been historically. So I think it becomes just a very active portfolio management of that. And importantly, I think we should not forget data centers are in the business to provide quality customer service and being not only stewards of important infrastructure for them but also important community members as well.
- [16:48]
C
Volimar, I want to dig a bit deeper into the way KKR is going about doing this. You've alluded to many elements already of how you're going about and investing in AI infrastructure. But there is one thing you said a while ago that I wanted to dig a little bit deeper into and you mentioned something along the lines that you KKR wanted to rethink how data centers are being built, focusing on what you called molecule to the rack. And I just wanted to get a better sense of what you mean with that and what you guys really are trying to do differently there.
- [17:21]
B
What we see today in this industry is what we call it a coordination tax or problem. If you just think about any given moment in time, pick your favorite number of high quality data center operators operating across us. Let's just as a market pick your favorite number of, number of power companies, number of capital providers. Each one of them is talking to the hyperscalers, to each other and to others creates incredible opaqueness in this industry. Double, triple counting at times. And that lack of coordination from our perspective is slowing down the system. Elongating the cycle of deployment efficiency and timeline to deployment is incredibly important for hyperscalers in terms of their economics. Shortening the time from where they can light up a data center that is spending capex being unleashed where they can provide services to customers by three months, six months means hundreds of millions of dollars to them. And I think this industry is not today operating it's too siloed and operating in these verticals and there's very little coordination. What we're trying to think about it is okay, how do we position our data center platforms with our power generation platforms. How do we think about creating an efficient flywheel of communication between them? How we use our data center platforms with our relationships with utilities both on the generation side and the transmission side. We formed an incredibly, I think exciting partnership with ECP to leverage their power generation assets. And we announced this very which I believe it's the blueprint of how things will be executed. We this deployment in Bosque, Texas where we are co locating a data center with two hyperscalers taking capacity under long term adjacent to a power plant. The power plant is providing capacity to the hyperscalers and we have engineered a structure. This is an important element of how do you work with regulators, community members and hyperscalers to figure out sustainable solutions. So there's a lot of, I think rightfully so noise about grid being stretched, not enough capacity and worry about what happens in certain weather events or other events because so much capacity is being used frankly, not just for data centers, but everything that's happening in the US in terms of onshoring, reshoring, manufacturing resurgence. And so we've generated this really interesting structure where if things on the grid happen and there is effectively a high utilization of the grid, the power plant will divert electrons from powering the data centers back into the grid and we will use our on site generation capacity to provide power to the data centers itself. And this was complicated because it required negotiation and engineering with our customers to be able to accommodate that. So that is I think, a blueprint of an example of how we think about re engineering this mousetrap to create an efficiency rather than inefficiency. And a lot more of those solutions hopefully are still on the common and we're pretty proud of being able to pioneer some of that. Yeah.
- [20:30]
C
Let me try and unpack just two points that you've alluded to, which I'm going to class as efficiency and flexibility. And let me start with efficiency, because in a recent interview with us with Infrastructure Investor, you pointed out something that I found really interesting, goes to some of the points you were just alluding to. And you basically highlighted at the moment the cost per megawatt that the hyperscalers are currently paying can range anything from $12 million to $18 million for what is essentially the same compute. And of course you were arguing in the interview with us that there was an efficiency gain to be had here. Again, can you give us a sense of when you think the industry will be in a better position to capture that sort of efficiency gain?
- [21:13]
B
Yeah, it's really interesting. I mean I think the base built for data centers. Now you can sort of debate on there's nuances because some data centers may not require maybe as much backup generation or others. But I think on average the cost to build a data center per megawatt ranges from as I mentioned, 12 million to up to 18 million. Certain parts of Europe it can be on towards the higher end of that. That includes building data centers with the state of the art liquid cooling technologies. And for the longest time data center costs would be in the $12 million range because liquid cooling is so important for the latest GPU technology. And we are so early in that there has been a pretty wide range of outcomes or what we see in terms of customers. How do they think about procuring for that and engineering for that. And that could be an increment of 2 to 5 million dollars per megawatt, which is pretty large range, right? We think that because it is so early and everyone is so drinking from the fire hose in terms of procuring this capacity, the industry has not really helped standardize that and we think that that's coming. Right? So there's a little bit of a question. Clearly, if you're building a data center for someone, you could have people paying very different rates for effectively the same type of a solution. If you were just to say if I have two hyperscale data centers with two different hyperscalers in the same market, one could be paying because it's ultimately sort of you pricing it on the dollars per kilowatt relative to the capex that you're incurring, right? To some sort of a stabilized yield of, let's just say in the 9% range, right? In an unlevered basis. So if someone pays 30% more, their cost per kilowatt per month is significantly higher. It makes you think about the releasability of that asset when their competitor could be at a much lower cost per yield because the facility is not as over engineered for that. So we think that efficiency will come in the years, in the next few years. But today there's a pretty wide range in terms of that. This is why I'm so excited about this approach we're taking in terms of the molecule to rack. If you are just a bit of a cynical person, you would say, gosh, the data center industry may not necessarily have the incentive to innovate here, right? Because if I'm generating, let's just say 9% return on asset, if I spend more dollars, I get more EBITDA for that. So why would I necessarily push this innovation? This is not about the lease that's in front of you. This is about building long standing relationships and becoming embedded within your customers as a solution provider and not just a vendor. I do think that long term that's what's going to differentiate the big winners and losers in the sector is those who are working with customers to provide innovative solutions, innovate, save costs, also operate assets at quite high quality standards versus those who are just trying to extract the most value on an individual opportunity that's right in front of them.
- [24:15]
C
Yeah. On the second point that I wanted just to pick your brain quickly because the example you gave of your investment with a colocation of generation, the possibility of releasing power to the grid by using your own on site generation when needed, that to me seems to kind of fit into a larger conversation about the flexibility of these assets. That in the beginning was almost a little bit taboo. Like you couldn't talk about it. The tagline was that the Hyperscalers needed power 24 7. There was no room for flexibility. End of conversation. Now you have a FERC order actually talking about expediting connections if they there is flexibility. You have people talking about using on site batteries, other forms of generation. Is it fair to say that the conversation here is evolving and hyperscalers are opening up more to the idea of flexibility in how they go about doing business?
- [25:06]
B
Yeah, necessity is the mother of all inventions. So I think you're right, it is, it is happening. And that's why I'm so optimistic that I think despite the fact that there is a lot of noise around us being power short, I actually don't think that that's necessarily the case. We have to figure out a more efficient use of the resources than we have than necessarily going on a massive hundreds of billions of dollars of investment cycle which doesn't do anyone any good. Right. And so I'm encouraged by some of that. I'm encouraged by operators like ourselves who are thinking about innovating and I'm also encouraged by the way that the hyperscalers are thinking about this flexibility wise and innovation wise. So we are certainly not there in terms of, you know, battery storage to be able to truly utilize renewable resources to power up data centers. That's not quite there yet, but I think that evolution will happen and eventually we'll get there. I do think that this innovative solutions of how do you think about engineering the data center and the power uses? Now it gets a little complicated, Bruno, where think about providing an enterprise centric hyperscale solution. It may be an AWS or one of its peers. And if they are, if you're building a data center for them in a major urban area, they are selling that capacity to other enterprises and they have of course, their own obligations to make sure that the data center is online. None of us would tolerate if our cloud services go offline. So in some places the flexibility may be slightly lower than in other places. But I do think that one shouldn't treat all cloud equally. Right. And I think the hyperscalers themselves, and I give them credit, are thinking about this the right way of how can we be part of the solution to some of these problems and are innovating alongside with the infrastructure providers and of course regulators as well to be able to accommodate solutions that can help that.
- [27:00]
C
And Valdemar, final question for you because we've been discussing this massive AI infrastructure build out and we're talking about trillions of dollars and a lot of gigawatts to be brought online. But we are also facing, and I'm going to generalize here, but everything from supply chain problems to growing opposition in communities, you could call it NIMBYism even if you would want to affordability issues raised by the spike in electricity bills that some of the industry is helping to cause. Not the sole cause of it obviously, but put it all together and you've got a very long list of what could turn out to be some fairly formidable obstacles to the rollout of this very, very ambitious capex. How concerned are you by some of these obstacles and what they can do to slow down projects, derail them entirely, et cetera?
- [27:54]
B
It's a really interesting question. You're really. Bruno, you were just on point this morning. It's really interesting. I think of it as there is a sort of, I call it, there's a collision course of ideas, policy collision. Right. It's a one hand is to win the AI race and we need to be able to build at the speed of China, which is very difficult and at the same time protect ratepayers, which is how do you think about building 50 year assets and having 5 to 10 year tech visibility to do that. And so I think that is definitely happening. US has some challenges. We need to build something like 70,000 miles of high voltage lines in the next decade and we build at the current pace. It will take us 200 years to do that. So we do have some challenges, but I'm still optimistic. I wouldn't bet against us. I'm just an optimist right about that. So I do think that there is ways to solve this and some of the federal policies are already taking place. I mentioned earlier we can maximize grid efficiency. There are some costs that are being contemplated of what is the more equitable way of sharing some of the improvement costs. There's certainly federal fast tracking authority which is helping spearhead some of that. There are discussions around the use of some of the federal land or excess capacity on federal land to be able to build inventory. There's of course flexibility of how to think about behind the meter generation and storage generation. So I am incredibly optimistic that I think this is going to work out. Everyone has a vested interest to be able to provide this capacity and help ensure that US continues to be really the epicenter of technological innovation. But I do think that it will require bringing all of these constituents together, which is, I think you have to be able to work with the hyperscalers, you have to be able to work with the regulators and work with the local communities to engage on sustainable solutions. And a big part of this is, I think, educating people on what I think what data centers do. How can they be a contributor to local economies, how can they help solve environmental issues? And I do think that oftentimes the data center industry has not necessarily done enough of a good job to be able to advocate on its own behalf. And I think we need to do a lot more of that. I am optimistic that's the case you mentioned on the supply chain basis, I think you can get today if you go and you want to buy maybe not transformers or maybe turbines if you wanted to get those from ge, Vernova and others. But certainly you can get basic mechanical electrical equipment to be able to build data centers at scale. The big proof point to me over the next few years will be there will be a huge barbell in terms of distribution of quality operators who have approached this with prudence, who have allocated capital in a, in an intelligent way, who are able to deliver to customers and those who are maybe chewed off a little bit too much in terms of building these mega facilities and maybe struggle with delivery. So I would be watching for that in the next few years as some of these mega facilities come online. But I am an optimist and I think the best days for our industry are still in front of us in opportunities like this. It's about discipline, pattern recognition, incredible communication, and really utilizing all of the resources that you can to to stay heads.
- [31:18]
C
All right, that's a great place to end this. Thanks very much for your time. This morning. It was great having you on the show.
- [31:25]
B
Br, thank you for your time and I appreciate all the coverage of this incredibly interesting, you know, industry. Great to be part of it and participant and hopefully shed a little bit of light. Thanks for having us.
- [31:34]
C
As always, thank you.
- [31:41]
A
That again was Volodymor Slezak, KKR's global head of digital infrastructure, and I'm BrunoAlvis, editor in chief of Infrastructure Investure. We're going to be talking a lot about digital infrastructure at the Infrastructure Investor Global Summit taking place in Berlin from the 24th to the 27th of March 2026. We have an early bird deadline coming up on the 12th of December which you should definitely consider, so don't miss your chance to lock in discounted registration and join me and over 3,000 infrastructure decision makers, including 1,000 LPs from 50 countries at the iconic station Berlin. You can find a link to the Early Bird in the blurb accompanying this podcast. Finally, to hear more of our episodes, head over to infrastructureinvestor.com podcast or you can search and subscribe to the Infrastructure Investor podcast wherever you like to listen.
- [32:35]
B
Sam.