← The Infrastructure Investor Podcast
The Infrastructure Investor Podcast
DigitalBridge: Industrialisation of AI will need 500GW of compute over next decade
00:22:48
Loading summary
Transcript22 lines
- [00:04]
A
Hi, I'm Bruno Alves, editor in chief of Infrastructure Investor, and welcome to the Infrastructure Investor Podcast. In today's episode, I welcome back Mark Ganzi, chief executive of Digital Bridge, for a special episode of the Infrastructure Investor Podcast, based on a live interview Mark and I did in late March at the Infrastructure Investor Global Summit in Berlin. On stage in Berlin, Mark took on concerns about AI data center obsolescence, comparing fast GPU refresh cycles to radio upgrades on cell towers. He talked through the huge arbitrage opportunity for operators that control their own power, stressed the importance of tight contract discipline, and highlighted the hidden risks many new entrants tend to overlook. He also looked ahead to the next wave of demand, lifting the curtain on robotics and the rise of AI factories.
- [01:02]
B
Hello, everyone. It's good to see you all. Mark, welcome to our new stage. It's a pleasure to host you again and do this interview.
- [01:10]
C
Thank you. Good to be here.
- [01:11]
B
A lot changes year to year, so we have plenty to talk about and I want to get stuck in quickly. But I'd be remiss if I didn't ask you about something that happened at the end of December, which is that SoftBank announced a definitive agreement to acquire Digital Bridge. I know the transaction is not closed yet, but I want to just pick your brain a bit on what that is going to mean for you. What's changing, what is not changing? Give us a sense of how you're seeing it.
- [01:38]
C
Well, first of all, as a public company CEO, I have to do this thing every year where I have to scribe my time allocation, which is a complete waste of my time. I spend 62 days a year running a public company and I spend the rest of my time running our funds. So imagine I get two months of my life back to focus on the things that are really impactful, which is for me, working with our CEOs, working with customers, building platforms, getting back to really our entrepreneurial roots. For me, that's really exciting because that's the work that I really cherish and that I like. Can't talk specifically about the transaction itself, but I think for us, the ability to get back to being private at this moment in time is super critical. I see the alternative asset space really being two camps. There's folks that are in the arms race to go get to trillions of Aum, and there's folks that are out investing and returning capital and building returns. I think for us it was a decision, it was a fork in the road, which is my goal is not to get to a trillion of Aum My goal is to be the best builder and developer of digital infrastructure and to create the highest returns and best outcomes for LPs. So when I started the business 13 years ago, that was the intention, was to go build great companies and to work with our customers. And so the intentionality of working with Masa and going back to our roots of building infrastructure and powering the global economy for the challenges ahead, which are real and are fundamental, is really exciting. The intellectual engagement with their team in terms of what we see, what's coming down the road and what's ahead is really unique. There is a vision for the future that he has and that other people like Jensen and Masa have and Elon, that is really quite clear. Around the next leg into AI. And so we've spent the last five years building these large language models. We're moving now into the phase of inferencing and delivering those industrial use cases. But ultimately what is around the corner is robotics and ultimately the proliferation of AI into use cases that monetize because what's slow to happen is the monetization of AI. I think Wall Street's waiting for that. And that's really what's exciting for us, is to work with an industrial partner like SoftBank and to really focus very intentionally on returns. Yeah, to me, that is why we're here. That's why we're at this conference is working for LPs and delivering returns, not aggregating assets and management fees.
- [04:10]
B
Mark, I want to go into the AI infrastructure topic. You know, there's been a lot of excitement over the last few years, as we all know. I think there's, as you just alluded to, there's skepticism creeping in to do with revenue models, but also I think for the people in this room to do with obsolescence and the durability of, and I'm thinking here specifically of AI infrastructure data centers. And what I wanted to get your take on is we all know that there is a GPU refresh cycle in these boxes that's not done by you. That's the clients that have to do it. But it's fairly. It's done every three to five years. It's very costly. If you think about something like laying track, a railroad or something, you're not refreshing that every three to five years. So frame a little bit the durability of these AI infrastructure data centers. Know for an infrastructure crowd, how durable are they?
- [05:00]
C
So there's the physical infrastructure, which is the buildings, the power, the land, and then there is the IT infrastructure, which you referenced, which is happening inside the data center. I think for us, we've always felt like that owning the physical infrastructure is where you want to be, and where you don't want to be is you don't want to be in the IT stack making that bet. And so intentionally, we're not making that bet. And the analogy that I draw is if you go back to the late 90s when we were building cell towers, and you look at some of those towers that we built in the late 1990s, I started in 1994, many of those towers have gone through seven or eight iterations of radio technology upgrades, which actually follow a very similar trajectory path as a gpu. So when you think about radio standards that change every five years on a cell tower, there's an opportunity, and the opportunity is to obviously sell more fiber at the base of that tower, introduce another set of antennas, introduce another set of radio technology standards. And what that's led to is it's led to more revenue and more opportunity. The same opportunity exists in data centers, particularly as power density and compute gets tighter. And in the smaller spaces, we can free up space and we can lease to other providers in that data center. A great example of that is one of our investments in Las Vegas in Nevada, Switch. You know, if you looked at Switch has been around now almost 17 years, and that company has been growing very consistently at about 30 to 40% CAGR on its EBITDA growth. And we really haven't dramatically changed our footprint in terms of land and data centers. How have we done it? We've done it through engineering, we've done it through modern cooling, which is our new EVO self containing cooling system. And so what used to be one room with maybe 500 racks of computer, which was one megawatt, I can now do in three racks and create one megawatt. And then I've created, in essence, another 20,000 square feet of RSF raised square footage. And so now what we're able to do is release that space and put other applications into that data hall. So on one hand, technology can be expensive and our customers are constantly upgrading their GPU stack, their IT stack. But again, I don't play in that space. That's a space we don't play in. What we are in the space of doing is optimizing the experience for the customer and most importantly, creating more space where we can lease to other applications. So that's how you achieve growth and keep the same footprint.
- [07:22]
B
Okay. And when you have to do those kinds of retoolings is that contractually that's all laid out, you can pass that cost on, et cetera.
- [07:29]
C
I think for the hard physical infrastructure where we're on the hook for the cooling, we're making that decision to create that cooling experience for the customer. But they are paying for it in their lease. The lease rate for an EVO data hall and the lease rate for forced air is very different because the customer is putting in a different compute set and most importantly, they're putting in different GPUs. So how you cool an H200 versus how you cool a Blackwell's chip is very different. And so at Switch, we have different experiences for different customers. That ability to tailor the experience and work with Nvidia hand in hand in deciding how we built that cooling system was three years in a lab with Nvidia thinking about how to cool that chip and how to do it in a way that candidly didn't give up any water. Because one of the big precious commodities that that we'll start talking about in the next two to three years is what happens to water. Where do you get water? A lot of the delays right now, for example, in Northern Virginia is about water. In Arizona, it's water. Everyone thinks it's power. Yes, it's power, but it's also water is becoming a very precious commodity. Our ability not to waste one drop of water in a switched data center is differentiation. And it puts us at the front of the line in leasing with people that care about the waste of water, like Google and Meta. These are topics that they care about the customers.
- [08:45]
B
Yeah, I mean, water is super important. Power is the bottleneck that everyone is kind of front of mind and we can't escape from. And I wanted to get some practical examples, I guess from you that the trend now is really towards bringing your own power, especially in the U.S. but I want you to give us some examples of what you're doing, what that looks like on the power side. How are you powering your data centers? Maybe if you want mix up some regional examples, just give us an idea of what that is looking like.
- [09:13]
C
Well, first of all, I don't think the power issue is just a US issue. No, I think it is a global issue. I wasn't here yesterday, but I heard Sean intimating that compute and data centers are reaching a global moment. That's been happening for a while. Actually the global moment for compute has been ongoing and I think what we're seeing is the opportunity to create in front of the meter and behind the meter solutions that can differentiate yourself. We take in our best applied learnings from our portfolio companies. So for example, one of the largest micro grids that we own is actually sitting in Sao Paulo, Brazil, just outside. And there we have close to 800 megawatts of power that is 100% hydro. And so there the concept was we leased transmission infrastructure, we had two sources of hydropower, we brought it into our microgrid in Tambor, and then from there we have 14 data centers within essentially 5 kilometers. Now those data centers only consume 180 megawatts of power on a sort of annualized basis. We're selling close to 600 megawatts of power back to the Sao Paulo grid. Now why is that interesting? It was interesting because our connection dates got moved up a lot faster because we were able to be a net promoter to the grid. So understanding that baseload study and understanding the challenges that the Sao Paulo electric grid had and our ability to put power back into the grid put us at the front of the line. While Equinix and Digital Realty were waiting for their connection date, we didn't wait. And so we've taken some of those applied learnings two, three years ago. We've done the same thing in Nevada where we have a 1.3 gigawatt microgrid. And what's interesting about that case study is there's not a single bullet that solves power. If any of you think there is a single bullet, that's not how it works. You have to be really pragmatic and you have to draw multiple sources of power. So there we have solar, we have wind, we have hydro, and now we've just introduced LNG. So that microgrid that's 1.3 gigawatts is bigger than the entire city of Reno. So the amount of power that Switch controls at its supernap is larger than an entire city. And we're completely grid independent. On an average day, we sell somewhere between 100 to 120 megawatts back into Nevada power and lights grid. Why are we doing that? We're doing that because there's a, a good opportunity. We're selling that power between 9:00am and 5:00pm at Spot Retail, somewhere between 65 per megawatt hour and 115 per megawatt hour. The minute it turns 11:00 clock at night, we're buying power back at wholesale, somewhere between 34 and 38 per megawatt hour. So there's a massive arbitrage when you can control your power. We have almost close to 400 megawatts of battery storage dealing with intermodulation across 28 data centers. And that ability to store power enables you to have this kind of spick it on, spick it off. So during peak, where I have a ton of solar, I'm selling power back into the grid. And by the way, if I don't like the rate that Nevada Power and Light's giving me, I go to Puget, I go to Portland General, I go to Pacific Gas and Electric. I have multiple sources of power. Why? Because I'm interconnected. So this notion of having bess and interconnection and your ability to pivot, essentially trading power every 2/10 of a second, which is what we do, is really differentiation. And that differentiation also manifests itself in our customers. So when two hyperscalers came to us and they had a choice to self perform or go to switch, they went to switch. Now why did they go to switch? Look, these two hyperscalers, without naming names, can build their own data centers, but the efficiency at which switch could light that power and bring that Data hall online 12 months quicker than they could do it themselves. That's differentiation. And it wasn't so much about power. It was about a customer solution set that enabled us to activate for them quicker and, and create really strong connectivity and relationship currency with the customer. So now I can take that relationship in Reno and I can transport it to Texas, I can transport it to Ohio, I can transport it to Atlanta, where we're competing against other people for space. And we say, hey, we delivered that 600 megawatts for you in Reno, we're going to deliver for you again in Atlanta. And customers go where they can trust. Trust is a critical currency in dealing with customers.
- [13:31]
B
You're talking about deliverability. And that's exactly one of the topics I wanted to hit. Because when we talk about this big corporate capex program, which is around 650 billion, 700 billion this year or so some figures go, I wanted to talk about what's actually coming online. And I'll just put some figures out there to frame what I'm referring to. But a recent report from consultancy sideline Climate, this was in February, highlighted that between 30 and 50% of of large data centers set to come online this year are expected to be delayed because of power constraints, equipment shortages, local opposition, all of them, you name it. And of the 16 gigawatts planned for 2026 globally, a quarter of those 140 projects haven't even disclosed how they plan to be powered and only 5 gigawatts is actually under construction, primarily in the U.S. what does this look like on the ground? Are you seeing these deliverability issues? What, what's the reality and what, what is, you know, the hype here?
- [14:28]
C
So you're right. The US grid effectively turns on about 5 gigawatts of new power every year. The data center sector is absorbing that. But at the same time, last year we leased 16 gigawatts against 5. You basically have a supply and demand imbalance of about 11 gigawatts. At the same time, there are power projects, roughly about 8 gigawatts being built in the US that come online sometime next year. And just in our pipeline alone, in the stuff that we're doing with our joint venture Partner, we have 9 gigawatts of microgrids that we're building right now and that's in 12 different locations. And that does not include the Portsmouth location, which was announced with SoftBank power last week of 9.2 gigawatts in Ohio, which is a project that we're involved in as well. So if you add that 9 gigawatts, we're building 18 gigawatts of power. And mind you, we're not a power developer, but we have become a power developer because out of necessity to the customer. So there'll continue to be this supply and trade imbalance which is what the US grid can bring and versus what the hyperscalers want to lease. The challenge around that is making sure that you can be ahead of the customer and you have the solution sets that they want. I think this notion of delays is real. What's happening in investment committees around the world at different GPs that have made data center investments or LP investment committees, they're saying, well wait a second, what's a delay? What's a penalty? And you got to dig into the contract. Right? I was having this debate with somebody earlier today and someone said, why in the world can Swisscom terminate their lease with Inuit? And that's supposed to be a 25 year contract. And I said, well, it's the quality of the contract, which I've been saying to you for a long time, since you and I have known each other. There's a big difference between an MSA and a lease. A master services agreement doesn't hold up in bankruptcy. A lease holds up in bankruptcy. And so if you've entered into a lease with a customer versus a services contract, it's a very different piece of paper. And if you don't understand that, you shouldn't be in infrastructure. The second piece to that is once you do sign a contract with a customer, you really have to be familiar with the data hall delivery conditions. Because until you've delivered all of the deliverables in those cps, you know what doesn't turn on rent. And if the data center is not cash flowing, it's worthless. It's just as fundamentally simple as that. And so a lot of the new kids on the block that are getting into the data center business are learning this the hard way. And you saw, for example, that one of the projects in Abilene, Texas, got to Data Hall 2 and the customer left and they went somewhere else. Well, they went to our data center because we know how to build a data center. And so this notion of, you know, veterans versus tourists is a real issue. Once you understand air rights, water, zoning, permitting, the difference in a will serve letter and a connection date, which is huge. These are details that determine whether you're three months late, six months late, nine months late, and by the way, if you're a year late, you're accruing hundreds of millions of dollars of penalties. And so what you thought was a 12% unlevered return now all of a sudden is a 6% unlevered return. And what was a good investment becomes a bad investment because you didn't understand how you wrote your contract. So details are becoming more and more important. And keep in mind, the Hyperscalers have been signing leases for 15 years. They're pretty good at it. They've been doing it for a long time. And so when somebody new comes into the business, they've never negotiated a lease before. Advantage Hyperscaler, Disadvantaged gp More important today, to be with someone who's done it and understands the battle scars.
- [18:07]
B
Mark, we have a couple of minutes left. I wanted to ask you to peer into the future a little bit for us because we're all talking about AI data centers. There's plenty of stuff around the corner. Tell us what you guys are looking at as the next wave of investment. Where is it going to come from? What's it going to look like?
- [18:24]
C
The biggest thing that we see ahead is robotics. If you're not paying attention to what's going on in robotics, you're missing a really big aspect of what's coming in. Infrastructure today we're at about 200 gigawatts of compute, which has taken us to where we are in AI today. The next hundred gigawatts of compute we'll be inferencing, which is the Ability to move the compute from a large language model into the field, into enterprise use cases, into machine to machine learning, and then the next leg up to that is about 500 gigawatts of compute. That'll happen over the next 10 to 15 years, which is robotics. And that'll happen in the form of AI factories. So the reason this project in Ohio is getting a lot of attention is that it is a 9.2 gigawatt AI factory. An AI factory is ultimately the industrialization of AI and moving those use cases into monetization models that make sense for an enterprise. And so that bridge there of that infrastructure of taking a large language model and then deploying it into inferencing and then turning it into actual an industrial use case requires a lot to the IT stack. It requires more up top in terms of software, hardware, GPUs, but also the ability to deliver that compute requires a lot of fiber, requires a lot more towers, a lot more what I would call interconnection capabilities. We have a lot going on in terms of looking at the future, but for us, it's about robotics. It's about AI factories coming back to enterprises and speaking the language of using a hyperscaler and using a big corporate user and putting those use cases together and delivering the infrastructure, whether it's delivery of drones for Amazon, working with waste management to go completely autonomous in terms of their delivery trucks, taking a business model that had nine hours of productivity into 24 hours of productivity. This is no longer science fiction. The future of infrastructure is automation. The future of infrastructure is robotics. The ability to do that is going to require massive compute, massive interconnection and a ton of capabilities. And so part of this coming back to where this conversation started is with SoftBank. Having an industrial partner that understands the future and what that looks like, and then bringing together the power of our portfolio companies and the ability to form capital and to go execute those ideas is what we're really excited about.
- [20:42]
B
I mean, we're out of time, but I've seen a lot of questions come in and I just. Since we're talking about the future, are you considering data centers in space anytime soon? That's one of my questions.
- [20:52]
C
No, I think I'm going to leave that to Elon. I think. Look, data centers in space. I really enjoy Elon. I think he's really smart. I think he's one of these metahumans. But at the end of the day, remember, satellite infrastructure has this issue called latency.
- [21:06]
B
Yes.
- [21:06]
C
And so the propagation characteristics between 14 strands of dark fiber versus using satellite infrastructure at 300 to 400 MHz. It's a vastly different experience. I love the question up top. Anonymous Can I say it to the crowd? Because this is funny. As an electrical engineer, I have no idea what a microgrid is compared to a grid. It just seems like marketing. Blah blah blah to me. Is Enron back in business? Okay, I don't know who you are, but I really like you. I'd love to have a conversation with you. This is not blah blah blah because blah blah blah was a fake contract with a fake counterparty creating fake cash flows in our micro grids that we have. We have long term 15 year, 17 year unbreakable PPAs set at rates with escalators with investment grade customers. I want to apologize to you. That's not Enron. That's called a blueprint for securitization where you can form capital at 3 to 4%. A very different business model.
- [22:04]
B
Well Mark, I'm afraid we'll have to leave it there. Thank you very much.
- [22:07]
C
Thank you. Good to see you. Thank you both. Really interesting.
- [22:18]
A
That again was Mark Ganzi, Chief Executive of Digital Bridge, in a special episode of the Infrastructure Investor Podcast, recorded live at our Global Summit in Berlin. To hear more of our episodes, head over to infrastructureinvestor.com podcast or you can search and subscribe to to the Infrastructure Investor Podcast wherever you like to listen.