DataBank CEO Raul Martynek Talks Data Centers

Speaker 1:
Welcome to TD Cowen Insights, a space that brings leading thinkers together to share insights and ideas shaping the world around us. Join us as we converse with the top minds who are influencing our global sectors.
Michael Elias:
My name's Michael Elias and I'm the communications infrastructure analyst at TD Cowen. I'm joined here by Raul Martynek, who's the CEO of DataBank. Raul, thank you so much for being here, really appreciate it.
Raul Martynek:
Thanks, Michael.
Michael Elias:
So, I want to kick things off by talking about hyperscale demand. So, we've seen really strong demand over the last two years. I'm curious how you've seen the demand picture evolve year to date? And perhaps as part of that, any acceleration or deceleration to call out as we look at the hyperscale demand backdrop?
Raul Martynek:
Sure. So look, just for perspective, DataBank, obviously we're a data center developer here in the US with 65 assets in 25 markets. The hyperscale segment's about 30% of our business when you include the backlog, so we see a decent amount of that demand. What's interesting about our demand with them is that it ranges from 200 kW to 40 megawatts, because we don't just scratch collect the core demand, but we also do smaller type of deployments.
That being said, going into 2025, obviously it's well-known that Microsoft's taken a pause, AWS has taken a pause, so from that perspective, it's been slower than what we saw last year, but at the same time, you have other hyperscalers like Meta and Oracle that are still very active, and you have what are hyperscale-like companies like OpenAI and CoreWeave stepping into the fold. So, I think this year it's going to be ... it will really depend on what happens over the rest of the year and how much of that demand comes back to market, but it's still active. But yeah, it's not as active as it was last year, because last year everyone was looking to get capacity, because obviously we were still kind of in this initial phase of the AI craze.
Michael Elias:
It does seem like there's an element, to use a basketball analogy, of the fourth quarter, right? It seems like there's a swing factor as we think of the year in terms of where we shake out from a demand perspective is do some of these hyperscalers come back at the end of the year? So, it sounds like that's one of the things that you're getting at here.
Raul Martynek:
Exactly. I would say they're already starting to indicate that they're preparing to come back, right? So it's not like it's 100% return, but the indications are they are. It was our view from the get-go that this was going to be a temporary situation, it was going to be temporal, it wasn't going to be a long-term situation. The long-term demand trends are still fantastic these companies are growing at. In AWS's case, 17%, 18% year-over-year growth on $120 billion base business. In the case of Oracle, 49% on a smaller base, right? So the growth is still significant in these platforms and the absolute scale of them is staggering, so it's only a question of time from my perspective.
Michael Elias:
Let's shift a little bit and talk about the nature of the demand. So yeah, as I think about it, a lot of the demand over the last few years has been toward training, and we've put that in some remote areas and we've also put it in major markets. Now as we think of the complexion of demand currently, would you say that the overwhelming majority of the hyperscale demand is still a function of training? Or do you believe we're starting to see some of the early signs of inference deployments emerge, particularly within the major markets?
Raul Martynek:
Yeah, no, I think it is emerging, right? I think we know for a fact that some of the leasing that we've done with the hyperscalers over the last call it 18 months was related to machine learning or AI clusters that they need for enterprise deployments. I mean, they're all, again, a little bit different, right? Meta is deploying for its own purposes, so they're building large training facilities, and because it's really for internal consumption, they can be pretty flexible on geography.
With AWS, Google and Microsoft, that their underlying demand is driven not just from some internal requirements, but really a lot of third party, right? They're the ones supporting all these enterprises, they need to have that infrastructure closer to their current availability zone. So I think the inference use is what is going to drive long-term demand sustainably across the sector, because ultimately the amount of that compared to training is going to be just 10X, 20X more, and I think it's already there.
Then back to the users, we have 2,500 customers, obviously the vast majority of them are large technology and enterprise. We definitely think the enterprises are figuring it out, that they're finding use cases to incorporate LLMs into their business and to drive value. That's something we keep a real close track on, and we're starting to see some of those GPU deployments from enterprises show up in our data centers, right? So I think it's going to be an evolutionary adoption that is going to be revolutionary in the end, and it's just going to take some time. No different than public cloud took time, but long-term the trends are very solid in my view.
Michael Elias:
All right, let's shift gears and talk a little bit about enterprise demand. Now, obviously there's been a lot of macro uncertainty given the tariff dynamic post-Liberation Day. I'm curious, have you seen the tariffs or the broader macro have an impact on enterprise demand or your pipeline for that matter?
Raul Martynek:
Yeah, the short answer is no, not yet. Let me give you a little more details on that. We had a great enterprise quarter in Q1 where 185% of our sales targets, all of that was enterprise, signed our largest enterprise deal in history. A couple trends there, we're definitely seeing the enterprises need larger requirements. A couple of years ago it was rare to have an enterprise with greater than a one megawatt requirement, now we're seeing three megawatt, five megawatt, six megawatt, nine megawatt type of requirements coming from these enterprises. So number one, we think that's a really good trend.
Number two, anything that we did in Q1, almost by definition given the long lead time to deploy data center infrastructure, that project started 12 months ago or 18 months ago, right? So you could say that Q1 demand is really a rearward looking indicator, and really it's about what our enterprise is doing today that's going to impact demand in the future. But we haven't really seen much of a pullback. We've seen some cautiousness around data center terms, but ultimately so far it hasn't had an impact, but it's uncertainty. At the end of the day, it's certainly not benefiting us, but it doesn't feel like it's hurting us right now.
Michael Elias:
I would agree on the dynamic about the deal sizes getting larger. I'm seeing more call it five to 10 megawatt large footprint enterprise deals, and on top of that I'm seeing them with expansion options for a similar amount to the same the initial tranche and capacity. I just wanted to clarify, when you say you're seeing changes around data center term, is that a commentary about contract length or is that commentary about the underlying structure of the lease?
Raul Martynek:
Really around contract length, because ultimately I think they're perceiving that kind of contract length as that's the commitment they need to make, and in uncertain times, let's make a smaller commitment because of that, but you're still talking about five to 10 year contracts with these larger multi-megawatt type of deals.
Michael Elias:
All right, now keeping with the theme of tariffs, we're talking about the implications for enterprise demand, I just want to touch quickly on the implications for build costs for data centers. Now, I appreciate you build both hyperscale and enterprise facilities. Just as you look at your portfolio of builds, how are you seeing it shake out? Is there a distinction that you could make either in terms of when the capacity is coming online or the type of capacity it is? Any color you can give, that would be fantastic.
Raul Martynek:
Yeah, so just a slight add to what your comment there is, even though we're supporting both enterprise, large technology and hyperscale, we've actually standardized on a single basis of design that allows us to pivot between those different workloads with more modifications at the data hall level than at the core level, right? So the basis of design is very similar, but ultimately our capital budget this year is $1.5 billion. We're building out about 12, 13 active projects, and what we see across our projects, 'cause they're all known nodes, we've acquired the land, we've entitled it, we've already secured the power, we're in some state of development of that project. So we're very clear about our build costs, and the range is really about 10 and a half million to 12 and a half million is what our build costs are. We think that's a very attractive cost given where pricing is in 140 to 145, 150 range per kW, so we're able to generate greater than 10% cash yields on that stuff. So, that's where the cost is.
In terms of the tariff impacts, we've looked at that pretty deeply of course, and it's hard, because it's a moving target, it's constantly changing, but our assessment, the way we look at it is we have to divide our portfolio into three buckets. There's data centers that are in development and are so far along that the tariffs really aren't going to have any impact to them, and that's inventory that's coming online in 2025 or very, very early 2026. Then there's a set of inventory that's coming online starting in mid-2026, the tariff impact there is between 4% and 7% depending on the stage of the asset.
Then there's the assets that are going to be delivered after 2026. Those, it's hard to know, because you really haven't procured a lot of the components for that. There our answer is we're going to have to pass those costs, if there are increased costs, onto our customers. Then from that perspective, we feel comfortable about that, because that's ultimately what's happened across the sector over the last couple years as pricing has doubled over the last couple years as a function of demand, but also as a function of the supply chain issues that we had coming out of COVID that drove that increase in build costs from where everyone was in this eight to and a half million range up to where we are today.
Michael Elias:
Perfect. I want to build on that theme, the point that you made about going back to the supply chain shock that we saw. So the way I think about it is that from the lows in terms of pricing and yield in 2021, what we saw is 2022 build costs increased.
Raul Martynek:
Yep.
Michael Elias:
As part of that, the underlying pricing had to go up, but the yield didn't change. Then the cost of capital rerated, the yield went up, but it was just the pass-through of the higher cost of capital, so you could still earn that same return spread. Then in 2024, from my perspective, as the demand meaningfully outstripped supply, you saw underlying yield expansion relative to your cost of capital, so it became more profitable to build data centers.
As we sit here in 2025 and think about the supply-demand dynamic, do you think there's an opportunity for yields to go higher relative to the cost of capital such that it is more economically profitable for you to build data centers?
Raul Martynek:
Yeah, great characterization. I would just say that we started to see that yield expansion in 2023, because 2022, '21, obviously the Ukraine war of '22, that's when interest rates started going up, that's when supply chain was at its worst, so development costs were going up, so you had both cost of capital and development costs going up in '21, '22. '23 is when the demand, the tsunami showed up, and that's when pricing accelerated from below $100 of kW to above 100 kW, probably 125, and then in 2024 we saw that expand even further to $140 to $150 a kW. So at least from my perspective, pricing seems to have plateaued-
Michael Elias:
Okay.
Raul Martynek:
... for large enterprise deals, multi-megawatt in this 140 to 150 range. The supply chain is rational now as we're not seeing those costs increase obviously sans tariffs, and then obviously we all know where the five-year node is. So, I think those development yields have grown and we don't see them expanding right now. That's really going to be a function of where is the supply demand in balance towards the end of the year, right? Inventory's still very tight, which is good, but I think it's still a good story, but it's, in my view, not a story that's getting better and better and better.
Michael Elias:
Got it. So the way I think about that is that's impossibly saying that absence and underlying increase in build costs or change in the cost of capital, really we should see data center pricing be relatively stable, because that excess return spread isn't expanding.
Raul Martynek:
Yeah, so I think, again, cost of capital is where it's at, cost to build is stable, cost of capital is stable. That means that pricing should be stable, absent an increase in demand that would then drive higher pricing.
Michael Elias:
Got it. I want to shift over and talk about supply. So, we've seen a meaningful demand supply in balance here with demand outstripping supply. I'm just curious, as you are building the data centers, how would you rank order the major bottlenecks to delivering incremental data center capacity?
Raul Martynek:
Yeah, so our development pipeline delivers capacity between 2025 to call it 2027, right? This year I think things are going to continue to be tight, just simply because a lot of that excess slack in the system got absorbed in '23 and '24. That in '23 and '24 is when you had a lot of new development opportunities start to enter the market. That capacity almost by definition, you have given a 24 to 36 month time period, gestation period is only going to start coming online in '27 and '28.
So from a bottleneck perspective, obviously we all know about power, there's still a lot of very constrained markets. We know of course Loudoun County's still very constrained. We operate in Salt Lake, there's no power in Salt Lake, I mean, that's been tapped out. Rocky Mountain Power has a 2030 time period to be able to deliver more power in that market. We know Santa Clara is obviously very constrained from the power perspective, and then pockets in Phoenix and in Columbus, right? So, these constraints are still there.
What's happened obviously is that people have gone to slightly different regions like Atlanta, or South Dallas in the case of Dallas, to be able to access power with a ready-for-service date of '26, '27. So I think if you're looking at bringing on capacity in '28 and beyond, there "isn't a constraint," because it's so far off in the future that the utility's going to be able to do that. If you're looking to bring on capacity in '26, it's almost like that window's gone, because between procuring the equipment with 52 to 110 week lead times and being able to have a signed FEA, it doesn't work, the math doesn't work. So, that supply is going to be coming out in the future and not in '25 or '26.
Michael Elias:
So to your point, it sounds like getting power in 2028 is still possible in some of these markets. Are there any markets where ... you mentioned Salt Lake City, as an example, where it's 2030. In those markets that are super constrained, do you think that there's a role for the data center operator to play in terms of expediting access to power, potentially even doing generation, or do you think that's a bridge too far?
Raul Martynek:
No, that's what you see. In specific markets where there's specific demand and there is a constraint on power, we have seen a couple situations and have seen zoning [inaudible 00:16:21] for on-site gas generation, even up to 200 megawatts plus on a particular site. I still think those situations are relatively rare. Most of the power that you see out there is frankly not in the core area, it's 50 miles away, it's 40 miles away, it's 70 miles away, it's 100 miles away. So, there's a lot of powered land in "the middle of nowhere." There's less or no powered land in and around core, core zones, and that is obviously what keeps this supply-demand in balance continuing.
Michael Elias:
So to that point, as we get to the end of this, what I would ask you is, as I think about inference, I view it as latency-sensitive, and I've historically thought about it needing to sit similar to cloud within one millisecond or roughly 100 miles away from the on-ramp or the network node within the market. You made the point that there's power 40, 50, 70 miles away from the core market, so as I think about it, that would be the core of Ashburn, and then around the periphery, there are pockets of available power.
One of the things I've been struggling reconciling is for inference, given that the deployment needs to be in the major market, we're not absolved of delivering incremental power, and it seemed like those markets have been constrained. Would it be fair, based on what you just said, to say that the solution to power constraints in the major markets is to open the aperture up to the edges of that availability zone in that ring and there's still opportunity there for us to deliver inference?
Raul Martynek:
Yeah, I think so. A perfect example is we acquired 85 acres in Culpeper, Virginia, which is 50 miles south of Ashburn. Obviously given the constraints in Ashburn, people started moving to Manassas and in Fauquier County, and then right after that is Culpeper. There's a two, three gigawatt tech zone being developed there with ourselves and a couple other developers, and we think that's going to be a logical place of expansion in '27, '28, '29.
So I do think that will ultimately be how this gets solved is through people widening that aperture, both on the hyperscale side and on the large enterprise side, large enterprise and technology. I mean, South Dallas is a market, another perfect example where it really wasn't a data center market five years ago, but now even enterprises, we're having conversations with enterprises about our campus down there and they're positive on the location. So I think you're right, that's how it's going to evolve.
Michael Elias:
Final question for you. We've seen multiple evolutions of the data center market over the decades. I'm curious from your perspective, what do you see as the key opportunities and challenges ahead for the data center market over the next five years?
Raul Martynek:
So look, I'm very bullish data centers long-term, right? The way I describe it is data centers are the foundation for technology adoption. My view is humans are addicted to technology and that there's a limitless number of applications for technology, so long-term this is a great business. So, that is the opportunity is that it's a long-term macro trend that it shows no signs of diminishing.
On the challenges, it's the things that we've talked about. How do you navigate these ebbs and flows of supply and demand in the short-term? How do you pace your capital so that you can get the best return? How do you make the best decisions on site selection? Then obviously it's a very attractive market, so it's brought in a lot of new entrants, a lot of new players that want to participate. That competitive landscape is going to get increasingly more fierce, but hey, that's the sector. I've been in internet infrastructure for 35 years and it's never not been competitive, so I think the wheat separates itself from the chaff, and if you have a good business with a good strategy and good investors, you're going to do well long-term.
Michael Elias:
Raul, it's always a pleasure to spend time with you. Thank you very much for joining us.
Raul Martynek:
Thank you, Michael. Thanks for having me.
Speaker 1:
Thanks for joining us. Stay tuned for the next episode of TD Cowen Insights.