The Exchanges

Q The way I think about the archetype of, like, what were, what are the, uh, energy resources at a data center historically, was you would have a UPS system, and you would have backup power. And those are basically, you had a diesel generator and a UPS system. And that, like, every data center had those things. Do you think that'll change? It's like, there, is there a new archetype?

A I, I was hopeful early on, but I've never seen one built with, you know, the right type of SLA. So, said a bit differently, yeah, I wanted a straight in facility where the software would have resiliency to fail over, where I didn't have to invest in all the diesel generators or some backup system if the utility failed. I, I haven't seen one come to market, but we've always been watching that because there is a lot of I mean, believe me, I, I don't like to admit the secondary piece, which maybe a lot of the listeners already recognize. We operate almost three gigawatt of diesel generators today. And so finding that balance and, you know, we do utilize some of those for, for peak loads and, and peak shaving and things like that. But I would love to build an environment where for certain workloads, we can build a different type of data center, but just because of the SLAs, because of the requirements associated with these chips where they're liquid cooled. Right. And that liquid cooling, you want almost three in worth of reliability, where if that pump goes down, that those hot set of infrastructure that accelerated compute continually has the right type of liquid for a certain period of time where it doesn't damage that infrastructure. And we're talking billions and billions of dollars, you know, for 30, 35 megawatt build up to 50 megawatts. It's very expensive infrastructure tha…

AI assessment note: “So high hopes, but it never really came to fruition.”

Q most about, but then there were like Some other tier two regions behind that. And that part of what has happened in this new wave of excitement and AI and hyperscale data centers getting planned everywhere is that there's been a big geographic dispersion. And I, so I guess one question for you is, is that true or is it still really regions that drive the majority of the growth?

A Yeah. So I, I think you have to take a step back and look at the problem from two lenses, right? The first lens is, is What are the workloads coming to market? And I think that lens is interesting, right? Where the most simplistic terminology people hear about training and inference. I think training has forced a broader regional deployment, but that's for, you know, training these kind of frontier models. I would say that there's been a lot of growth in that, but we see that kind of leveling out where we really see the consumption of AI or inference that's driving that regional specific growth going forward. And I think that's where it's more embedded. And a lot of the existing, if you will, follow the clouds with availability zones, that's where it's really starting to be that investment growing and evolving quite, quite quickly. And I think you brought it up with Northern Virginia that Nova market has been one of the critical availability zones, which is now represented as a critical kind of AI growth sector going forward as well.

AI assessment note: “training has forced a broader regional deployment... where we really see the consumption of AI”

Q I wanted to ask you about that. So it's obviously, uh, location specific, but can you talk to me particularly relative to your history in this sector from, I don't know, from the beginning of development of a new site to operations of that site? What does that timeline look like? What's the, what's the range of timelines that that looks like today? And how does that compare to history?

A Yeah. So it is a challenging scenario where all things being equal, it takes about two years from concept to delivery to, to build out what I would say versatile data center. And by versatility, I mean, comprehensive portfolio of solving for the hyperscaler needs, but also solving for the enterprise customers. So that's a. 24 month window, but with the backdrop that we're experiencing, particularly with the power and the grid and just the overall equipment bottlenecks, I mean, Utilities are requesting aggressive kind of four year ramp projection that when you start to take that power down, you need to utilize it, which we've been very good stewards in a lot of these markets that when we do that master planning, we project that we are going to need 500 megawatts. We take down that 500 megawatts and operate that over a longer period of time. But some of these other, you know, interconnects are definitely no two markets are alike, but they're elongating even beyond the 24 months that it would take us to pull that together. So There was a lot of challenges there, and I referenced that at a high level, but some of the, you know, broader infrastructure constraints are transformer lead times are 50 plus weeks right now.

AI assessment note: “it takes about two years from concept to delivery to, to build out”

Q bifurcate even the inference workloads into things that are latency sensitive and things that are not, and the ones that are not Maybe you go take advantage of cheap, clean power, maybe even intermittent power out in the middle of nowhere, right? Which definitely exists, but is not where the rest of the data centers are getting cited. Do you view that as a viable approach given the actual workloads?

A It is, it is, and you know, it's great that we're going through the workloads, right? Like that workload is what depicts the infrastructure required to making it successful, and I think latency and throughput remain Many different things, and so I always try to double click on it a little bit. The amount of throughput required is what's challenging, right? And latency, as long as it's consistent for a lot of the workloads we see, they can operate fine, but it's that throughput, the amount of data that's required for delivering kind of an inference type of solution is something that is again proximate, not only to the consumer, but proximate to an ecosystem. So I'll, I'll, I'll hit on your second point where Yeah, we see a lot of text to text scenarios where that workload can be deployed, you know, in two or three markets throughout North America and service the entire market. And so that's a very simplistic kind of scenario where we're in the early innings of AI and the complexities hasn't, hasn't really come to fruition for the broader market. But as you see bimodal, some of these more advanced reasoning models where a token isn't just generated Against a prompt, and then you consume it, and it's done. A token may be generated inside of an AI world and go through multiple models to ensure that, you know, it's not hallucinating or that it has a mixture of experts or a depth exp…

AI assessment note: “It is, it is, and you know, it's great that we're going through the workloads”

Q I wonder whether, because you have the added constraint of really long interconnection lead times, Does it just mean that the interconnection is the long pole and the tent, and so you sort of, you have enough time that the transformer thing doesn't actually, or Twitch gear or whatever, doesn't actually delay projects because you happen to have another thing that takes even longer, or is it its own constraint?

A Yeah, there's, there's two high levels. It is a constraint. Don't get me wrong. There's two high level elements that we've been doing at digital for, I've been here 10 years, the company's been around 20 years, is vendor managed inventory. So not only understanding, hey, here's our portfolio in a single market, but really operating at a point where we're buying that switchgear, buying that infrastructure ahead of time, where we can alleviate some of the bottlenecks. But yeah, that secondary constraint, and this is why I referenced earlier the master planning. Showing and signaling to the utility operators and being a good steward of having, you know, top tier customers and, and creditworthy customers in our portfolio, which want to operate with us 10 plus years in that asset. Balancing that together is everything, and so that interconnect from the utility has become constrained, and, you know, some markets were always investigating, you know, different types of, you know, solutions via gas turbines and, Even those are backlog plus 20, 29, right? Like, that's, that's an extensive background as well.

AI assessment note: “It is a constraint. Don't get me wrong.”

Q wonder the degree to which that presents a challenge for folks like you, because you're, you know, you need to get stuff built, but on the other side of the table from you is a utility who's inundated, right? With load interconnection requests and needs to figure out which things are real and which things are not. And I imagine that sort of gums up the works a little bit.

A Yeah, no, you're, you're spot on, right? And I think I view that as three elements, right? Where, you know, the power, the amount of power, and even the amount of financing required to meet these upper end projections, it doesn't exist, right? And so you can't solve it all, even if you wanted to, but then Double clicking on aligning to your customer, right? And not all customers are equal and really understanding what their goals are and what they're trying to achieve. That's, that's the heritage of digital reality, right? And that's where we've been doing that in pretty much every theater on earth over multiple cycles, right? So AI represents a new cycle in a new wave that's bigger and faster than we've ever seen before, but it takes partnerships to really pull that off correctly. And I think, you know, I, I couldn't say it Better in that, you know, some of the works that we've been doing together collectively, and also with utility operators, having a communication with them to show that, you know, we, we, they won't overbuild unless they have a level of comfort that you'll take and utilize that infrastructure they brought to market. So that ramp is everything to them. Working with that customer to show them that we have the right customers. We have the understanding to support the workload is everything because There's going to be some probably written about very big challen…

AI assessment note: “having a communication with them to show that, you know, we, they won't overbuild”

Q Uh, use case versus how much you could build if you actually have a cluster, and it's all regional, and, and it's near all the other models. Like, that's going to be a much bigger opportunity. Do I have that about right?

A Yeah, no, you're spot on, and there's a confluence of events that are happening there. It's not just power. Um, the expense to stand this infrastructure up is, these chips are not cheap, right? And so, being able to utilize that over a longer horizon of workload and driving that utilization is also a form factor of, if I have it installed for training, and training can be very, a spiky workload, that I can embed inference in that capability to get higher utilization out of that investment, because the ROIC is real on this. And so, You see a lot driving that direction as well, but no, as we see these higher value kind of aggregators, if you will, of multiple models, right, multiple capabilities, that's really starting to become more proximate to one another, and again, data is everything to a lot of these environments, be it hyperscalers or be it enterprise, which we focus on both, having the ability to embed Algorithms or this accelerated compute infrastructure in close proximity to their existing data oceans or constant data creation models is everything to our customers.

AI assessment note: “Yeah, no, you're spot on, and there's a confluence of events”

Q training models need really big scale data centers. Um, but you know, over history, right? Like, you guys probably were developing 20 megawatt data centers 10 years ago, right? And now it's hundreds of megawatts, or you mentioned a gigawatt. What is the demand picture look like for you? Does everybody want the biggest possible data center that you can build them? Or like, what is the nuance to that?

A Yeah, it's, it's a good piece to dig into, right? Where there's a lot of noise, like one of the things we've been joking about is a lot of braggawatts. Oh, I have a gigawatt. I have a gigawatt. And there's just so much noise out there that you really want to get underneath the workload and the durability of the company behind the workload. And that's where, you know, being a publicly traded operator, we're constantly watching that. And not everybody needs a hundred megawatt data hall. And there's certain use cases where A contiguous set of GPU infrastructure, which requires a very discreet capability, which we have some of the strongest heritage of engineering talent within digital that have been solving this for the clouds. And now it has grown, but they want a contiguous hundred megawatt GPU array. So it's not just about the total capacity block, but then it's the densification of that capacity within the asset. And so we're always watching that, but what we're really seeing is inference. Can come in, in like, five-ish megawatt blocks, and you can solve for it a bit differently. Now, the densification is still there, and then the private AI pieces, there's hot spots where it can be, you know, a couple of megawatts as well, but they want to be embedded in their existing portfolio of assets, and balancing those two things is something we're always eyes wide open.

AI assessment note: “not everybody needs a hundred megawatt data hall. And there's certain use cases”

Q is, um, how quickly do we tap out these regions? From a power perspective in particular, because unless you tell me otherwise, I think that tends to be the thing that maxes out first, unless you tell me maybe it's labor, but like, let's take, we talked about Nova, right? Northern Virginia. How close to tapped out are we there? How much more can we possibly build in that region?

A Yeah, so you bring up, uh, great points, right, where tapping out is the right word, where in a lot of these markets, power has been tapped out. I mean, it is a phenomenal market. I mean, some of the most recent stats, it has .five percent vacancy rate, which is phenomenal. I mean, it's a multi-gigawatt market. I think, you know, one of the things that differentiates how we view these markets is coming in and master planning, not only with, like, the entitlements and making sure you have access and the rights to the land, but That master planning arc is sometimes five plus years, and a critical element of that is working with a utility operator so that they know that, you know, when we say we're going to need a gigawatt, like with what we're building right now, right next to the Dulles Airport, that they have an understanding of that power requirement, and in a lot of the cases, they're able to meet that, but in certain cases, particularly in Northern Virginia, which has been wildly, you know, publicized, is that, you know, some of the Not necessarily generation, but the distribution of the grid has been challenged. And so we're always working with different solutions to overcome those shortcomings in the short term, but then ultimately working with that utility operator so that they get an understanding of the future growth associated with these markets. Because again, this in…

AI assessment note: “where in a lot of these markets, power has been tapped out.”

Q a little bit about it, right? There are folks who are saying, I mean, the famously the Grok data center, uh, employed this substantially where you just say like, okay, the grid interconnection timeline is too long. And so I'm going to throw a bunch of generators on site and operate off of the generators as a bridge until the grid comes along for me. How common is that actually?

A Yeah, I think there's some outliers. Uh, very few of it gets covered in the press, um, because nobody wants to really go on to the market where you, the grid can't meet the customer ramp demands today. I mean, full stop. And I think, you know, one of the things that we're always looking at, and I, I keep harping on this, is that customer ramp requirements are an absolute key driver, but bridge power is one tool among many that we're always looking at, right? And, you know, natural gas, which is what you were referencing earlier, I think it, it has a solution in a shorter term horizon, but we're always looking at what are the longer term power generation capabilities that could potentially coming online and working with the utility operators on if it is grid constrained and they couldn't get the resiliency in the grid, or if it's a generation challenge, we're always looking at how do you hit that peak demand with some of the batteries and some of the other technology that we've been seeing come to market. But yeah, it is an outlier. I think A lot of the utility operators are starting to understand that, hey, this is real demand and I'll align to it. They're not chasing the noise because nobody wants to invest in a bubble, right? Like I'll, I'll go on record to saying that we're always looking at to ensure that we're not aligned to a bubble and that long-term durable workload is …

AI assessment note: “Yeah, I think there's some outliers.”

Q When you say availability zone, what is the, what's the promise? What's the availability promise that's being made? Because this is what's driving the It's regions for this reason, right? Like, it is a promise of a certain level of availability.

A Yeah, and so I, I think, you know, I always go one step further on that training inference, but it's monetization. Those availability zones were foundational and set up gravity, if you will, of, you know, what was driving that is SLAs and consumption to the enterprise, right? And I think that's where you see a lot of these capability and infrastructure being invested in these zones all around the globe where, you There's a major city center. Um, it's usually closer to the, the CBD because there's proximate requirements with throughput and latency associated with it. But those availability zones are what has built that kind of, if you will, first wave of cloud infrastructure coming to market. And, you know, availability zones are slowly evolved. They're not everywhere. They're not in these tier two, tier three markets, but we're seeing a lot of AI applications being embedded inside of those availability zones. And, The last piece I'd leave you with is that AI is an and and not an or to cloud, right? I want, I want people to go really comfortable with that is that a lot of these AI capabilities are being embedded in the cloud services you're consuming today, like Copilot and in some of the early capabilities coming to market, but that's how we really see a lot of these markets maturing over time.

AI assessment note: “what was driving that is SLAs and consumption to the enterprise”