Wes Cummins, Founder & CEO of Applied Digital: Ellendale’s Purpose Built AI Factory in North Dakota

Yeah. Thank you. So Applied Digital, we are, I would say, one of, if not the leading company for what we termed a few years ago, next-generation digital infrastructure. So that's high power density, the cooling, the power, the location, the fiber to go along with that. But we find location, we design, we build, and we operate full stack for data centers. And then we primarily cater to hyperscalers. So we announced last month a large contract with CoreWeave for our facility in North Dakota. We're building that campus. We have two other campuses in North Dakota, one in South Dakota, and several in other states. But really the goal for us is right now large-scale, you know, gigawatt-style campuses that are going to be, you know, right now mostly used for training but also work for a lot of inference as well. But we kicked off this journey back in '22, prior to ChatGPT being introduced, and we've just accelerated, you know, through now.

Sure. So we did, we started out, I would clarify that we were always a data center-focused company. So we built Bitcoin data centers starting in 2021. We built 500 megawatts of Bitcoin data centers. We did not mine Bitcoin ourselves. We just ran data center operations for Bitcoin miners. Marathon, being our largest customer, is the largest Bitcoin miner in the world. So we kind of started out doing hyperscale Bitcoin mining data centers. Built that out very quickly. And there's similarities and then there's stark contrast to these. So the similarities are we need large amounts of power and then that's kind of where it stops. And you use some of the same power infrastructure to build these. You are building them at high density, but the cooling is very different, the redundancy is different, the need for fiber, you know, Bitcoin, you don't really need connectivity for. So back in 2022, so we started building all of this capacity for Bitcoin and then in '22 when we started focusing more on what we called HPC at the time and now we've evolved that into AI factories because really the builds are very different than HPC builds of the past. You know, we started focusing on, you know, the ingredients that we needed for a site, which is, you know, land usually isn't the issue. Power is the number one issue. So power, and not just the power we were using for Bitcoin, because in Bitcoin you can use interruptible power. For high-performance computing and AI factories, you have to have firm power that stays on, you know, 24/7, 365. So that was different. And then you need robust fiber connectivity to the sites. And that doesn't mean that you have, you know, one fiber line to a site. It means that you have, you know, two or three diverse paths of significant bandwidth running from the site that go, you know, north, south, east, west. And so we started to focus on that because there's a significant difference. But the builds are dramatically different. And we've kind of been through the iterations of the build where we first said, okay, let's see if we can retrofit our Bitcoin facilities. And we, you know, pretty early on we stopped doing that. So we did that back in '22 and then in '23 and late in '22 and '23 we were building kind of a stripped-down version of what would be a Tier 3 data center. So no backup gen. Some of the other things were taken out of it. We built a small scale of that at our Jamestown, North Dakota facility. And then we decided that was not the way. And then we're building our current version, which is really a Tier 3 data center but for extreme high power density.

Yeah. So with Bitcoin mining, it's not far off from a chicken coop, right? So if you went and looked at our facilities, and I think we build really good Bitcoin facilities, but long, you know, fairly skinny building. You have a row from bottom to top of Bitcoin miners, a single row. You have a filter wall on one side and then the wall is open on the other side and louvers, so you could close it on both sides, but and then you blow air through it, right? Airflow runs through the entire length of the building, creates, you know, dust in North Dakota, sometimes snow gets in, you don't control for humidity whatsoever. But the A6, like the Bitmain A6, are pretty robust. And when you were doing, with Core, when they were doing Ethereum mining, you were using much lower cost GPUs. So maybe the longevity didn't matter as much. But I think, as you know, and as probably as everyone in this room knows, the GPUs that we're putting in these facilities now are extraordinarily expensive. And so the customers that we're hosting in our facilities, they want those GPUs to last a long time. And so you're looking more like a semiconductor clean room with really tight temperature controls.

Sure. So the deal that we announced is 250 megawatts. It's 100 megawatts that'll turn on October and November of this year and then another 150 that turns on in June of next year. There's a third building behind that for another 150. So this is 400 megawatts of critical IT load, which equates to, you know, when people talk about power, there's utility power. So that's like 530 megawatts utility power that goes down into 400 megawatts of critical IT load.

So it's not the max, but if we, so let me just step through the, you know, the 530 down to 400. So, you know, the excess power is used for mechanical and cooling. So the measure we talk about is the IT load that you can actually run, how many servers can you plug in, and then there's some, the other power has to be used for how much you have to use for the cooling and then some mechanical. But so these facilities and then the way we've designed and the way you have to design for the customers that we're going for on the hyperscale side is you need five nines of redundancy. So these can basically never be down. And you need to be able to design the facility and the campus, by the way, to be able to run at 100% on the hottest day of the year in North Dakota, which maybe is still not that hot. It does, you know, get hot during the days there, but in the summer, but you have to design it for that. Typically though, your customers are going to run at like 70 to 80% of the workload capacity and then it's not going to be the hottest day in North Dakota every day of the year. So you're going to have a lot of excess power left. But for these, these aren't mission-critical workloads, but you still run them at five nines, so you have to build for that kind of spec.

Yeah. So I don't, off the top of my head, I don't have the breakdown, but I think the important part is, so the way we've started the company and then the way that, you know, a lot of, I think, capacity for AI workloads can be built in the near term is we utilize what we term as stranded power. So it's power that's actually generated and not used or can't be transported out to a workload. So North Dakota specifically, not a big population base in North Dakota. It's the sixth largest wind-producing state in the nation even though the population is low. You know, it's flat land. The wind blows. It's cheap land. So the incentives in the US as far as building wind have been, you know, build wind where land is the cheapest and the wind blows the most. And North Dakota really checks those boxes. So there's a lot of wind power generation there, but there's not enough grid capacity when the wind is blowing to transport all of that electricity out to be used by an end user or a load. And so what we've done because of these workloads, we actually go locate next to the source of power. We utilize a lot of this power that would go unused. We sit right by the substation where about 2 gigawatts of wind feed in. And that's how we're, you know, we're building this. And so we have multiple locations that are like this in the Dakotas. They're around the US. But they're being used up. And I think that's kind of the short-term solution for, everyone's heard about there's not enough power to power this. This is kind of the short-term solution for that.

So you had to go and find them. And the history is there used to, you know, people didn't necessarily locate things because of power availability. This is, Bitcoin was the first for...

You want it close to the customer. So Bitcoin was the first kind of big electrical load that could go anywhere. And now we're following on the heels of that with AI factories. They again, they can't go everywhere that you could put Bitcoin facilities. But where you can, if there's fiber, where there's power, then you can build these facilities typically.

So we spent the last, you know, five years hunting for these locations. We have quite a few of those locations. They're mostly gone at this point though, as you could imagine. So they're the...