Inside Applied Digital: Speed, Scale, and the 600-MW Leap Forward with Wes Cummins

Erin, thanks for having me. I appreciate it.

Yeah, I don't know if I have an inspiring story, but I appreciate you saying that. So, you know, I think some people know, but it's an interesting start to the story because of where it finishes. Hopefully not finished yet, but so I grew up in a small town of 200 people in Idaho on a potato farm. And then I always make a joke of besides working really hard, we worked on the farm, it was still family farm times, but the one thing I learned from working on a farm besides hard work was that I didn't want to be a farmer when I grew up. Uh because it is a lot of hard work. But the so I left went to college then I became the got found myself into the investing world. I was a tech investor been a tech investor for 25 plus years now. And as I spent a lot of time mostly in tech hardware, semiconductors, networking, a lot of infrastructure. And then I saw an opportunity come along which was actually in Bitcoin mining and it was building data centers for Bitcoin mining back in late 20. Started the company in 21. Went aggressively into building data centers in the US for Bitcoin miners. And then in 22 we started decided to start to add high performance computing. It was a much different industry at that time in the early part of 22. So we hired some guys that were building actually data centers for Meta at the time. Designed a new purpose-built high performance computing facility, much smaller scale, and we we started building it in in North Dakota at our site in Jamestown, North Dakota. The idea was we would add, you know, small increments of high performance computing alongside our Bitcoin facilities, high power density, you know, high power needs, cooling, which is great for North Dakota. And so we were building 10 megawatts in Jamestown. By December of 22, we were running some GPUs out of that our facility in Jamestown. And we had a few customers, mostly universities that were doing machine learning and deep learning on the GPUs. And then chat GBT happened. And so chat GPT didn't really change anything immediately when it happened. It was you know was great to play around with and you know it was novelty and that it went to a million users faster than any app and it just kind of kept going. But then when Nvidia introduced the H100 in March of 23 our world started to change then so I was out marketing data center capacity. We'd happen to have built some capacity. It was a pretty good fit for H100 50 kilowatt rack densities. And then it was started getting introduced into uh frontier model companies, AI labs and what they wanted was just GPU capacity. They want to deal with data center. So we started offering GPU capacity by July of that year. We had a little over a thousand H100 GPUs online. Then we had customers for that. But in that time because of what I saw in the market, I went back to my data center design team and I said, "Guys, look, I know I said we were going to do 10 or 15 megawatts at these sites. I think we need to go back to building hundreds of megawatts of capacity." And so we worked through design. We actually worked with Nvidia's data center team, which I think at that point a lot of people didn't know that really existed. We worked with the Cray unit over at HPE that has a long history in supercomputing. And we designed a facility really around Blackwell. It wasn't Cole Blackwell when we designed it, but we were looking at the Nvidia road map. And so we designed the facility around that. We started building it in 23 because we just thought Nvidia is the clear leader here and we want to just map to that. And you know that facility had turned on last month. The first capacity turned on last month. The remaining capacity in that building turned on this month. And then we'll have another building following that. Another one following that that we've contracted all the core weave. And then we have another campus that looks similar to our Ellendale North Dakota campus also in North Dakota that we've contracted to a investment grade hyperscaler. So we target Coreweave plus the five investment grade hyperscalers which are Microsoft, Amazon, Oracle, Meta, Google. And so it's one of those. We haven't been able to disclose which one of those it is. So we're off and running and we're building at large scale and there's a lot of work and we can talk about what went around that. To bring this full circle is we're doing this in small agricultural-based towns, both of them right now in North Dakota. We're looking at other towns, but for me it's a really, you know, personal journey and interesting for me to be able to bring the absolute leading edge of technology to a town that's similar in size to what I grew up in. Right? I mean, Ellendale's five times the size of what I grew up in, but still a very small town. Great people, great community. We work really closely with the community and really try to be instrumental in positive change and the projects that we're doing are just such a perfect fit for these communities. And that it's a bit of a headache when we're doing the construction part of it, but at the end you bring a few hundred jobs, maybe three 400 jobs into a location. You don't change the face of the community. You create a a massive amount of tax revenue for the community. So now Ellendale, North Dakota becomes the wealthiest school district in the state that the community things that you don't think of. So you go to the these towns and now all of a sudden they were doing special tax assessments to fix their sewer system, to do their freshwater, fix a road, and now all of a sudden they have a budget that they can fund all of these programs without taxing the residents. Because what you'll find in these small towns is typically property tax is higher than it would be, you know, in a larger city because you have to provide these same essential services. But it and then it also has been rewarding in that it has actually allowed people who grew up in Ellendale, North Dakota, left because there was no real job prospect in Ellendale, North Dakota. And they've actually come back to work at our facility and their families there. And it's really interesting to me and rewarding and how it's kind of went very much full circle coming from a small town. We were the people left behind. We were the last ones to get the internet. I think my parents stopped using dialup internet in like 2018. So, and they still live there. But, so it's it is a good full circle story.

Polaris Forge One is in Ellendale, North Dakota. Polaris Forge 2 is in Harwood, North Dakota. So in Polaris Forge One is where we've contracted Coreweave for 400 megawatts of critical IT load and in Polaris Forge 2 is where we've contracted the investment grade hyperscaler building the same designs you know our building design we put a lot of work into this we're on our fourth generation of data center designs so we've made a lot of improvements I would say both in you know the scale the flexibility the speed at which we can build these now and the cost at which we can build them and then efficiency and the flexibility is key for the future, right? Because we're I I always make a joke that operating in the AI ecosystem is like operating in dog years and it truly is and the change is so fast. And so we're building assets that we expect to last for 30 to 50 years. Then we need to try to futureproof those as much as possible. So we give those as much flexibility as we possibly can in these. But yeah, we've made a lot of progress. So we started the company 5 years ago and now so we still operate 300 megawatts of Bitcoin data centers but now we have 700 megawatts of critical IT load under construction 16 billion of contracted revenue between two customers over 15 years. So it's been an eventful year. I mean it's been a lot of work and a lot of stress but it's been a really good rewarding year so far. But the year's not finished yet. Let's see how it finishes out and see what else we can do.

There's no real secret other than we were early. So, you know, in 2023 and it was only 2 years ago, but again, dog years, right? And so, when I saw what was happening and I saw the demand level and at that time the real focus was getting could you procure H100s from Nvidia, right? There was no ability to procure these. there was such so if you placed orders you weren't getting delivery for 6 plus months and so everyone was very focused on procuring the GPUs and me with my history and background as a hardware investor I thought to myself that you know they're going to fix wafer throughput at TSMC they're going to fix the co-as packaging the advanced packaging for the semiconductors they're going to fix bottlenecks in transceiver manufacturing and so like they'll fix all of this but then what will happen is the GPU supply chain itself self will balance out with the demand. But I was looking at the data center space and I'm going where are we going to put all of these GPUs and eventually I think I tweeted about this in the summer of 23. They said data center space is going to be the issue. Eventually that's going to be the issue. And I I thought it might be all the way to 25. It could happen in 24. that there was so little of this high power density data center capacity in the market cuz the vast majority of data center capacity was built for streaming and all of our video apps and pictures and messaging and all of those are really CPU based data centers and they only need like I think the average density power density is like 7 kilowatts a rack and then if you want to be efficient with H100 you need to be at 50 and then when we look at Blackwell now where we're deploying it's at 135 kW a rack and so the power densities went up and so now you need different style of cooling because you like air cooling works to about 50KW and then you need liquid cooling after that. And so I was just working through this in my mind and going how are we going to get power? How are we going to get permits and how are we going to build and deal deal with the supply chain and I'm going the the semiconductor supply chain knows how to ramp up really quickly. How do you deal with all of that? And so we did that early and we started early and then we went and built a team really early before the demand kind of is what it is. And so we put a great team together and we you know we stubbed our toe in how we did things. So we kind of started with we're going to retrofit Bitcoin facilities and then we abandoned that pretty quickly and then we were going to build a scaleown version and this was before the chat GBT had showed up and so we' kind of been through iterations and by the time we went to large scale we said look we have this design we worked with the leader for GPUs in the industry the real leader in AI innovation from a hardware perspective and so we designed to that and then we went out and secured supply chain around that and so now we're sitting I thought you know 18 months months ago, I was super optimistic securing, you know, 6 or 700 megawatts per year of MEP, mechanical, electrical, and plumbing for our supply chain. And now I don't think I was optimistic enough. But we went and secured these so that we can actually build, we can deliver for our customers because there's a lot of news and there's a lot of big contracts, a lot of announcements. And I've been saying lately, you know, it's really fun to make big announcements and signing contracts, but you have to deliver on those contracts. like that's the next step. And so we're actually delivering. We turned over the first half of our building on schedule. Turned over the second half of our building ahead of schedule and we're working to really deliver that second building ahead of schedule as well because our customers are really needing the space. So we're trying to pull that ahead and I think we will pull that ahead. But the key here is, you know, not only signing the contracts, making the design, but actually building and delivering a product on time and for the customers so that our customers could be successful as well. So the real advantage here is we started early. We've made some mistakes. We've ironed a lot of those out and we have a really streamlined process now for supply chain for building. We have a lot of power from a site perspective. So we're just in a really good position.

Yeah, so like I said with our team that manages construction and does design and does operations, we did that early and so that was great. We have really fantastic people in place for that. And then there's dealing with the local labor that actually builds the facility. And one of the things that we have done, so we've carved out this spot for ourselves in North Dakota and then and we looked at North Dakota. We said, "Look, it's abundant power. They're generating 8 to 10 times the amount of energy in the state versus what they use." The former governor Doug Burgum who's now secretary of interior, one of his biggest agenda items in the state when he was governor was we want to use the power, the energy that the state generates for the economic benefit of the state rather than exporting it. Here we are, right? So, we bring that use right into the state and we started with Bitcoin and now we're doing it with AI data center. So, you have that you have a cold climate which is phenomenal and the cold climate allows us to run at a much higher efficiency. So, PUE so the power that we use goes mostly to the IT. So we published a white paper back in June because of the location with the abundant low-cost energy because of the weather we actually can save our customers versus southern states on 100 megawatts roughly 50 to 65 million per year in operating costs. That's savings to our customers. That's a huge benefit over time. And so we're signing 15-year contracts. So you can just do the math and even, you know, that's on 100 megawatts, $750 million on the low end that you're saving over that contract length. So big deal there. And then a lot of fiber in the state from a connectivity standpoint because you have a lot of cross-country routes that run through North Dakota. But here back to the point of the question, we're creating a data center region there and we're the only ones doing it. So we have really good handle on the local labor force for building. And even better, we're not competing with, you know, 20 other data center sites like we would be if we were operating in Texas or if we were operating in other states that are much bigger data center states. And we're creating the region ourselves. We're creating the data center region in North Dakota. And now we have two hyperscalers in North Dakota. I think we'll add more to North Dakota. We'll add some to South Dakota as well and just create that region. But have that labor force that's more captive to Applied Digital instead of us plus, you know, an electrician or a plumber can just drive 20 or 30 miles down the road and get a 10% raise. Now, we do spend a lot of time, you'll laugh at this, but we treat the construction people on site really well. I was having a conversation two weeks ago, with Brad Barton, who leads construction for us about the catering and curating menus so that the food is really good on site. And so we do these things and I'm like but with like there's a lot of things on my bingo card that I would have never thought that I would be doing, you know, all going alongside of building data centers and curating menus and what the catering looks like on site was one of those things. But we do, you know, we want to be a great place to work when you're on site so that we retain the labor that's there and is already done. If they're doing building one, we want them for building two and building three and building four so it's just easier to retain that labor. But it is in a lot of areas it's a huge issue. Huge issue.

Yeah. So the day I started the company, my number one goal from a culture perspective was I wanted to create a company where people were excited on Sunday night, right? That that was number one for me. I want people to be excited on Sunday night and I want them to, you know, be feel like we're making a difference. They get to see their impact inside the company. And we're we're a small, you know, we've gotten to be much bigger, but we're still my view a small company. We very much try to create an underdog mentality, which is, you know, it's always great to have that. I'm a big sports person, so you know, I always talk to people like I feel like I'm now the the head coach of the football team, right? And and so it's but we have a an underdog mentality. And then I would just say that the culture outside of that is we just get stuff done, right? We make it happen. I have a lot of problem solvers in inside the org. We've hired just the right people. You know, we've been growing fast. So you don't always hire the right ones. We've had some mistakes there for sure, but we hire the right people that can get the job done, put them in the place, in the seat, and just let them run. And so I try to remove those roadblocks for people to let them do what they do best. And you know, we hired Todd Gale, who runs data center design for us. That was a huge hire. And he's just been an absolute home run for the company. Brad was at DPR. He was building data centers for hyperscalers. On the construction side, we brought in Laura Lello who came to us from Honeywell was at Lenovo prior to that as doing operations for us. She's been great. Then my co-founder Jason has just really come up to speed and he's been he's a fast learner. You know, he went to Harvard, so he's a smart guy, but he's been, you know, really instrumental in getting the company to where it is. And there's just there's so many people I can't list, but getting the right people in the right seats and then removing roadblocks from them and everyone's excited. You know, we're building some of the largest scale projects. We're building the largest scale project in the history of North Dakota for sure, but some of the largest scale projects in the country and in the world and it just creates a lot of excitement inside the company.