Wes Cummins - Experience in Bitcoin Helped Us Manage Supply Chains and Move Quickly | IDCP

Morning, Andy. Nice to be here.

Sure. So as you already mentioned, I'm a founder and CEO of Applied Digital. We started the company, it's approximately four years old. We started building Bitcoin data centers. We've never been a Bitcoin miner, but provided data center services for Bitcoin miners was where we started. And then we started down the path for high performance computing data centers in 2022, and we've really accelerated that build. And so as the CEO, I set direction and kind of do deal making for the company, and there's been a lot of that lately, which we'll cover soon.

Yeah, so I started my career on the investment side in the investment industry, and I've been a tech investor for over 25 years now, anywhere from research analyst to portfolio manager to running my own fund. And so I spent a lot of time, especially early in the first technology revolution of the internet. I guess there's prior ones in semiconductor and things like that. But so we started in late 90s going into 2000, saw the big boom of the internet and then the big bust. It then turned into another boom, and I spent a lot of time primarily in hardware. So semiconductors, network gear, a lot of telecom equipment. So I was predominantly a hardware investor, and it's just been an interesting cycle where when the first internet part started, we were building all the infrastructure for the internet. Stock valuations obviously got way ahead of where the businesses were, but that kind of worked itself out over time. And then in the early part of the mid part of the 2000s, hardware went very much out of favor, whether it was semis or network gear or whatever it might be, and everyone was very focused on software. Software got big valuations. And then it's interesting now because we've kind of rotated back to hardware, semis, network, and now data centers are a big part of that, fiber is a big part of that. So we're back in that. As we saw that start happen, I originally started seeing it with the boom of the Bitcoin network and early in the Bitcoin days, and then that's rolled into we were diversifying into HPC. But for me, it's really just been focused on primarily tech hardware investing, and now we're in another big cycle for tech hardware. And so I've pressed in harder than just an investor at this point to really participate in this new infrastructure build for AI specifically.

But I think things will go a little more lockstep this time. And so there's the infrastructure hardware, and then there's going to be a lot of other hardware that goes along with this. I've been talking for probably nine months or so at this point that I think one of the big AI applications that's coming, and it's going to start to be deployed over the coming years, is humanoids on robotics. And so that's another big piece of hardware, but there's a lot of software that goes along with that hardware. But it's going to be a really interesting time because you're getting into a place where the hardware writes the software. And you know, you train these frontier, these foundational models. We saw the new Grok model come out, was it yesterday or the day before? And the software coding capabilities of these models is really interesting. And so it's almost like you're going to be going lockstep with software and hardware at the same time, which will be different than the last cycle where we built out a bunch of networking, we built out long distance fiber, short distance fiber, big data center, and then it was all driven by CPU compute. And CPU compute was fairly mature, and so it was really what software could you write and what applications could you bring on top of CPU compute. But this time, like I said, a lot of the hardware is going to write the software. And so it's going to be a really interesting time, but I think we're early innings in the build of both of those.

Yeah, so we started really down the path of building Bitcoin data centers. There was a big demand, and there's some parallels here too, Andy. So when we were building Bitcoin data centers, what was going on was there was a real shortage of the new Bitmain S19s, and it was really hard to get your hands on. And they were trading for big numbers in the aftermarket, and people were able to sell, you know, kind of arbitrage and buy and sell equipment. But as we went through 21, we got into 22, what happened was the supply really caught up on the S19s, which, you know, let's call that effectively the GPU of what AI is, but this is for Bitcoin mining. And then you had warehouses full of S19s with no data center capacity. Well, we started building data center capacity in 21, and then we caught up a lot of that in 22 and 23. But I've seen some of the same elements now. There was an element of luck in this for sure, where we started going down the high performance computing route in 2022, pre-ChatGPT, and high performance computing was really a niche market at that point. You had a handful of applications that used it. We were building much smaller scale data centers. That was our plan. We built one of them. But what happened for us is kind of by we started building our HPC, we hired a new team that had experience in high performance computing. They were people that were here. They were working for a contractor in Dallas that was building data centers for, I believe, Meta at the time. So we hired them in 22. We designed a new style of data center. And then by the end of 22, we had a small GPU cloud business running mostly for university research. And then ChatGPT happens. The world didn't really change for us until NVIDIA introduced the H100 in March of 23. And then we were introduced into our first cloud customers where we were running a cloud service, which was never really our intention, but it just happened to be that that's the type of service that they wanted. We were already running a small cloud version, so we leaned into that. And then I saw the amount of demand and where this market was going and went back to my data center development team in the middle part of 23 and said, guys, we need to go back to building hundreds of megawatts of capacity like we did for Bitcoin facilities. And so we designed a new, updated version of our data center, much larger scale. We relied somewhat on the data center team at NVIDIA to help with that design. We used some of the team at Cray, which is part of HP, to build the design because it was pretty new and we wanted to future-proof this as much as we possibly could. And then we broke ground in October of 23 on this 100 megawatt critical IT load facility, and that will come online this year. We've made a huge amount of progress, and we have plenty of video of this facility up online for people to look at. But that's really how we've leaned into it. We saw that demand coming forward, but the parallel that I saw was it was really difficult in 2023 to get your hands on H100, and then subsequently H200. But typically on the semiconductor side, fixing wafer throughput, fixing any yield issues, fixing advanced packaging issues, in my mind, that's going to be a lot easier than it's going to be to get power, get permitting, get supply chain to build data centers to house these, because the data center needed to be dramatically different than what traditional data centers have been from a power density, from a cooling perspective. And so that was really the key, the light bulb moment where we really leaned in on building the AI-focused data centers, the high power density data centers.

Yeah, so like our experience there really helped us just in how we manage supply chain and move very quickly. You know, we were one of the larger purchasers of a lot of this electric gear. We use a lot of similar electric gear, right? Step-down transformers, high-voltage switchgear, those types of things. Now, what I will say, the piece that we went and found and added to our company was experience in tier three style data centers, because Bitcoin facilities are wildly different than tier three data centers. And so, you know, we needed to go add that capability inside of our company, which we did in 22, and then we've supplemented that since. But, you know, that's the piece that I would also point out. While there are some parallels, it's very different building tier three data center capacity, you know, with 59 type reliability versus, you know, Bitcoin facilities where you're really trying on Bitcoin to build as cheaply as possible. Those ASICs are pretty resilient. You can put them in environments you could never put GPU servers into. We actually went down the path of, hey, can we retrofit our Bitcoin facilities? Realized pretty quickly that I think that was not a good idea for us. And then we said, okay, well, let's build a lower cost HPC where we're trying to build for, you know, five or seven million dollars a megawatt versus, you know, traditional tier three at ten to thirteen. And then we've moved on to, you know, traditional tier three builds. So we kind of went through this learning curve of what we needed to do. And I think we've landed in the right place. And, you know, time will tell if we have, but I think we really have.

Yeah, so we have our Ellendale, North Dakota campus. We have our 100 megawatt facility that we expect to come online this year. I think it's the largest, you know, only facility of its kind left that's in the market to come online this year. It's a unique design. It's a three-story design. And really what it was designed for was around the networking requirements for the GPUs, for the NVIDIA GPUs specifically. So, you know, there's a, when you move to these types of data centers, you kind of move away from where latency outside the data center was always the most important thing. Now you're moving into latency inside the data center. So network latency inside the data center. And then InfiniBand has certain limitations on distance. So this is built with kind of network core running through the middle. And you can, in theory, connect every GPU to every GPU inside of this for a 100 megawatt critical IT load and create a massive supercomputer effectively. So that's a unique build. But that campus goes to 400 megawatts of critical IT load. And so we expect 100 megawatts this year, 150 megawatts in 26, and 150 megawatts in 27. And then we have, I think we've talked about, about 1.6 GW of utility power outside of that at other campuses and other locations that we are marketing and hopefully will be building in 25, 26, and 27 as well.

Yeah, absolutely. This is a really unique opportunity to take leading-edge technology to parts of the country that were definitely left behind by the first internet tech revolution. And so you can go into rural communities again because that outside the data center latency is not as important. It's still important, but it's not as important for these types of workloads. What's important is getting access to large amounts of power. And so what we do, what we've done in the past and how we've built the company is we're good at finding large amounts of low-cost power. And we do that by finding, for the most part, what we call stranded power. So this is power that's generated but oftentimes goes unused. And there's not enough infrastructure in the grid to transport it from the point of generation to the point of use. So we take the use case to the point of generation. And so this one in particular at Ellendale has roughly two GWs of wind that feed into the substation. But when the wind is blowing, there's not enough transmission capacity to take all of that out. And oftentimes the owners of those wind assets have to curtail the turbines because you'll run the grid very negative pricing. And they'll run the grid down to about negative thirty-five bucks a megawatt, which ends up meaning low-cost power for us, but also using renewable power that typically would go unused or previously would go unused. So it happens to be that these are oftentimes, especially with renewable energy assets, in smaller communities because you build wind farms in wide open spaces where the wind blows a lot, the land is cheap. And so North Dakota is a big wind producer, and they produce over two times the amount of electricity typically that they use inside of the state. And a goal in North Dakota is for the state to use that energy inside the state for their own economic growth rather than exporting it to neighboring states for them to use it for economic benefit. And so it's really kind of getting two pieces here. You bring a huge development project to a town, a small community. You revitalize some of these communities. You bring high-paying jobs versus what the typical job is in that community. But I think this is important too. It does a lot of things. So it increases tax base for them. So you've got better town, county services and education. But it brings jobs, like permanent jobs. So you have the construction jobs. We have, I think, over five hundred people on site every single day. And so you get that boost to the local economy. And then you'll roll over to kind of ongoing a few hundred jobs, which it's great to have those few hundred jobs. And people love to talk to me about creating a lot of jobs, but it's the right amount of jobs for these size of communities. You don't want to overwhelm them and completely change the face of the community, but really just be an enhancement of job opportunity and economic benefit for the local community. And we've worked really closely hand in hand with the city of Ellendale to make sure that this is a win-win for both of us.

Yeah, so the thing that's been great for us over the past twelve months is our, I would say, our cost of capital has declined significantly, and we've been able to raise larger and larger amounts. And these are big projects, right? That campus for us in Ellendale will be about a four point five, four point six billion dollar development, which is, you know, when you step back and think about that, I kind of make jokes to my wife. She'll show me things. She was showing me recently that, you know, the newest Disney park, I think in Japan, it was like the Frozen Park, and come on a tour with us with a new one point nine billion dollar park in, I don't know if it's Tokyo, but I think it was in Japan. And I tell her we're spending four point six billion dollars or four point five in Ellendale, North Dakota. It's not going to look like that. But, you know, the scale is impressive of the investment. And so for us, you know, raising capital is a key for what we do. And the big change happened first for us when NVIDIA invested in the company back in September of last year. And then we were able to do an attractively priced convertible debt offering, I believe, right at the end of October or the first part of November, where we did four hundred fifty million dollars in debt. And then, you know, more recently in January, we announced a five billion dollar equity commitment from Macquarie, I believe, if not the largest, one of the largest infrastructure investors in the world, also big in the data center market. So they bring more than just capital, and that's equity capital. So that five billion dollars of equity capital will, in theory, unlock about twenty-five billion dollars of total capital for us. And the other piece of that capital is the debt portion, where you do project finance down at the asset level that then typically will flip into like an ABS. So think of, you know, you're building a house, you get a construction loan that flips into a mortgage, and that's effectively what we'll do here. That'll be, you know, roughly, you know, seventy-five to eighty percent of the financing will be in that style of, you know, low-cost debt commitments. And then to follow on that, we saw, you know, SMBC, which is a big lender in this market, about ten days ago for us, we announced that three hundred seventy-five million dollars with SMBC. That'll be a loan that flows into the, you know, basically into the project finance for us, where they'll lead that for us as well. So, you know, we have the capital machine, you know, in place for what we need to get the right cost of capital and the right type of capital for us to go execute on these projects.