The quarter ended and so did EV tax credits. We survey which automakers delivered before the deadline. In the race to build for AI, chips are grabbing the headlines, but what's the latest with data centers? And are we in a bubble? We break it down with Applied Digital's founder and CEO Wes Cummins. And Spotify, two new CEOs, double the horsepower or just no change at all? For Friday, October 3rd, it's Brew Markets Daily and I'm Anne Barry.
More market details to come, but first, when is just one CEO not enough? Well, apparently when you're Spotify. The music streaming giant announced this week that co-founder Daniel Ek will step down from the CEO role from January 1st, elevating two co-presidents, Gustav Sodarm and Alex Nordstrom, to the top job. Ek will become executive chairman.
Now, Sodestrom and Nordstrom have so far served in two seemingly distinct lanes at Spotify, with one as chief product and technology officer and the other as chief business officer. Now, I got a little deja vu reading about this because the other major streamer, Netflix, has a similar setup. Since the start of 2023, two co-CEOs have brought different strengths to the shared seat. Greg Peters, an engineer by background, has run product, while Hollywood titan Ted Sarandos has focused on programming. Meanwhile, their predecessor, founder Reed Hastings, became, yes, executive chair. Well, it's worked for Netflix. The stock price is up from around $360 at the time of the change to over $1,170 today.
Then there's Oracle, which had two co-CEOs for six years. Safra Catz, who just stepped out and into, yes, the executive chair role, shared the CEO seat with Mark Hurd until his death in 2019, and she has been replaced just last month with, yes, two co-CEOs, one leading Oracle's cloud infrastructure business and the other focusing on industry applications and sales initiatives. But overall, this setup is pretty unusual, so the headlines would maybe lead you to think otherwise. But here's what tends to happen. Benefits of co-CEOs may include double the horsepower in the key leadership role aboard, therefore less stressed about succession issues. You've got an heir and a spare. It's what we say in the UK. And for a period of time, it may help with the transition from the force of personality of one charismatic founder to a future, however far away, with just less of their involvement.
On the flip side, it can leave organizations asking who's really in charge. The structure may open the door to internal politics, which probably explains why the co-CEO roles have recently been shared by folks with very different expertise. And sometimes it can just mean a change in title, but little real change. To go back to Spotify to bring this to life, the new co-CEOs will continue to report to Ek. That's been made abundantly clear. The two have worked together already as co-presidents since 2023, together partnering with Ek to see the number of users pop to 700 million. Spotify's announcement makes it clear that as executive chairman, Ek will, this is now quoting their announcement, determine capital allocation, map the long-term future of Spotify, and continue to provide support and guidance to his senior team, which bluntly sounds exactly like a CEO's job description.
And in his own letter, yes, Ek sent out his own letter separately, he went out of his way to say, quote, "I will be more hands-on than some of my US peers who have a chairman title." In other words, he's not going anywhere. He's still the boss. He's probably just got more time on his hands. Now, given that Spotify's share price has run up by 180% over the past five years, perhaps it's just as well. We're going to keep watching this one and all of these other CEO succession stories.
Well, coming up, one company started life serving Bitcoin miners. Now, it counts AI giants like CoreWeave among its clients. That's Applied Digital. We asked CEO and co-founder Wes Cummins exactly how it got there. But first, a word from our sponsor, Capital Group. Now, our producer John and I were talking about some other podcasts that we listen to.
That's right. You know, Anne, some of the best lessons don't come from a classroom, they come from experience.
On the power of advice, a new limited series from Capital Ideas and Capital Group. You'll hear candid conversations with leaders and innovators who've built brands, reimagined careers, and learned a lot along the way.
What shaped their journeys? Who gave them advice that changed everything? Tune in and hear what they've learned and what they're passing on.
The power of advice, published by Capital Group, Inc. Practically every day on this show, we report on a new partnership between AI companies, chip makers, and their investors. Now, it's time to go beyond those headlines and talk about the infrastructure of AI, the engine that's truly driving all of this. Now, meeting the accelerating demand for computing power requires physical data centers and energy to go along with them. Construction and operation of those sites brings up issues of location, water usage, sustainability, not to mention how these massive effective factories are funded and integrated into the communities in which they find themselves.
So to get some context on the real-world footprint of AI, I'm delighted to welcome Wes Cummins, the CEO and co-founder of Applied Digital, the publicly traded company that designs, builds, and operates data centers for AI, cloud, networking, and blockchain workloads. Wes, thank you so much for joining us here from Dallas today.
Yes, thanks, Anne. Thanks for having me.
Well, whereas to set the stage for this, this is an industry that does run the risk of being highly jargonized. And what we want to do is make sure that our audience has, you know, some of the level-setting items under their belt. So, I'm going to start with a question from one of our listeners, Marissa in South Carolina, who wrote, "What is the measure? What is the metric for measuring data centers? I've heard some of the size of football fields. That's the physicality of it. Others advertise the number of megawatts. At this point, how do you think about it and what does it mean when it comes to hyperscaling?"
Yeah, absolutely. So, I think it's important when we think about AI factories that we're building. So, you have the old style of data center that was really serving up mostly video applications like this, and then your Netflix, your YouTube, all of those types of video applications that we've all grown to use significantly over the past 20 years. Those were based in metro areas. Those were around square footage. They didn't require the amount of power that the AI factories really require. So just to give context, those types of data centers would deliver about 7 kilowatts for an entire rack of compute. We're delivering at our first facility in Ellendale, North Dakota, about 135 for a rack of compute. So significantly higher power density versus what you've seen in the past. So that means that you're going to use a significantly higher number of power over a smaller square footage or smaller footprint. So you really should be looking at the power loads of data centers when you're measuring how large they are going forward. When you're specifically talking about AI factories, this doesn't mean the old style of data center video streaming that's not going away. So both of them continue to grow, but with the infrastructure we're doing, it's really around the amount of megawatts that we're delivering.
Let's talk about how Applied Digital started life a little bit differently, using energy for a different focus, which is actually crypto-focused infrastructure. Talk to us about that shift from your crypto-focused to now your AI-focused, and what actually makes you able to play in both of those arenas.
Sure. So we started building data centers for Bitcoin. And it could be any crypto, but it was really Bitcoin miners. We were not a Bitcoin miner ourselves. We haven't been a Bitcoin miner. But we built large-scale data center capacity for Bitcoin miners here in the US. That started in '21. And then in '22, we started adding what we called at the time high-performance computing. And high-performance computing and AI factories kind of blend together. The reason we call it AI factories, Nvidia calls it AI factories, but also it really captures the scale of what we're doing. HPC historically, and even when we started, was, you know, you're looking at 5, 10, 15 megawatt builds, and now we're doing hundreds and hundreds of megawatts. And so that really captures the AI factory. But, you know, we went and secured a significant amount of power, and then we added a different group of people back in 2022 to start going after high-performance computing, which has grown into AI factories. But we had already started building large-scale data centers, and again, that was around Bitcoin. But it was a lot of, you know, similarities in supply chain. There's massive differences when you look at the type of build they are, one versus the other. But we were one of the first companies down the path of building very large-scale AI infrastructure, AI factories. And it's just been a really, you know, young market. So we started this in '22 and really started building our first really large-scale facility in '23. But our start was data center builds for Bitcoin miners, of which, you know, our biggest customer is the largest Bitcoin miner in the world, which is Marathon.
And so talk to us about how the decision to pivot actually happened. Did you have one of the hyperscalers come to you and say, look, forget the Bitcoin play, work with us instead? Did you actually make the strategic decision to change and then proactively go out to the AI universe? Which came first?
So it was proactive on our part. We didn't have anyone come to us. It was really, we have this power asset. We know how to build high-power-density data centers. We spent a lot of time looking at high-performance computing facilities, and then we spent a significant amount of time hiring the correct people, getting the right supply chain partners, and we've continuously refined that. But we did this pre-ChatGPT. So it was really, what else can we do with this power asset that we have and our ability to go and build at scale and build quickly? We started down this path, and then it was just really, you know, I'll call it a little bit of luck for sure, where ChatGPT hit in December of '22, and it was probably in March of '23 we realized the demand is going to be extremely large. And then for us to break into the hyperscale market, I thought for a newcomer, what we need to do is like eventually there's going to be a significant amount of demand for this type of data center capacity. We started building it ahead of having a customer. We spent close to a billion dollars at our building in Ellendale, North Dakota, before having a customer. Now we've signed a customer for 400 megawatts there. We have another campus going in Harwood, North Dakota, which is just outside of Fargo. And we're already doing construction. We expect to sign our second customer for that campus in the near term. And it really put us on the map to have this capacity available. And the thought process inside the company was the GPUs were really hard to come by. So the H100, when those hit the market in March of '23, it was really almost impossible to get your hands on those GPUs, and it was supply chain issues. But internally inside the company, we thought, hey, it's a lot easier to fix wafer throughput for semiconductors, advanced packaging. That supply chain is going to get fixed a lot faster than we realized what type of data center infrastructure was going to be necessary for these workloads. And if you just thought forward, you know, 18 or 24 months, you go, the data center capacity is going to end up being a bottleneck because getting power, getting permitting, getting supply chain, and then actually building the facilities we thought would be the issue. So we started early and we started big, and it's really worked out with how the market has come together for us.
So you took the strategy, which is a risky one, of build it and they will come, right?
Yeah. It was a Field of Dreams strategy for sure.
And that hasn't worked. I mean, that hasn't worked out for Intel, for example, if you look at their foundry model. So it is risky when you've got a capital-intensive construction model, which is the one that you have.
But a customer did come. So talk to us about that. Talk to us about who signed up with you and when they arrived.
Yeah, so we signed CoreWeave about three months ago. We signed CoreWeave initially for 250 megawatts. So we have a 100-megawatt building. This was the Field of Dreams building, if you will, that we started building back in 2023. And that will turn on actually this month. We're the 1st of October. So we expect that to turn on first capacity this month. It'll be, I think, the largest, we've done some work around this, but let's call it one of the largest supercomputer deployments, the largest or one of the largest liquid-cooled, direct-to-chip liquid-cooled deployments in the world. And we designed the infrastructure really to be that large deployment. And it was really first of its kind in liquid cooling as well. And so it'll be one of the largest deployments. And then we have a second building that's 150 megawatts that's under construction, and a third building that is not far behind that, that's also 150 megawatts, that gives us that 400 megawatts of capacity. And we've now started on with a 280-megawatt campus in Harwood, North Dakota, that we think expands beyond that. Ellendale itself, we think expands well beyond a gigawatt. So we have a lot of power locked in place, but we have supply chain, which we should discuss more, is going to become much more important, I think, in the year ahead. But yeah, that's the scale that we've built at these facilities.
Well, let's talk about supply chain, and let's just take CoreWeave as an example, where as a framework within which to dig in. Just clarify for us, in your relationship with CoreWeave, for example, who is actually responsible for getting hold of the chips and supplying those versus everything else in the infrastructure side that you just talked about?
So our customer in this instance, CoreWeave, is responsible for the chips. They choose which server is going in, who the chip manufacturer is, who's manufacturing the servers, network gear. They decide. So that architecture, excuse me, they decide. What our job is, is to build a highly technical building that runs at five nines reliability, that provides the appropriate cooling and the appropriate power with the appropriate redundancy. And then our job is to build it and then also make sure that that building operates. You know, their job is they'll bring their equipment and they decide on what that equipment is and they own that equipment.
So, let's talk about the cooling. Let's talk about the location that you've chosen. Let's talk about building in the Dakotas. We've got to talk about the proximity to energy. Walk us through how you worked through that.
So, a couple of things about North Dakota and why we're building there. We're really the company that's building that into a big data center region. So we have the 400 megawatts contracted with CoreWeave. We think that site expands. We have another one, like I said, in Harwood. We think that site expands. The reason we love North Dakota, cold weather. So you use a significant amount of power there, it creates a lot of heat and you have to cool. So what does the cold weather help us with? That helps us with efficiency and to cool much more efficiently than you would in, say, Texas or in Louisiana or somewhere in Florida. And what that does is it drives the PUE, which is the efficiency measure, close to one, and one is perfect. So that means that you're using all of the power just for the IT load and not for cooling or mechanical around that. Now, it's really impossible to get there. But then there's also water efficiency. And so our facilities are actually, because they're in North Dakota, they're designed to use no water. It's a closed-loop system. So we don't use water. Now, the last piece of North Dakota is it's a very energy-rich state. So they're producing six to eight times more energy inside the state than they're using inside the state. So they're a big net exporter of energy in the state of North Dakota. And so what we have is access to power, and in many cases what I call the right kind of power, which is a lot of renewable. It's a big wind energy state, as well as being an oil and gas and coal state. So we locate next to substations, and in the instance of Ellendale, two gigawatts of power feed into that substation. Two gigawatts of wind power feed into that substation and oftentimes go unused. And as we attract more customers to North Dakota, we expect it to really grow into a data center region, and we expect to be the leader in that state.
