Dan Roberts, IREN — theCUBE + NYSE Wired: AI Factories - Data Centers of the Future

[00:00:00] Speaker ?: . [00:00:06] John Furrier: Hello, I'm John Furrier with theCUBE. We are here at our NYSE CUBE studios overlooking the stock exchange here in New York. And of course, we've got our Palo Alto studios connecting Silicon Valley and Wall Street. Tech and money together. Again, we are here as part of the AI Factory in our world kickoff series of our media week. Ongoing series, we feature the leaders in AI factories, the future of the data center. As distributed computing continues to go, you know, continues to expand. Large scale systems are powering it. Daniel Roberts here, co-founder and co-CEO of IREN. Doing extremely well in the public markets right now. Stocks at a 52 week high this week, congratulations. Daniel, thanks for coming on theCUBE. Thanks for having us, John. So you guys are really a great illustration of an AI factory business. You're also a great illustration of a data center business because you have a lot of data centers, a lot of power. You've done some Bitcoin applications, but now AI factory and the energy story is obviously now understood. Stock prices like almost like literally straight up in the past eight months. Business is good. Everyone's talking about more power, data center stories, front page of all the top news outlets. It's this billion going to this state, another billion going over there, a hundred billion going to open AI from NVIDIA. I mean, it really, it showcases the fact that data center demand is at an all time high. I mean, you got to love that. [00:01:33] Daniel Roberts: Oh, look, it's a good time to be in the sector, absolutely. And it's, I think it's both sides of the coin. It's the demand side. We're seeing that continue to escalate, but also on the supply side, because what we're dealing with AI factories is fundamentally a new asset class. So legacy data centers based in metropolitan areas, just aren't geared up for the level of power required, the rack density and the architecture. [00:01:55] John Furrier: Yeah, and we're seeing the enterprise market like getting ready and it's not fully as opened up as we thought it would be this year. We think maybe next year, but we see massive growth on the hyperscalers and then the rise of the Neo clouds or AI as a service, GPU as a service, anywhere anyone can get horsepower or, you know, any kind of like compute for these applications for training and inference, those are the top applications. How has your AI factory changed? What was the moment you knew, okay, Bitcoin's great. Bitcoin miners, they're bounded by power. AI is bounded by power. What was the moment it was like, okay, AI's here. And then when did it click in on the AI factory piece? What, take me through that mindset, 'cause that's a huge business decision. [00:02:44] Daniel Roberts: Yeah, look, it is, but seven years ago, when my brother and I founded this business, it was on the premise that the digitization of society was coming. We were moving into the cloud as humans and we're seeing all these exponential digital adoption curves where things are going zero to one overnight. So Bitcoin was worth nothing 15 years ago. Today, it's a $2 trillion asset class. Two years ago, AI was confined to the hallways of PhD researchers. Today, it's the next humanity-defining step change. And I think at its fundamental core is it's really hard to build out data centers at scale, yet these end addressable markets are going vertical in terms of their appetite for compute. And that's where we come in. We've got an enormous amount of data centers. [00:03:28] John Furrier: Dan, you know, it's funny. We just moved our new studio here, opening up a new hub in New York. Obviously, you're having the access point in New York. We hear things in the hallway here at the NYSE. Things like, maybe we should turn the data center into a commodities trade. Energy's there. So, kind of like, this is an asset class. I mean, there's already discussions of maybe if I pre-buy the energy now from the data center or buy GPU cycles now, I can resell them. That doesn't really, I mean, that's New York thinking. You guys are there with, it's almost options. It's option pricing. [00:04:02] Daniel Roberts: Yeah, it is because we're seeing that option be cashed in now. We're seeing next to insatiable appetite for GPU cloud capacity. We expanded from 1,900 servers at the end of June, and we've now got 23,000 either operating or being installed in the coming months, and we've got capacity to 3X that in short order [00:04:22] John Furrier: as well. Talk about the data center topology. How many data centers? Can you share a number, and what are you guys doing that others aren't? Because the old data center model was a bunch of REITs, real estate play, hosting providers, and then cloud comes in, then they need capacity, and then you start to see the real systems design of a data center as not just a building with power and putting racks of servers on it. It was like, whoa, let's design this as a supercomputer. That is now obvious, but it wasn't seven years ago. You guys saw that. What do you guys do? Take us through the mechanics of your business. [00:04:58] Daniel Roberts: Yeah, so I think the first thing we did was lock up a lot of power. So we've got almost three gigawatts of secured land and power, and we've got about 800 megawatts of operating data centers today. Where, because we've got those data centers operating, we're actively swapping out Bitcoin mining ASICs and installing NVIDIA GPUs every day at the moment to service AI training, AI inference, and even to your earlier point, we're starting to see the emergence of some enterprise demand via our partners. [00:05:27] John Furrier: It's interesting. You did all the hard work up front. Now everyone's spending a lot of dough to do that. You guys have it. [00:05:35] Daniel Roberts: That was a tough few years, but yeah, right now we're in a really good spot. [00:05:39] John Furrier: Tailwind's an understatement. What's the role of the Dell AI factory? Because one of the things we're seeing, first of all, Jensen Wang coined the term AI factory two years, I think two years ago at GTC, and then this year he said KVCache is the operating system, which was really a telling sign. The networking is a part of it, so it's really in the weeds there. And then Dell co-opted that. So Dell used to sell servers, they still do, but now an AI factory is just a lot of servers. And there's packaging, there's all kinds of things that they've done. What was the role of NVIDIA and Dell technologies in your deployments and your success? [00:06:15] Daniel Roberts: Yeah, look, Dell's role is far more broader than just the servers these days. They've been a really valuable partner on the AI factory side. So their air-cooled variant, the XC9680, has a lot of practical benefits over other OEMs, so we've had a really good success with that. But as we move into liquid-cooled systems, the way they view the data center is similar to us, it's the data center is the unit of compute, and it's about how the whole data center interfaces together. So it's the servers, it's the network, it's the cooling, it's the GPU, and how all that comes together in terms of reliability, resiliency. So we're in the process of rolling out an NVL72. So these are the Grace Blackwell 300s from NVIDIA in partnership with Dell at the moment. [00:07:00] John Furrier: Yeah. You know, we interviewed Michael Dell, we've been covering Dell for a long time. But a couple of folks years ago, EAB and John Rose saw this well, and they talk about it the same way you do. It's like a factory is a system, and then there's a collection of kinds of software on there, scale up and scale out servers, and systems, and fabrics, and storage fabrics, and network fabrics. But what's interesting was is that they go by OCP standards. And what I noticed with your business, I saw the video of the time progression, when you do the swap outs, it's just like, it looks easy. I mean, I know it's not easy, but it's not hard. I mean, it's not like you're tearing the building down to the studs, or in this case, complete ripping out. So talk about that switching cost, because this is where I think the AI factories plays well, because you can just drop it in, I know you guys do that. Take us through what that was like, scope that plan, that transition to AI factories. [00:07:52] Daniel Roberts: So we built multi-purpose data centers, specs for AI factories from day one, in terms of the end rack density levels. So for us today, we've got 160 megawatts of data centers ready to go for these AI factories. So we're actively swapping out those mining racks, installing black wells, and a few AMD chips. [00:08:13] John Furrier: And then the customer side, okay, so you've got the factories up and running, it abstracts out the complexities. When you see a factory, you think output, value. What are the customer value points on the AI cloud service? What are some of the things, and I'll say that you're probably the door being knocked down, what are some of the examples? [00:08:31] Daniel Roberts: Yeah, and the thing I love about the term factories is every other prior industrial revolution has been defined by factories and the workers in it, right? And you come to the fourth industrial revolution, which is It's all about leveraging human intelligence, and within these AI factories, the workers are the GPUs. They're the clusters, right? And they don't clock off. They don't tire. They work. They can scale up to millions, and that's how we're scaling human intelligence. And the ability for us to do that is super exciting. [00:09:00] John Furrier: And the performance side, you have to hit that SLA. You've got to make sure you've got good performance. How are those factories working? What's some of the result? Can you share some anecdotes or stories? [00:09:09] Daniel Roberts: Yeah, so this is where we're relatively unique. We're fully vertically integrated. So we own the land. We own the substations for the power connect. We own the grid connection. We own the buildings, all the way down to the server level. And what that allows us to do is to provide a very resilient, low cost service to customers, because we don't have intermediaries. We don't have third-party co-location fees that we need to pay. We don't have to get on a call to a co-location partner under their SLA. We have boots on the ground, people in the data centers. Our customers have direct access to those people, and it creates a really seamless, reliable service. [00:09:44] John Furrier: And they get what they want. [00:09:45] Daniel Roberts: They get everything they need. At the end of the day, that's the most important thing. [00:09:48] John Furrier: On the business model, one of the things that I love about your logo, it's green. Green is money. Your stock price is doing good. You guys are healthy on the business side. Great momentum, so congratulations. But the sustainability piece is huge. [00:10:00] Daniel Roberts: Can you explain that portion of your business model? Look, 100% of all power we've consumed in the last seven years has been from renewable energy sources. So it's been an absolute defining part of our business. And when you look at the projections for this data center industry, McKinsey forecasting another 100 gigawatts of data center demand over the next five years, you need to do that sustainably. But our view is go to the source of low-cost renewables and monetize that into GPU cloud. [00:10:28] John Furrier: How does that, how did Dell and NVIDIA translate to that? Because, you know, people who aren't informed would think, they think that these AI clusters are power suckers. They just suck all the power out of the earth. What's the partnership with NVIDIA and Dell around aligning with the eco-friendly approach? [00:10:45] Daniel Roberts: Look, a lot of it comes down to efficiency. It's efficiency of the compute layer and it's efficiency of the ancillary power that you need to consume. And the level of innovation, the progress they're making around that, the ability of us to integrate that into our data centers and keep that efficient is super important. [00:11:02] John Furrier: And now that you got the nice AI factories going and the success of the business model, what's next? What's on the horizon? What are you guys looking at? You're vertically integrated. You control your own destiny. You got a lot of power, which is in high demand. Love that story. You got the eco-friendly. You got the sustainability. What other AI initiatives are you guys looking at? Because you got the footprint and you got everything you need. [00:11:27] Daniel Roberts: So you may have put an unintended pun in there because we're developing a liquid cooled data center called Horizon. Horizon 1, Horizon 2, they're capable of supporting 19,000. A lot on the horizon. There is a lot happening. But the immediate focus is filling up the current 23,000 GPUs that we've ordered. That can scale up in the very short term to 60,000. And then we've got our liquid cooled facilities coming online for the end of this year capable of supporting the GB300s. [00:11:54] John Furrier: What are you guys doing that's different than others? Because when you look at like some of the things that are happening, you see people spending all this money. They're buying all this gear. They're billed that they will come. You know, that movie Field of Dreams is kind of a cliche. It's been around for a couple of decades. There's a lot of people saying, "Whoa, whoa, they're buying a lot. Where's that capacity?" Most of the hyperscale people know that they can go in and get a good nine-month projections. But people are worried that the risk may fall on the real estate or the vertical integrated. And so a lot of people think there might be a little bit of a bubble on some of those over the big money capex spends. What's your reaction to that? Because you guys are, again, you're controlling your own destiny and you own everything. [00:12:37] Daniel Roberts: And I think that's key because we can throttle up with demand. We can throttle back when demand slows down. But what we've seen noticeably in the last six weeks is demand going up. And it goes to the mid-market segment of the market. So AI labs, AI companies that are scaling up a level of inference enterprise via our demand partners. But what we're also seeing is appetite at the hyperscale level. [00:13:02] John Furrier: And the thing about the news this week with NVIDIA's, or last week, the NVIDIA's $100 billion investment, OpenAI, and all the data centers being built, is those are ground, they're breaking ground. They're building. So it'll take some time. And I was asked, what does that mean? I'm like, well, it means it's demand, right? So talk about that piece because you guys are up and running. How long do you think that demand will be? I know you probably can't say that being a public company without any forward-looking projections, but like just your gut feel, the demand curve. [00:13:33] Daniel Roberts: Look, right now, sitting here today, demand looks exceedingly strong. It's very robust. But at the end of the day, to forecast out a year, three years, five years, we don't know what's going to happen. This industry is so fast-moving, but at the fundamental heart of what's happening in this sector is this dislocation between the real world and the digital world. So as I mentioned before, all these digital exponential growth trends, but the ability to build new data centers, to get more power online, it's very hard. [00:14:02] John Furrier: Yes, that's why I started with you guys being in the data center, because even if demand might shift, say, inference and reasoning changes, or some sort of new architecture, there's still need for a data center. They don't go away. So the power and the physical plant is the asset. [00:14:20] Daniel Roberts: Did I get that right? Spot on. And case in point is the ability for us to bootstrap this business with Bitcoin mining, where we sold the Bitcoin each day, it was a cash flow play. And today, a higher and better use case for those data centers is emerging. So we're swapping out those ASICs, replacing them with Nvidia GPUs, to provide that GPU cloud service. [00:14:39] John Furrier: Okay, so you guys got the nice playbook. What's different with you guys? How would you describe that? Look, there's probably a couple of things. [00:14:45] Daniel Roberts: One is, I guess we came early to the sector and we locked up a lot of power and land. So we've got an enormous amount of growth. So even those 23,000 GPUs I referenced that we've got on order or are operating, less than 2% of our entire footprint. So enormous room for growth. And the other aspect I would say is our ability to execute and deliver has been noticed by both investors as well as customers. We've never missed a milestone. If we say we're going to deliver something by a certain date, we at least hit it if not exceed it. You mentioned you started this coming with your brother. [00:15:17] John Furrier: Hmm. How'd that go? How's that going? That's great. Yeah. It's great that brother started coming. I started coming with my brother years ago. [00:15:26] Daniel Roberts: Yeah. Look, it's been a fantastic journey along the last seven years. Who goes in and does the cabling? That's what I wanted to know. Fortunately, we've got people better at cabling than us today. [00:15:36] John Furrier: Early days, you're probably in there. Again, final question. As you look at the factories, this is like what we love about this is that we think it's a step function change. I like the asset class angle, but the demand and the apps coming are still waiting. I mean, it's almost the innovations of the infrastructure right now. That's why these data centers are so key. How do you see the developers and some of your use cases with customers? Are they chomping at the bit of like, give me more? What are some of the patterns? Can you share any patterns that you're seeing with the market? I saw Sam Oldman give a talk and he's like, well, when we started ChatGPT, we didn't thought it was going to over here, but then people were just talking to it. Then we realized that was the use case. This is like going back to 2018. [00:16:17] Daniel Roberts: Yeah, there's probably two ways I think about this. One is the demand that we're seeing live time in terms of the customer conversations right now. And this can change quickly, but right now it's exceedingly robust and there's a lot of interest. The second one is when I step back and look at this sector, AI generally, we're still in the first innings, like we're talking about large language models. Have you tried to render a video or an image? Look at the time it takes to do that. That is an indicator of how much more capacity that we need. We should be able to click a button and render an image in two seconds. Instead, we're sitting there for 20 or 30. So what does that need? That needs more GPU compute. That needs more AI factories. And as we get better at it, as we make it faster, as we make it cheaper, you unlock more demand and you get that self-fulfilling growth. [00:17:01] John Furrier: Daniel, you are an AI factory builder. What's your advice for folks out there that need to do this for their business? Whether it's a data center for an enterprise, or an enterprise trying to figure out how to take their existing data centers, if they still have any, most still do. Most of the people with data, like the banks and other people, they have data, but they also have to work with services. So what's your advice for as an experienced factory builder? What's your best practice? Some people are like, "Oh, this is maybe too much CapEx." What's the reality? [00:17:30] Daniel Roberts: The reality is never take a shortcut because it will come back to bite you. Build for the long term, build sustainably, build with a long-term perspective because that way you underwrite a reliable platform. The second thing is customers are number one. Listen to them, provide the most reliable, resilient, cost-efficient service, and you should do well. [00:17:50] John Furrier: Well, thanks for coming in. I appreciate it. Thanks, John. We've got a factory builder here on theCUBE. The AI factory wave is here. As the factories get built, they'll be large, medium, and small ones, but they will be pumping out value tokens, sustainable energy usage, a key part of the design. Again, these are large-scale supercomputers. We are in the supercomputer era. Of course, we're doing our part to bring that data to you. I'm John Furrier with theCUBE. Thanks for watchin'.