As we continue down our AI infrastructure track, happy to have with us Nick Robbins of CoreWeave, who wears many hats, but I think it might be best if you kind of introduce yourself and all the hats that you do wear at CoreWeave.
I wear whatever hats I'm told to. I am a VP in Corp Dev. I lead the equity and equity-linked side of the financing operation. I also lead Investor Relations and do whatever else I'm told to do. I've been with the company officially for like seven months but have been working with the founders for the last few years, first as an advisor back at Morgan Stanley, then I was an investor in the company pre-IPO, and now I'm an employee.
All right, let's jump into it. I think, you know, revenue's gone from, you know, $200 million to $5 billion in a couple of years, expecting to double that again this year and double it again in 2027. You know, backlog of $66 billion. The business is kind of executing on all cylinders. On the other hand, it feels like the market keeps trying to gravitate to some new issue or overhang, you know. I guess, what is it about CoreWeave that you think the market is just generally misunderstanding?
You raise a good point in that there have been, I would say, a number of different referendums over the years, and the good news is we've worked through a whole lot of them, right? I think even a year ago, you know, there was, "What is the differentiation of CoreWeave?" I think today there's a lot more conviction just, you know, globally that, yeah, we've built something different. We've built something that's more performant by doing it kind of ground up for this type of cloud and this type of workload. But you're right in that there are still more questions, right?
The reality of our business is we are really the first company to build a hyperscale cloud natively. What I mean by that is the hyperscalers, right? They all had these other big, beautiful businesses that gave them kind of one free cash flow, and as a consequence of that, the luxury of almost abstracting away the economics of the cloud. Like, I'm not sure the world has really ever seen cloud infrastructure built as quickly as it has, right? First cloud to north of $5 billion of revenue, or in the time that we did it. Then also kinda what the margin profile looks like along the way.
Recognizing we incur costs before we generate revenue, but we de-risk that, right, by signing these longer-dated take-or-pay contracts that effectively give us the ability to see into the future, right? For that $67 billion of revenue backlog, you know, the parts of it that aren't yet stabilized 'cause we are in hypergrowth, we know that we're gonna incur some costs before we hit stabilized margins, which are in the mid-20s at the contract level, and we're willing to do that, right?
Because we know that, hey, if our weighted average contract is five years, we know that we are going to generate sufficient cash flow in that period to service and repay any debt that would be required to deliver that contract, right? To pay for all, you know, contract level operating expenses to service that contract and after doing those things or while doing those things, deliver significant free cash flow to us, right? Creating value for our shareholders along the way. I think people hear the words, but to see it in motion is something new. There's, you know, I would say less cognitive dissonance today than there was a year ago, but we're working our way through it.
The issue at hand today seems to be kind of uneasiness over the credit markets. Now, I feel like it's kind of a bit overblown, especially for you guys. Walk me through it.
Yeah. I would say I think the market tends to misunderstand the primary means through which we access the credit markets, right? Which is via these delayed draw term loans. That's funding at the asset level that constitutes the overwhelming, you know, majority of the CapEx funding that we seek. We don't access that market speculatively, right? It all comes down to the predictability and the contractual nature of the underlying customer agreements we're signing, such that, hey, we signed a five-year contract. That customer is contractually bound to pay us the fixed price per GPU per hour for every day of every week of every month of every year for that five-year period.
When you have something as ironclad as that, you are able to take it to the capital markets and access capital at very efficient prices, well inside the cost of capital of CoreWeave, Inc., right? You know, our bonds have traded up, they've traded down, and you know, at points in time when they were lower than we are today, we would get the question of, "Hey, does this impact your ability to finance?" Again, the answer to that was overwhelmingly no because the three things that matter when you are raising at the asset level via these delayed draw term loans is quality of the contract you've written, quality of the customer you're serving, and quality of your execution, right?
In terms of quality of contracts, you know, from 2022- 2023, we spent close to a year industrializing, right, the HPC infrastructure contract, understanding how to write terms in a way that ensured financeability, right, in the capital markets. Quality of customer and their creditworthiness. You know, the majority of our backlog today is tied to investment-grade customers, and, you know, the rest of the backlog is largely tied to incredibly creditworthy customers who are the leading pioneers of AI that are growing like a weed, that have immense access to capital, right? Like, take OpenAI. They raised over $100 billion in a few months. It's unbelievable, right? The third kind of rung of this is quality of execution. Right.
When we did our first DDTL, right, DDTL 1, creatively named, I know, it was at SOFR+ 962, right? It was for a super high investment-grade customer. When we did DDTL 3, right, and this is publicly disclosed, it was for OpenAI. Between DDTL 1 and DDTL 3, we shaved off our cost of capital by close to 600 basis points. That was not because of OpenAI's creditworthiness versus a super high investment-grade customer, right? It was because the market saw us execute and learned, right, and understood that we are best in class at this, and we know how to deliver.
You are gonna see us continue to over time, and it won't be every deal, but right in the weighted average way, drive down that cost of capital further because of the comfort, right, that the capital markets have with the quality of contracts we deliver, who we serve, and the quality of our business and execution.
We've seen some kind of bank financing deals, SOFR+ 250, some even less. You know, is that where you think this trends?
I think for the right customers, like, yeah, I think, you know, and I think I'll say as much as I can say here, right, which is there have been things written about, you know, new financings we may be pursuing and, they've speculated on prices that are even inside of that. I think all we've been in position to say at the moment is, we'll talk about that at the right time when it's appropriate to, but we're excited to do so and for the market to see us continue to unlock new parts of the capital markets and drive down our cost of capital further.
When you issue guidance for the year, as you did this year, and say, "We're gonna raise or spend $30 billion on CapEx," you know, to what extent do you have confidence in you being able to access the necessary amount of capital to fund that?
Yeah.
Through equity, DDTLs.
So.
However, you know.
Absolutely. Again, primarily, right, through these DDTLs, and then we'll use some customer prepayments and then opportunistic financings at the topco across high-yield convertibles, etc. to do so. We have extreme conviction, and here's why. Another really important thing we said in that guidance when we provided it on CapEx is substantially all of it tied to contracts that are in our revenue backlog already, right? We are not speculatively raising money and going out and spending it aggressively. Rather, we are signing those revenue, those customer contracts where they are, you know, there's a certain amount of cash flow and levered free cash flow that underpins them, and we are levering that.
When you continue to use that, which is probably the most disciplined and conservative way to finance your business, you will have continued and efficient access to the capital markets.
Talk about the maybe percentage of CapEx you are financing. You know, I think you guys are financing, what, 90% of it?
The way an individual contract works is, yes, at the delayed draw term loan level, as of now and to date, let's talk about to date. We finance, call it, yeah, up to 90% or so of that CapEx at the, you know, the DDTL level there. That's right. That leaves kind of for a given contract the remaining 10% and where is that gonna come from, and that's where you see things like customer prepayments and operating free cash flow. We generate, you know, billions of operating free cash flow a year, as part of that, and then opportunistic topco financings as well.
Switching gears to maybe the power side of the equation, which is, you know, where we spend most of our time, you know, we think it's a massive bottleneck for AI in general. You guys have been ahead of the curve in securing power with over kinda 3 G W of secured power plus, you know, a partnership with NVIDIA to, you know, build 5 G W of new data center. What do you guys look for when trying to, you know, find a new site to build a new data center?
Sure. Time to power and source of power are probably the two most important components of that, right? Another thing which goes back to just kind of risk discipline around our business is in everything we do, we focus on kind of the near term part of the risk curve, right? We typically contract power 12-24 months before it comes online, and we typically bring that to customers, call it when it's within six to 12 months of coming online. That is probably the shortest term capacity that is out there. We're not the guys who are in 2026 going to talk to data center developers and customers about things happening in 2029.
We're focused on 2026, 2027, maybe beginning of early or early 2028. 2026, the market is tight, right? You know, I've joked before and I think I mean it, right? If you can find me like 100 MW of contiguous power in the United States that comes online this year, like I will give you my unborn child, right? Like, and again, unborn, so it's easy for me to say. The 2027 market, you know, feels a lot looser. There's still a lot of opportunity there, right? Like up until kinda fall, maybe later fall of 2025, we were signing 2026 power.
When you see that there's access to power, then it becomes what is the quality of that power. CoreWeave is a big fan of grid power as opposed to behind the meter. The reason for that is to deliver, effectively, you know, the uptime and the continuity of power that's required, you need to solve for a lot of redundancy. Effectively what that means is, right, if you're building exclusively behind the meter, you need to, for 100 MW, you might need to build 180 MW-200 MW of actual power at that site just to solve for that 100 MW. When you're building with the grid, you really only need to solve for the backup gen part of it.
That is more efficient. That is easier. We also, I would tell you, our view is that is that, I think people conflate the two, not that there aren't enough electrons in the grid right now. I don't think that is the power shortage we've been working through over the last couple of years. Rather, I think the power shortage we've been working through has been one of there's not enough physical infrastructure to deliver those electrons and turn them into a power shelf. There's not enough long lead time equipment, right? Transformers, you know, backup batteries, power gen, et cetera.
There's not enough, you know, transmission lines, right? There aren't enough workers, right? Like, I like to, in every investor meeting I do, literally ask people, like, "Put your hand up if you know an electrician below the age of 35.
Right.
Like, the hit rate is, like, two in 150, you know? The faster we want to build and the more we want to build, the more stress gets put on that system. The good news is that is not new.
Right.
Like, we've been navigating this pretty artfully for, like, the last three years, and it does not feel today like things are any more onerous than they were a year ago, right? I think we've developed the right playbook. I think we've gotten more efficient along the way, and we export our learnings, right? We work with, you know, close to 40 data center development partners, right? You know, no single data center developer is even 20% of our portfolio. When we build something over here and we see it's better, it's more efficient, there have been solutions and workarounds, we export it to the next guy, and the next guy, and the next guy.
When you talk about your portfolio, there are some projects that you guys are kind of doing on your own, right? You look out a couple of years, do you expect that to be a bigger portion of your data center build strategy?
I do think.
Buy versus rent approach.
Yeah. I do think it'll be a bigger part, but I think it'll take a very long time for it to be the overwhelming part. You know, of the 3.1 GW of contracted power that we have as of 12/31, virtually all of which comes online, or substantially all of which comes online by the end of 2027, that's very largely leased. I think as we look ahead.
Will the old Merck campus come online before then, the New Jersey one?
Kenilworth, yes. It will come online in, yeah, in that timeframe. Yeah, I don't think we've given specific guidance, but I think it's safe to say it'll be in the next, you know, one to two years. That'll be our first self-build online. We're working on one or two others. When you fast-forward to 2028, I think you'll see a bigger percentage be from self-build. But also to be very clear on our approach to self-build, what we do is, okay, we find powered land. We effectively jointly acquire it. We become the minority equity partner in a joint venture, and the long-term lessor or lessee or tenant of that facility.
That is our means of having greater control over the site of recognizing some operating efficiency at that site, investing in our platform and footprint, and trying to build, like, the crown jewels of our fleet, if you will, while also not taking on an overwhelming incremental amount of CapEx burden or obligation, right? That, that'll sit on the majority equity partner's balance sheet.
I mapped out all your different data center locations, and it's clear, like, you guys are, like, spreading out, like the roots are growing. Is that by design? Is that to, you know, if we open up our eyes in a couple of years, like, you suddenly have locations in every core market and, you know, when inference becomes even more thing, you're there.
It absolutely is, and you can kind of think of that as domestic expansion and international expansion, right? I would say on the international side, much like on the technology or product side, like we're customer-led in everything we do. We entered Canada, we entered the U.K., Spain, Norway, Sweden, et cetera, not because we speculatively kinda said, "Oh, someone's gonna want demand there," but because customers we were working with at scale already said, "We could use you here. We would love to work with you here."
I would say with regard to international expansion, you will see that play a role and also kind of the monetization of our cloud stack or licensing of our cloud stack to internationally play a role, the asset-light version of CoreWeave. Domestically, yes, we've built the diversified portfolio. That's a combination of a lot of sites that are probably smaller in, you know, population-dense MSAs, right? We can deliver that kind of low latency inference that certain parts of the market require, and we can do that contiguously across the country. As well as bigger campuses that are probably less co-lo and a bit more single tenant, right?
Where we're working with a world-leading research lab or hyperscaler to develop something that they're gonna wanna use for years and years and years. That's probably a bit more unique, not in the way it's built in that our technology is fungible, but in scale for them.
At this point, I don't think you guys have, like, 1 GW site anywhere, at least on the IT side. You know, would you envision you guys having 1 GW+ site in the near future?
I think it would stand to reason that that would be in the cards, but the way we're gonna get there is not by just saying we're building 1 GW site. Like, the way we're gonna get there is we have several very large multi 100 MW sites where there are options to grow and options to expand, and I think you're familiar with some of them. The way we're gonna get there is by the customers we are working with there saying, "We really want this additional power. Let's build it together," right? Like with again with everything else we do, it's gonna be done with discipline and kind of in a risk management-oriented way.
I wanna talk a bit about the CoreWeave NVIDIA relationship for the 5 GW. How does that look? Is that, like, you sourcing the power, them renting the data center or the lease or some combination?
Yeah, so the 5 GW, we put out an announcement for those who are less familiar in January about expanding our relationship with NVIDIA, and we packed a whole lot into that announcement.
Wow.
It's funny, the market saw a $2 billion investment and kind of gravitated towards that, right? I would argue that there were things that were far more noteworthy in it, right? I think for the first time, gave a deeper look into how we see CoreWeave evolving not just over the next two years, but beyond that, and that's where you got into we expect to, you know, build and add an additional 5 GW of AI cloud by 2030, right? In addition to that, there were two other announcements, right?
One was NVIDIA's intention to take our reference architecture and software to validate those with the intention of ultimately making them available to other NVIDIA Cloud enterprise and sovereign customers, right? That's the asset-light monetization of CoreWeave cloud infrastructure in other people's data centers where we can make money on other people's GPUs. The last part of it was partnership around the procurement of land, power, and shell to effectively accelerate our ability to meet our customers' needs, right? 'Cause I would say or argue that if we've failed at anything over the last couple of years, it's been to keep up with demand.
Like, we are turning people away at the door. It is overwhelming and insatiable and relentless and what other adjectives have been used. How does that part of the NVIDIA relationship work? I would say it's a tool in the toolkit, right? We added close to 2 GW of capacity in last year alone. You will see us continue to do that to self-build as well. When we see strategic opportunities to lean in and further accelerate our ability to grow, having NVIDIA as a partner who's gonna help us do that, in a way that allows us to, you know, meet or exceed kind of our ambitions for the rest of the decade is a pretty powerful tool to be able to leverage.
You're not a small company. As you look at the add-on services, you know, part of what you were saying with the CoreWeave cloud in other people's data centers, along with the storage, CPU, everything else that maybe you'll charge that just hit a $100 million run rate, how big could that get as, you know, a percentage of your business? You know, and how much of a focus is growing that very high margin, the CapEx associated with it.
It's absolutely a focus and it's part of a long-term roadmap, right? Like the easiest corollary to provide, right, is AWS, right? One of the most amazing businesses on Earth, right? Started as an infrastructure business, and then they started cross-selling additional products and services because their customers were bringing more compute workload to them, and they needed more from them, and those were margin accretive and cross-sold and you grew with them. Like, that is the inspiration. The only difference being that we are not building that cloud for browser hosting, right? Like we are building it for artificial intelligence and for powering the future of the cloud, not, you know, the past of it, right?
What we are seeing from our customers is undeniably that they want a unified cloud solution such that they're coming to us and looking for storage. They're coming to us and looking for, you know, networking and ultimately, right, for some CPU too. But again, not for kind of a, what I would call a more legacy use case for CPU, but I think what you'll see as agentic workloads get rolled out over the coming quarters and year or so, right? A CPU being close to the GPU, right, is an important part of that, right? Like, it's even NVIDIA, right, is talking about Vera CPU, right? As another step forward in enabling AI for exactly that reason.
I think whether it's storage or CPU or software and development tools, you will see all of those increasingly become a part of the kind of contribution, right, of what makes up our P&L. I do think, right, the bare kind of—or bare is the wrong word, but the foundational kind of where we started of effectively selling AI cloud, of starting with GPUs is growing so quickly that, hey, you know, when the base business has grown from $5 billion to, you know, $12 billion-$13 billion this year, that hey, even a business going from $100 million of ARR to something well in excess of that, like it's hard to make a dent. That takes time to catch up.
Yeah.
It's pretty awesome, you know, when you think about managing a bunch of different businesses that are, you know, growing to be or if not exceeding nine-figure businesses and on their way to something much larger.
If you look out five years or so, how do you see the customer mix changing? Obviously, you're still gonna have the big, you know, AI labs, hyperscaler customers that are gonna renew, add more capacity, but the long tail, what are you seeing in that cohort that kind of gives you confidence that that mix is gonna be much more evenly skewed, you know, over the long term?
I can tell you what I know, and I can tell you what I don't know. What I know is that the lines of enterprise customers, of AI native customers in our pipeline or signed as CoreWeave customers is multiplying rapidly. You know, in the fourth quarter, we talked about how we had signed more CoreWeave cloud, you know, customers, or about approximately double as many as any quarter in our history up to that point, right? Like demand is proliferating, and we see it in who signs up, we see it who's exploring, we see it who's productionizing AI, right? And so that gives you a conviction that growth is on an incredibly rapid path.
You also see, you know, the leading research labs, right, adding corporate customers like crazy. Like look at the growth of an Anthropic or an OpenAI, right? You see adoption taking place there. That's what I do know, which is all the demand signals, the depth, the breadth of them, the long tail of potential customers are all pointing in the direction of we are years away from balance, and growth will be unbelievable. What I don't know is the primary means through which artificial intelligence is gonna be consumed, right? Like some people think that it's all going through some of the leading research labs.
Well, the good news is we sell the most of those and, you know, so we have that distribution channel open, right? Some people think they'll be primarily consumed through the hyperscalers. Well, we sell the most of those guys too, right? Then, you know, I think another view would be. It seems quite logical that those enterprises will build and consume directly. We sell directly to enterprises and to AI native companies as well. We're setting ourselves up to distribute to all of them, recognizing we're not exactly sure what the balance is in terms of who consumes through what. What we do know is consumption as a whole is skyrocketing.
It feels like we've made a lot of progress across AI over the last year, right? Especially like Anthropic's getting a lot of buzz with Claude. You know, I think it's interesting that, you know, I would assume most of the compute being used is still probably coming from Hopper, like if you had to guess across the space. We haven't even had, you know, the full deployment of Grace get plugged in yet, let alone Rubin. Like the pace of deployments should, I would assume, inflect materially higher. You have like a similar view?
Yeah, I do think Blackwell is either has or is very quickly, you know, displacing Hopper, right, in terms of just the prevalence of it and its, you know, its presence in the United States and globally. I do think the world is just beginning to see the power of leading models trained on Grace Blackwell, right? They talked about it with like Codex 5, you know, 5.3, for example, and obviously, 5.4 coming out last week. I think the world is going to, in the coming months, continue to see the proliferation and impact of AI that is built upon the latest generation of Grace Blackwell, you know, technology, which is inspiring, to say the least.
I do think that innovation will continue in VR into the back half of this year and into 2027. Like I think we will continue to see all of that, but that doesn't at all mean that people are gonna stop working with Ampere and Hopper. Like I actually think quite the opposite. I think, as you see enterprises adopt, as you see that long tail that we were talking about of customers come into this ecosystem, oftentimes they're starting with Hopper, they're starting with Ampere because that is what their kinda workforce is a bit more familiar with.
That's where there might be a bit more open source software available, and it's efficient for them, and the ROI is incredibly attractive, right? I'm not sure it's a one or the other. I do believe that the latest generations of technology will unlock further capability, and that will continue to scale. I also think kinda broader consumption of both prior generations and current generations of technology can accelerate simultaneously.
When you sign like a data center lease for a specific site, maybe with some power coming on this year, next year, or the year after, whatever, do you at that point in time know which GPU is gonna get plugged into that site? Or is that then dictated by, you know, NVIDIA's kind of schedule?
I would say there's definitely. Taking a step back, right? The biggest, I think, change that we've seen over the last couple of years, and probably the largest one we expect any time soon, was the transition from air-cooled to liquid-cooled, right? You know, you could run H100s and B200s even on air-cooled, right? GB200, GB300 as you densify the rack require liquid cooling.
Yeah.
You know, it's pretty challenging once you've begun to build or you've built an air-cooled data center. You're gonna have to take it offline, retrofit in liquid cooling, and it takes time, and it costs money.
Yeah.
I do think when you know you're building something that's liquid-cooled, there's probably some more flexibility to say, is it gonna be, you know, GB or VR that goes into this? I think there's a timing element more than anything else of recognizing, you know, our customers wanna compute as quickly as possible and as much of it as they can get. If a data center is coming online, you know, and ready for deployment in August or let's say in May of this year, you know, it's unlikely that they say, "I'm gonna wait a few months until you can give me VR."
They're gonna say, "Yes, I'm gonna put GB into this and then give me VR for the next thing." But I think when you're signing those things two years out, like, in my mind, it's a bit more about like, okay, is it gonna be air-cooled or liquid-cooled? Then what does that do from a constraint perspective on what technology we can put in here? As opposed to being a month or two smart on once you know the physical build and the archetype, if you will, the generation of SKU that you can put in there.
Awesome. I think maybe just one last question, I think we're over a bit. 2026, what are you most excited about?
Execution, right, I think there's an awesome road ahead of us where we will be scaling some of the largest kind of compute builds the world has ever seen, while also delivering on later generations of technology, right? I think the market, if anything, underestimates the pace of execution and the sophistication of execution that's required to deliver AI cloud at scale. My money's on the world better understanding that kinda towards the end of this year versus the beginning.
Awesome. All right.
Awesome.
Thank you so much.
Thanks for having me. Thank you.
Yeah.
All right. No, but it's good.
I guess this is the joke I'm gonna do.
We're miked up here.
Number three. It's very brief. Hello, everyone. How are we doing today? Has everyone learned something at least? As you know, if you come here, and you don't learn something, it's like, I hope you at least got to meet someone new and interesting. Before we start, I'm just gonna give you a quick dad joke 'cause I like dad jokes. I started hanging out with electrons, great energy, but they're always so negative. Anyway, it sets the vibe. We're very excited to have here joined with Michael Gordon from Crusoe. Based out of Denver, Crusoe is a leader in AI infrastructure across the board, developer, operator. They've got incredible advantages across the board.
We're gonna be excited to dig into all of them, you know, their affinity to energy and software, and Michael's background in software. We're gonna look at how they're growing, how they're leading the charge, and really how they're thinking about the landscape. They've been a Cantor client for a few years now. Some people in the audience have known Crusoe and some of the people at Crusoe for a few years now, which is something that we like to do, build long-term partners, especially now with my brother and I in our new seats.
You know, my brother likes to say we don't really have political goals, but it's really hard to say that at the moment, so you don't really know what's gonna happen. As we've just learned about 14 months ago, you don't really know how this is gonna go, but we're gonna be in our seats for a long time, so we're looking for long-term partners and looking for long-term friends. Anyway, Michael, welcome. It's great to have you here today.
Great to be here. Thanks for having me.
Love to get into it and start. When in Davos, I spent some time with Chase, your CEO, and he mentioned that the name Crusoe came from Robinson Crusoe, which was a nod to stranded energy and looking for stranded. Basically, excess energy that's not being used. They would set up their Bitcoin mining business, like, eight years ago, utilizing this excess energy to drive it towards the compute. Eight years on, you know, how have you seen the landscape evolve from your perspective at the company, and what are the current secular drivers for its growth?
Yeah. First of all, thanks for having me. It's great to be here. The company's had an incredible story coming up on its eighth anniversary. As Kyle mentioned, started Bitcoin mining, two friends from high school, one with a quant trading AI background, one with an energy background. Harnessed the power of stranded energy and methane flares in the oil fields to mine Bitcoin. Incredibly creative, incredibly resourceful. The thing that I think is probably most important for Crusoe today is this sort of energy-first orientation, right? 'Cause if you now, like, bring forward to 2026, we offer a pretty interesting and diversified platform that is all about AI infrastructure, right?
We're providing an AI infrastructure platform where we serve, at one end, the most demanding, largest scale customers, hyperscalers who want data centers built for them, to power AI, and all the other way, we continue and we run our own cloud, for the leading AI companies like Cursor and Fireworks AI and Baseten , who are the sort of AI natives. I think there's this very interesting combination of, you know, fully integrated, and we can talk a little bit about the vertical integration.
Sure.
That energy-first orientation, we'll get into it. We own our own manufacturing for a number of things, so that also helps. Lastly, this sort of diversified piece, so it works quite synergistically across the portfolio.
Great. Let's start at the top. Your competitive moat is obviously your vertically integrated model. How does that differentiate you with your clients? How are you talking about it?
Yeah. It comes in a couple of different ways. First and foremost, I think it's at the highest level where we build better data centers for the hyperscalers because we run our own cloud, and we understand what that's like. Then conversely, we run a better cloud because we actually know what it takes to build the data centers. You can think about this just in some pretty simple ways, in terms of if you actually have the workloads, you understand what they mean and what the impact is on the physical infrastructure of the data center. I'll just give one example, because I think it's pretty interesting.
We have a facility that many people know or have heard of, the first Stargate facility in Abilene, Texas. It's a 1.2 GW campus. It was designed to run as a single cluster, right? If you think about the mechanics of what that means if you're doing a large training set where all the chips are acting in unison, you create a certain level of all those oscillations create a certain harmonic interaction that puts and can put enormous stress on the turbines, as an example. We came up with an engineering solution that dampened that. That's now actually approved by ERCOT. It's actually the only way that a manufacturer will now honor the warranty.
For these large intensive workloads. It's that sort of understanding of the compute and the workloads, but also how it manifests itself in sort of the physical world. I think that's one of the things that's pretty different about us. I think the other thing is on the manufacturing side. That Abilene campus that I mentioned, we built in record time. This predates me, so I can't claim any of the credit. If Chase were here, he would tell you he signed up to do it in just under a year. He went to go order some switchgear parts, and they said, "Well, that'll be a 100-week lead time."
He's like, "Wait a second, I agreed to do this in under a year." Last I checked, that's 52 weeks. We gotta come up with a different way. We now are vertically integrated and actually own our own manufacturing for some of those low and medium voltage switch parts. It controls key choke points in the supply chain, and that's been really helpful. We've also, though, similarly used that manufacturing capability for our own cloud business.
We've got a product called Crusoe Spark that you can think of as like a data center in a box that you can ship and is sort of powered and ready to go and sort of light up megawatts of capacity for our own cloud business. Those are just some of the ways in which we're vertically integrated.
Sure. Obviously, how you get into being vertically integrated is you don't like hearing no. You gotta make sure that you're hitting every angle, understanding every which point that someone might tell you no, so you can speed up your time to delivery.
Yes.
Try to build the best product for your customers. Let's think about this. You've got a very diversified client base, obviously. What's your typical client profile? What are they looking from you for? How does your versatility really lead you to meeting that diversified demand?
Yeah. I think about the continuum, and I think about it even though it manifests itself in quite different ways. I think about the offering as providing AI infrastructure. What that means to a big large hyperscaler is different than what it means to, you know, a venture-backed startup. We try and meet our customers where they are. I think that's really sort of the key part, is bringing our expertise, bringing our points of differentiation to them. On the data center side, certainly people, you know, care about, you know, power, speed to delivery, quality, cost, you know, all those kinds of things are all, you know, incredibly important.
On the software side of the cloud business, you know, obviously, in today's compute constrained environment, first thing they wanna know is do you have capacity? But beyond that, they really wanna know, like, talk to me about your software stack. What can you do? We introduced in the fall a product called Managed Inference. In addition to the Infrastructure -as -a -Service that we're offering that we have that has all the, not just compute and network and storage, but sort of the sparing and the proactive fleet management and things like that, this is actually a further level of abstraction whereby you're just consuming tokens, and you're thinking about throughput and performance.
You can interact in these three different models where, you know, we can build your data center, and we're gonna have a conversation around how much you're gonna pay per megawatt. You know, you can, on the infrastructure side, it's probably a conversation around the cost per GPU hour. Then on the Managed Inference, it's more about what's my price per token, and so the factors vary.
Sure.
I think the one thing that we don't do that's maybe a little bit different than others, and it's important to talk about, and you know, while we're trying to avoid saying no, sometimes, you know, you have to say, be really clear about, you know, where's your best positioning.
Of course.
We're not trying to be all things to all people, and so there are a number of players, who will sort of do kind of classic, what I would think of as, like, kind of GPU leasing deals, right? I'm Microsoft, I want some capacity, you know, for a few years. I'm gonna sign up and go do a deal like that. We have not found that to be important, or our spot of differentiation.
Sure.
In that situation, candidly, we're better off building a data center. Now, not everyone can do that full continuum.
GPU financing might fall more in our court.
Exactly. Yeah, exactly. Exactly. That's a little bit about how we think about things.
Given your software background. It's very impressive. How have you, in your first three months in the role, thought through the software component of Crusoe's stack?
For those who don't know, I spent the prior just under a decade helping build a software company called MongoDB, a database software infrastructure player. I joined MongoDB at $40 million in revenue, and then just under those 10 years, grew it to $2 billion in revenue.
Amazing.
Took the company public. You know, we ended with over 50,000 customers and was really a great story. I have a lot of fondness, to your point, for infrastructure software.
Sure.
Understand the value of that stack, and I think really it's about making sure, you know, you have high quality products and engineering, making sure that you're being really thoughtful about understanding where your customers are, what do customers need, how do we meet them there. I think we're making really good progress. We've hired a bunch of people really across the board. Chase and Cully, the two co-founders, have done a great job of assembling talent, but that includes in the software realm. I've been really pleased. Like I said, we launched Managed Inference in the fall. We just launched a thing called Mission Control, which really gives you more sort of visibility or observability.
If you're a developer building in our cloud, you wanna understand what's happening with the health of my fleet, things like that. I think, like anything, it's just rooted in understanding your customers and then sort of building for them.
For sure. Let's go back to Abilene. That's a very famous project. How is the narrative, I guess, and how many people are thinking about it, talking about it as it being sort of a North Star, how is that really playing into your plan, your narrative, and your growth strategies?
The Abilene campus, for those who don't know, is, like I said, 1.2 GW, very large, you know, project. We built it in record time, delivered the first couple buildings. Still, you know, working on additional buildings.
Incredible project.
Yeah, it's really crazy. I actually have had the good fortune to visit it a couple times.
Oh, nice. I've been invited. I gotta go down.
Okay. We'll definitely get you down there.
Sounds great.
It's a good time. It's really. When you see the scale of it really, you know, starts to hit you. The way that I've tried to sort of help people, you know, for some people, 1.2 GW means something, right? That's accessible. Prior to taking this job, that was not a unit of measure that was helpful to me.
Sure.
As a New Yorker, when someone says, "Well, it's roughly the size of Central Park," I was like, "I can relate to that." If you're in retail, and you say it's about four million sq ft.
There you go.
Okay, I know what that means. Real estate, right?
All the metrics out there.
Exactly, right.
Yeah.
Things you can relate to. The size is really extraordinary. There are, you know, sort of 8,000-9,000 people on site every day.
Amazing.
To your point, like, the scale really hits you when you're there, and you see the sort of 5,000-car parking lot. You, you know, are driving along, and you see a giant tent, and you're like, "Oh, my God. That looks like a lunch tent." You see the sign that says, "Lunch tent number 5." You know, it just sort of gives you a flavor for.
Right
You know, it's like a little mini city. To your question, you know, I think it's been very important. I think it certainly is one of the things that people sort of first identify Crusoe with, if you're familiar with the company. You know, we don't have the benefit yet of being a household name, but for those who know, they often think of that site or that facility. It also, I think, created a lot of momentum around some of these gigawatt style campuses in general, sort of proved that it was possible. I think it's created the real hub in Abilene.
You know, we've had incredibly good relations with everyone there. It's local, state, federal have been super helpful. I think that sort of showed the art of what's possible.
For sure.
I think it's been really a tremendous success so far.
You've mentioned ERCOT.
Yep
Obviously the changing regulatory landscape. How should this room be thinking about, you know, your approach to navigating those obstacles? How should we be thinking about it? How's ERCOT sort of specifically affecting Abilene?
Yeah, no, they've been a great partner to us. I think they're. It's a hard job, right? Given the demands.
For sure.
Given what everyone is talking about doing and their desire to obviously both be constructive but also protect the grid and deliver for their consumers and everything, you know, makes a ton of sense. If I take a step back, it's really interesting. One of the things that I learned as I was doing my diligence was trying to understand, like, why has Crusoe been successful? Talked to a number of people who'd worked with the company, customers of ours. One of the things that I heard pretty consistently was doing what we do requires skill sets in three different disciplines that don't normally go together.
Sure.
Right? There's this sort of like data center construction, real estate side. There's this AI, cloud computing, software, you know, angle , and then the third part of the Venn diagram is this sort of energy and power piece, right? When we get to ERCOT, like, that comes in. We mentioned the company's heritage and the stranded energy and the Robbins and Crusoe angle, and I think that energy first orientation, that energy fluency, has been very helpful to us.
Sure.
I think it's always been important to us to be, you know, a good partner. We think about, you know, when you're building buildings like this, like, you know you're gonna be in the community a long time. You wanna take that, you know, seriously. We've had very effective partnerships with them. I think that we've tried to take, like I said, this sort of energy first orientation. We're very supportive of this sort of bring your own energy, you know, approach to things.
Bring your own energy, I like that.
Yes. When you look at the Abilene facility, you know, like, as I mentioned, 1.2 GW, but it started with just a, I think a 200 MW substation on site. We had to build a 330 MW natural gas plant right on site. Again, that's not a thing that just everyone does, and so there's a comfort and a familiarity and a desire to work with, you know, regulators and really all the constituents.
Well, that's pretty much what I've been hearing is data centers cover all sorts of different topics.
Yes.
In order to work on a data center, you need lots of different types of companies.