The Rundown - CoreWeave Co-Founder on Sold-Out Compute and What the Market Gets Wrong about GPUs

Episode Date: August 23, 2026

CoreWeave co-founder and Chief Development Officer Brannin McBee joins The Rundown to explain why the company's compute is effectively sold out through 2027, and what's driving demand at this scale. He breaks down the take-or-pay contracts and self-amortizing debt funding the buildout, and answers the "circular financing" concerns head-on. He also makes the case for why customers like Caterpillar are leaving AWS and Azure for CoreWeave, and why older GPUs like the A100 are holding their value far longer than the market expected.

Welcome back to the rundown, interview edition. Today, we are talking to Brennan McBee, the co-founder and chief development officer at Corwee. Corweave has become one of the biggest players in the AI cloud space. Business is booming right now, but so are questions about their business model. So in today's conversation, we get into it all. We talk about how Corweave makes money, how they differentiate themselves from the hyperscalers, why the company believes their debt load is manageable, how take-or-pay contracts work, and why older generation GPUs are holding up their value.

This was a very interesting conversation. It got a bit nerdy and technical there in the middle, but I think you guys are going to really enjoy it. So let's get into it. All right, guys, today we are joined by Brennan McBee, the co-founder and chief development officer at Corweave. Brennan, welcome to the rundown. Thanks. Thanks for the opportunity to join.

Looking forward to it. Hey, I'm super excited for today's conversation. Before we really get into it, I don't want to rehash the core weave origin story. I think a lot of people are familiar that the company pivoted from being a crypto miner to being an AI cloud provider. What I'm curious, though, is more about your role. What does a chief development officer do?

Yeah, it's one of those titles who are like, what are you developing? The development organization of what I run, it's capital origination, it's M&A, it's Vitchips. Right. So our group raises all the money for the business. So in these public market debt and equity transactions at UC, we run M&A processes for the business. And then we have a group that does direct into company venture investing as well. It's a group that's dominated by, you know, ex-private equity, private credit, investment baking guys were pretty much fully based in New York. We qualified as a, you know, pretty small strategic team that just goes and get shit done.

I love it. I love it. I'm kind of curious. You guys are at the forefront of AI infrastructure right now, spending billions of dollars in the process. I mean, just high level, what do you think is the hardest part right now for the company? Is it just, you know, getting the chips, building the physical shells of the data centers, getting the power, the labor, the financing. Like, what's the hardest part right now? Yeah, yeah. And it's funny. That, that has evolved over time, right? At the very beginning, it was absolutely getting the chips and the allocation. I'm talking like, you know, 20, 21, 22, time by.

like very early days as we were scaling the cloud. Then it became financing the business, right? Like you have all this intense demand for products. And that demand was really scaling quickly back in those days. But how do you go finance? And I'm sure we'll get into the financing of the business a little bit later. But we provided and created these very innovative strategies and market to get that financing together to provide the growth vehicle for the business. Where is it today?

Today it really sits on just getting access to more powered shell, right? And powered shell, we qualify as the data center itself. As you guys know, we predominantly least data centers. We have some self-development as well. But the reason why I'd say it's so hard today is because what we have access to now, we had to predict how much we needed two years ago, right? You can't buy and get access to these leases in a slot basis, right? Like 2026.

So that can't get to find any more leases out there. 2027, pretty much gone at this point, right? Like there's nothing there. Like 2028, very competitive. And I'm talking, you know, most of the U.S.-based deployments. But for getting access to incremental data center space, that will be online so that we can move our infrastructure and our software technology into those sites. that's a tough thing to predict into the future.

And I mean, I remember two years ago when we were planning for 20s, we're like, man, this is a lot of capacity? Like, is this right? And we were getting all the right signals from our client base that I think has, you know, evolved pretty materially from two years ago as well to keep growing, keep growing quickly. But, you know, if you had extra 100 megawatts, extra 500 megawatts, whatever that number is today, gone in an instant.

It's just so funny you say that because, like, I remember two years ago when like the, you know, the, the, the, the, the CappX numbers really started going up across the board and everyone started freaking out about it. What's funny is that everyone still didn't invest enough because there still isn't enough compute because I think everyone was so worried about overbuilding at the time. And honestly, those Cappax numbers should have been double what they were back then. And maybe the compute constraint wouldn't be so so bad. So you guys have a tough job, I got to say. Yeah, yeah. And look, I think that people are so freaked out by Cappax numbers, right?

They're huge numbers. We are moving out such an unbelievable velocity. of technology and capital. It's truly some curriculum efforts that are happening across the industry, right? Like, it's truly phenomenal watching our peer set execute.

It's phenomenal watching our team execute. The scale of these projects is so mind-boggling when you go to these sites and the fact that we're able to bring this all together and deliver the intelligence, the infrastructure for the intelligence, creation

of intelligence, like that's just such a cool thing to be a part of it. Yeah. And so let's let's think into some of the Core Weaves is actual business. You know, you guys are coming off a big quarter. Revenues doubling year over year. Over $100 billion in backlog. Like you said, near term capacity, pretty much sold out. And then you guys are spending, you know, 30 plus billion dollars on CAPEX. The other eye-watering amount, though, was the debt on the balance sheet, right? I think over $30 billion. Why is that something that investors shouldn't freak out about? It's, the, the, the, There's a number of things. And let's spend a few minutes on this. I think why don't we start

with how does our debt work? How do we finance the business? Right. So I think that's a point that is a little bit misunderstood. We break the financing into two categories, right? There's Parenthood or Topco financing. And then there's Asset Co financing. Asset Co financing is where the overwhelming majority of our debt sits. And Asset Co. as the name in first is where a lot of our assets are as well. So that's where GPUs predominantly sit is down at ASAGO, right? And what we came up with, and this is what I was alluding to, you know, a few years ago where we first brought these things in the market.

This is where the kind of collateralized GPU financing comes from. And the idea we had, it's actually, you know, we kind of pulled it out of other infrastructure financing that other sectors use, like take liquidified natural gas. They use this type of financing to do their terminals. And the premise of it is you go out there, you sign these multi-year take-or-pay agreements with your clients, right? Which means they must pay over that entire duration of the contract. And the contract is fixed economic terms, right?

They pay the exact same rate every single month. It's for a fixed skew and fixed set of infrastructure, right? It's not like they can upgrade the infrastructure or change the volume or anything like that. And they, there's no variability in utilization for us in terms of our economics either, right? They pay the same amount if they use an unpercent infrastructure or zero percent of infrastructure. As we expect, everyone uses 100 percent of infrastructure. Right. So now you have this really attractive stream of payments out that you can, that you have

contracture, right? We take the stream of payments and you pair it with the infrastructure and you go, collateralize these things together. And the way it works is you go to the creditors, you say, all right, here's our data center contract, here's our GPU update contract, here are the GPUs. We've set it up to where revenue comes in to this facility, a box, so to say. Revenue comes in off the contract. It pays down, first and foremost, the debt, right? It amortizes. So it pays the principal and it pays the interest. on the debt radically, right? It's called a self-amortizing instrument, right? Not all the debt is paid off

at the end. It paid off throughout the term of the contract. Then the next stream of payments goes to data center operations, right? So your next dollars pay off the operation of these GPUs. And then the final amount, the remaining amount, the remainder, as it's referred to, gets kicked back up to the parent pill. Right. So if we break that down into like, that's the profit, right? That's the profit. exactly that's the profit that goes into the parent and so if you break this down into dollars right like if you have a dollar revenue that comes in you have 75 cents that goes to paying the debt paying the interest and paying the data center operations and then 25 cents a contribution margin of 25 goes up to the parent code and that's that's the profit so it's a self amortizing facility that

takes care of its debt in other words like if we didn't build another GPU cluster, didn't take out another contract, all of that debt at Asico gets paid off by the existing take or pay contracts that we have. And these are, you know, I think like four and a half to five years in duration or so as weighted across our hundred and I think four billion dollars of backlog that we disclosed at the end of Q2. It's been a wonderful way to finance the business. And we've liked this way of financing because it doesn't require anything. mince amount of equity, right? Like, if we had to go do all equity for this instead, it dilutes the shareholders immensely. It's a way of financing business that doesn't require us to need

upfront payments, to rely on our clients to finance our business. Like, of course, like, you know, clients are willing to give us 50, 75% plus upfront payments to finance everything. That works great as well, but man, that's a tough way to build a business because it puts all the leverage to the client because they can show up someday and say, I don't want to give you enough front payment anymore. Like, and then you're kind of left with, well, how do you finance your platform? And if there's something about the credit markets, it's like they deeply appreciate a track record of execution and participating in the credit markets. And that's something we've done really well. We've gone through a number of these facilities. We've successfully brought them online. successfully paying them down. And that's why the credit market continues to show up in larger and larger size for us at a decreasing cost of capital over time. It's a very important financing mechanism that I think is quite beneficial for the business. That makes the question, though. So the take or pay contracts are great for you guys because it kind of de-riskes the debt,

de-restes it for Corweave. But the follow-up question to that is like, you got you. You you still become reliant on your clients. And that kind of brings in the question of whole circular financing where like you're dependent on the hyperscalers to, you know, to sign these take or pay contracts. They're willing to do it right now because there's a shortage of compute. So they need that compute. They're willing to agree to these terms. And but what happens in the future three, four, five years from now, if we have better

compute supply demand, you know, balance, what happens? what happens then? What if they were not willing to sign these take or pay contracts? Yeah, yeah. So the hyperscaler side, right? And this goes a little bit to customer diversification and everything. I think that's something that we took a lot of well-directed criticism for in the quarter leading up to our IPO last year.

Right? We had a lot of customer concentration with one hyperscaler. It makes sense then. Obviously, we've driven so much diversification across our business, right? And the sectors sit in Enterprise. They sit in AI lab. They sit in hyperscaler. Hyperscalelor includes companies like meta as well.

And we announced Caterpillar as a client in this past quarter. I would think people hide that on their bingo cards. Yeah. Caterpillar. A lot of catapillar has tons of existing hyperscaler like cloud relationships. Right? You have to think about it. It's like core week must be doing something so much better than the existing

hyperscale cloud providers that Caterpillar is willing to move its workload requirements away from there in the core weave. Can we talk more about that? Like what is it? Like what is it that CoreWeave does that's so different than like the AWS or the Azure or the Google clouds of the world? Yeah. It varies.

But, you know, it comes down to a simple premise that. our hyperscale peers, they were optimized over a decade of engineering and investment work to host websites and store data leaks. And it's doing a great job of that. It's a phenomenal product they have for doing so, but it's an entirely different solution to run AI workloads. The analogy I've used in the past, which I still like a lot, is, it's kind of like walking into Toyota and asking them, why can't you produce a model while? And Tesla is going to look at you and say, no, no, I'm sorry,

Toyota's going to look at you and say, no, of course we can. They're going to take their camera, put a battery in it and say, here it is. It's called a Prius. But we all know that a Prius and a Model Y, while they're both EV, they're entirely different products, right? One product was built around being a product. electric vehicle. The other product was taking a traditionally produced vehicle augmenting a bit and asking their clients to take compromises, right? And those compromises

are what the hyper-scales ask their clients take with those platforms because it has these legacy components that at the end of the day, just don't have the performance of Corey's product. And you ask like, all right, like, where's the credibility in that performance? Look at our client base, right? We have, I believe it's 10 of the 10 top AI labs on the planet or on Corleaves platform. We have a number of the hikerskillers on our platform themselves. We have enterprises signing up from Corleweave every day. I think that the client base speaks to it. I think third-party research speaks to it through fantastic organizations like semi-analysis. and I think our supply chain speaks to it as well, right?

Like the data center operators are choosing to work with us. The chip suppliers are choosing to work with us because of who we are the quality of our product and market. Yeah, I mean, that makes sense. And I think that's something that, you know, maybe casual retail investors have a hard time understanding. They just see CoreWe, they see AI cloud provider. And then they're thinking, like, well, how are they going to compete with the mammoths like the AWMETs, like the AWMETs, like the AWS and the Google Clouds. And like, you know, and the other thing is these hyperscalers have advantages when it

comes to better financing terms and things like that because they're a trillion-dollar companies with massive cash flows. That's another disadvantage that, you know, that core we would have when it comes to competing with them. It is a disadvantage, but it's when we made so much progress on, right? Like, I think two and a half, three years ago, we were financing that SOFA plus 850 on like really attractive counterparty contracts, right? Now we're financing that.

So for plus 225, what a massive contraction of cost of capital, right? And we have this flexibility now of financing through the credit markets, of financing through equity, of financing through prepayments. Like we can choose based on the needs of our client as well, because if there's something I'm certain of, it's that the needs that the client is going to change, right? Like this reliance upon prepays as financing. of course we can accept it. It's just we can't build a business that relies upon it.

And the other thing on the credit or the debt financing side, I want to hit on before I forget about it, it's the incremental level of institutional diligence that comes with that process, right? Let me have a second to make this point. Equity investors, all they really see at the end of the day when a large GPU contract assigned is a TCD, right? Like 10,000. million dollar contract it's sent, right? They don't see any of the detail in there, like, what the delivery of that infrastructure actually looks like? What are the SLA penalties? Like, what happens if they don't bring up at a certain time? What are the cancelability clauses that are in there? There isn't a lot of detail that sit in these contracts. And a lot of detail

that really matters to inform like just how much value is there in that TCV. In other words, like, how realistic is it that you're actually to extract that full contract value? And when we go through this debt raising process, they see those contracts, right? They see every single detail that sits there. They see every detail that sits on the data center contract as well. Is that data center going to be available to take the GPUs to turn it into revenue? Right? So you're getting this to a really intensive incremental level of diligence that's occurring.

That just benefits the equity investor because an equity investor can look at our backlog and RTCV and say, I know that those contracts have gone through an excruciating level of diligence by people whose job it is to not lose a single dollar. Yeah, exactly. The debt world is not allowed to lose money, right? They're only going to underwrite things that they have a high degree of confidence and of execution and ability to make money not only for the lenders, but, you know, the business. We have to make money as well, right?

It's not good for lenders for Parenthood or Corweave to not make money. That was back to the 25% contribution margin that's out there. So I feel like it's just been a bit underappreciated of just how beneficial it is for the equity investors, for us to have built such a sophisticated financing mechanism that is able to interact with the debt markets at an increasing scale and a decreasing cost of capital. You do make a good point there because you're right. I mean, these are these credit guys. They're not, they're not just going to be lending anyone money if they're not, you know, sold on the contracts. They're doing their deep research on it. So that's a good point. I want to move, move along and talk more about like some of the other things that were mentioned in the earnings call, which was like the useful life of the GPUs. I think that's been a lot very surprising is like some of these older GPUs like, you know, that are three, four, five years old are still maintaining their value. And in fact, some of them are selling for higher. and what they were when they first came out. Let me see how I want to ask this question.

It's a fun one. Yeah, what I'm trying to understand is, would you expect that to happen? And is this, I guess, how much longer will these older GPUs maintain their value? Because, I mean, new ones are coming out all the time. You know, if Jensen gets on stage once or twice a year, has new products coming out.

So how much longer can these older GPUs maintain their value? Or is this just a function of like, the supply constraint that we have right now and like kind of like what happened with COVID where used car prices were being sold for higher than what they were when they were new. Yeah. Yeah. I think the easy conclusion is it's just supply demand

because without looking in details, which admittedly details are not readily accessible out there, but without looking at the details, you say, well, people must be forced to use these older generation of GPUs. What you see within our platform every day with clients coming in is clients asking for these older generation GPUs. They want Amper, they want Hopper. And they want those platforms because those are the platforms that are most efficient for the workloads that they're running. I think that there's been this misconception that everyone only wants the latest generation GPU.

and that there's one GPU to rule them all, right? And there's one model to rule them all. And why would you have anything different than that? But the reality is something entirely different. The reality is that there is a very broad array of different sizes of models. And those different sizes and models all efficiently pair with different sizes or different generations of GPUs. Right? Like we have guys who will come in for Ampier and will say, we only have hobbies.

hopper available right now. And they'll say, well, we don't want hopper. We only one amp here because that's the most efficient GPU for our workload. Right. And I think that we see that all the way from enterprise to AI lab to hyperscale client. I believe that's the piece of information that the market is missing is that people want exposure to these different types of GPUs. And accordingly, it is completely changing the expectation.

of useful life. We depreciate over six years. A number of our peers depreciate by them to have six years somewhere in there. We've always thought that that makes a lot of sense. What we're seeing today, though, and as you mentioned in our Q2 earnings, we touched on this, we signed a three-year A-100 lease. That pushes that A-100 useful life out to the end of 2029. And by the way, we're signing that skew at the same prices that we were signing in early 2025 as well. Now, the way to think about that, that means no pricing degradation on the A100, which is a 2020 skew from 2025 to 29, puts it in nine years of life. We're seeing that across our platforms. And by the way,

you've seen that in the cloud historically as well. I think in AWS, they still have Teslas and Volta's online. Those are late 2010 skews. It's not like they're writing those things unprofitably. Like they're making money off them. They have demand for them. You can pull it down otherwise. So I think that this whole debate around useful life being two to three years, it's just not based in any semblance of empirical data. Like they might be based in hopes and prayers for whatever their thesis on the market is, but empirical data for a long time at this point, as supported at least six years of useful life. And it's going for it. Well, I think the thought process is that, well, AI, the AI space is changing so fast.

There's new stuff coming out all the time. And I think people think back to like what happened in the 80s and 90s where these new CPUs were coming out so often. And that the old one, the old generation would be useless after six months to a year because this new one was coming out. It was so much better. But that's not what the case is with AI because these old ones are useful for other workloads.

And not everyone needs a Ferrari when it comes to running their AI models. They can use an older Honda Civic or whatever the case may be. And that's what's going on with. with the AI's chips. Now, what I would say is really important is, regardless of which skew the client is on, they're getting access to the same core technology platform that sits on all of this.

And the core of technology platform is this incredibly robust management suite that ensures uptime of GPUs. It allows for inference to be run on GPUs. It hardens the platform to the points where I think we have the best TCO, like total cost of ownership in the industry, or close to it. It's another way of saying, like, we have the most efficient platform for serving all of these skews on the market. And, I mean, you should look at our engineering org. It's one by some of those brilliant people have ever met in my life. And what they bring to market into problems they solve for our

clients every day is a, is only. by having access to this full suite of AI cloud infrastructure that stands from CPUs to memory, to storage, to GPU infrastructure, to high-performance fabrics, to a distributed set of data centers is what we have 51 data centers in operation. Global, like, and we think that that makes sense because inference is going to be distributed global. Like, you're not going to run inference out or just like one or two sites somewhere. Like, you're for, to bring in a enterprise adoption to be true global cloud adoption, you need a global platform. And that's what Corey viz. So you mentioned inference. That's one thing I wanted to ask you about is like, you know,

we're seeing this change right now in the industry where, you know, a couple years ago, it was all about training models. I think that's one reason why everyone wanted the cutting as chips from Nvidia, because those were the best for training these massive models. Now we're moving towards the world of AI agents. And that's why CPUs are becoming more important. That's why agentic workloads are, I think they've overtaken AI training workloads. Are you seeing that on your end? And how does that impact your business? We're seeing that every day.

Inference is exploding. Inference is the monetization of AI, right? I mean, the investment in AI was training. And CoreWeave was known as a go-to platform for training for years. Because training is exceptionally difficult. Stabilizing a 60,000, 80,000 GPU fabric. and enabling those types of training runs enormously difficult and takes a incredibly robust suite of software solutions that we've developed and are entirely core proprietary to keep these infrastructure online available for our clients. But what we're seeing now is this shift towards inference for our clients.

And the use case has changed, right? Clients aren't asking us for inference only or training only infrastructure. They just want AI infrastructure, right? And AI infrastructure can do everything. AI infrastructure allows you to run inference for 18 hours during the day as inference demand is peaking in whatever region you're in. And then at night, it allows you to run training and fine-tuning work loads on your platform. That's interesting.

So there's actually not a difference between like how a data center or a cluster is made for inference workloads versus training workloads? Not on the corey platform. that might be in the industry. There's some inference only or training only platforms, but we build AI infrastructure. And what we're observing is this kind of loop process, an AI loop that we refer into in our recent earnings call

where clients are consistently shifting workloads from training into fine-tuning into inference and just going around and around as they collect more and more information. Right. And we expect that that's going to be, you know, the new norm. across AI. And again, I think we're just widely recognized as having the best platform to enable the scaling of that AI loop. Yeah, I think that's, I think that's the thing that I'll be

keeping my eye on is like, what is like this explosion of inference going to do? How does that change the economics of the AI cloud business? What does it do to like, does it require changes to these data centers that are halfway under construction right now? Does it need to make any, do you need to make any modifications to it? And I mean, it seems like, for, For core weave, that's not the case. You guys make AI infrastructure, so it's, it works for all types of use cases. That's right. And, you know, a little on that point of like, what has changed in the data center?

It's liquid cooling, right? Like, that was the biggest shift in data center construction. Because data centers previously were like, no water inside the day sphere, which makes sense. But you can't just like, you know, drill holes in walls and walls and put pipes through it and say it's liquid cooling. Now it's kind of a first principle of engineering solution. with liquid cooling. And we have been very focused on all of our data center operations to the liquid cooling

for a while at this point. What else happens in the data center from here? They just get more dense, right? You're moving up the rack density of like how much power is going into the rack. But liquid cooling is incredibly efficient within those platforms. That's, we don't expect for other like major data center shifts. within the next few generations. Gotcha, gotcha, gotcha.

Speaking of data centers, I'm going to end with two quick questions. One, kind of a serious one, one more fun. The serious question is, like, you know, there's been a lot of backlash now when it comes to data center buildouts throughout the country, throughout, you know, throughout the world. There's talks of like data centers in space that kind of go around some of the backlash here. Does that impact your, how you guys are thinking about data center build out,

your projections moving forward? I mean, how are you dealing with some of the recent, commentary out there. Yeah. This doesn't impact our projections going forward, right? Because at the end of the day, demand for AI is only increasing and is not being impacted by whatever may be happening at a state or county level. Demand is still there. I think all that really changes is where does it get built, right? Space maybe right? Space maybe. I think it's a big literal lift to get to space. But like this infrastructure is going to come online because the demand there. The ROI is there, right?

And the ROI for our clients is massive, right? The ROI for us is incredibly attractive on these investments. I think it makes sense for states and counties to want to understand the development that's happened, but it's not going to stop the development on a national or global level because the demand profile is there. And thus it's a where does it get built question rather than will it get built question? Yeah, I think that's going to be the key question over the next six to 12 months. Last question, fun one.

So again, you guys started off as a crypto mining company. Do you still keep track of crypto prices at all or is that just not? Because are you still keeping track of Bitcoin on your home screen? Like what's your relationship with crypto these days? Yeah. I mean, legitimately, we started the business in 2018. I think we were focused on crypto for maybe 18 months, right? By mid-2019, we were all in on developing and fundraising for cloud.

Do I still follow crypto? Of course, like, it's such an interesting asset class. I think Bitcoin became this institutional quality asset. It's pretty amazing to watch something have scaled to there, but all my attention is focused on on core weave in building the infrastructure that drives the creation of intelligence. I mean, you guys probably pulled off one of the greatest pivots in corporate history going from, you know, crypto to AI cloud and timed it perfectly.

So congrats on that. And congrats on all the success. I appreciate you making the time today. I think I learned a lot. I think our audience had learned a lot. So hopefully we'll have you back on soon. And, you know, we'll do an update on what's going on with Coreweave in, you know, six to nine months.

Appreciate that today. Thank you. Thanks so much, Brennan. Well, all right, guys, hope you enjoyed that conversation with Brennan McBee. I thought this was a really interesting one because AI infrastructure is one of those areas where the bull case and the bear case are both pretty compelling. There's no question that demand for AI compute is exploding, but there are also legitimate questions around debt and financing and competition and the supply and what the business will look like when the market eventually matures. Let me know in the comments of what you guys thought about today's conversation. Did you buy the CoreWeave story or do you still have concerns about the Neo Cloud business model?

Drop your comments on Spotify and YouTube. And while you're at it, consider giving us a five-star rating as well. You know all that engagement really does help us out and it helps other people find the show. Thank you guys so much for listening, watching, and commenting. Shout out to Mike for all the work behind the scenes. And we'll see you guys back here tomorrow. Two and five Canadians will hear the words you have cancer. That's why every step and dollar raised matters.

On September 19th, join thousands in Toronto for the Princess Margaret Cancer Foundation Walk. Challenge yourself, friends, and family to walk 21 kilometers in support of life-saving research. Together, we can carry the fire and help create a world free from the fear of cancer. Register today at pmcf walk.ca.ca. Thank you.