Go back

CoreWeave Bonds Yield 15%. Is the Market Wrong? | Jordan Chalfin (CreditSights) #26

1h 2m 4s

CoreWeave Bonds Yield 15%. Is the Market Wrong? | Jordan Chalfin (CreditSights) #26

CoreWeave, a pure-play AI infrastructure provider, operates through off-take agreements with hyperscalers and AI labs, delivering GPU-powered compute capacity in data centers. Despite short-term operating losses due to ramp-up costs and timing mismatches, the company is achieving strong revenue and margin growth, driven by high demand and aggressive pricing. A surge in prepayments—up to 100% of contract value—reflects customers’ desperation for compute, with 9 billion in prepayments expected in Q4, including 3 billion from re-contracting existing capacity. The business is financed via secured delayed-draw term loans collateralized by customer contracts and GPU assets, minimizing exposure to parent guarantees. While medium-term risks like overcapacity and customer renewal exist, they are not expected to impact operations until 2028–2030. Credit investors are mispricing the company due to technical volatility and negative sentiment, ignoring robust fundamentals, strong pricing power, and favorable unit economics (16% pre-tax margin over contract life). CoreWeave stands out among neo-clouds for its non-convertible bonds and solid financial structure, with credit spreads compressed from 1,200 to 450 basis points due to improved fundamentals and funding confidence. The company’s execution, including scaling to 1.7 GW of active power and 4.6 GW of contracted capacity, signals strong growth and resilience in the AI compute cycle.

Transcription

8295 Words, 46256 Characters

English
There's kind of a self-fulfilling prophecy here when the bonds reach kind of double digit yields, investors get concerned about the ability for core weave to fund. It's capex, so customers are kind of so desperate for capacity right now that they're willing to make a pre-payment on an old generation of a chip. So there's too much capacity in industry, why would they renew with core weave? I don't think those play out in 2007. I think that's more of like a 2021 to 2030 issue. Welcome to another episode of Fix and Floating. We are recording this on Friday, that's 13th of November. In tune, CDS ensuring investors against a default by core weave cost about 450 basis points. Last week, they cost more than 1200. On Tuesday, core weave showed a beat on revenue and margins and also raced its guidance. Next day, stock rally by about 20%, but now the stock is barely up. Today, we're gonna look at what the release means for credit investors. I'm very pleased to welcome Jordan Charlfinn, head of technology at credit sites. Jordan, welcome to the show. - Thank you for having me excited to be here and talk about core weave. - Jordan, would you mind giving our listeners a bit of background on your experience? - Sure, I am head of technology here at credit sites. I've been here 17 years. I cover both investment grade and high yield and a bit of distress, including core weave credits. Prior to joining credit sites, I spent a few years as an equity research analyst at Goldman Sachs, as well as a couple of years on the buy side as well. My coverage is, I would say, mostly investment grade and that includes all of the hyper scalars as well. - So core weave is getting a lot of hate at the moment and you came out with a buy recommendation on the bonds. After Tuesday, you reiterated the buy. What prompted you to do that? - Sure, well, the buy recommendation is really driven by two things. It's one, the 15% yields on the bonds. And two, the fact that I feel strongly that the fundamentals are going to remain very strong through 2027. There are medium-term risks here, in terms of potentially having too much capacity when the cycle turns and also customer renewal risk as well. But I don't think those are going to play out in 2027 and our recommendations are generally one-year views. I would say that the Q3 results were very supportive of my recommendation, both on the operating trend front. So there's a revenue beat. More importantly, there's a very strong margin beat and the pricing on new deals is really strong and that's flowing through into the implied margin guidance for Q4 as well. They also address a lot of investor concerns on the call and announce that they're coming with a new IG-rated facility. And I think that certainly address some fears about their ability to obtain funding. - What do you do say is the market missing in your view at the moment? - Yeah, I think, and it kind of transcends across both investment grade and high yield. But the market right now is just, it's very much driven by technicals and sentiment as opposed to fundamentals which we talked about are strong and actually improving. There's just been such a plethora of supply of tech issue into the sector and a lot of headline noise, whether it's like open AI, doing an IPO, there's just a lot of noise out there and a lot of skittishness I would say from investors. - In December, CDS peaked at around 880 basis points while the stock was well above 100 bucks. By June, CDS was down to around 450 last week. It was about 1200 stock materially lower. What would you say changed between June and now? - Yeah, so that's I think a lot of the things that we talked about, right? The all of the new supply that hit the market and tech, the delayed IPO from open AI, they had a failed acquisition of of course scientific. There's a lot of short seller criticism out there, including Jim Chanos and Michael Barry. And so I would say just the technicals and the sentiment have been really bad and it really hasn't been, it really hasn't been on the fundamental side at all, as pricing continues to move up and be quite impressive. I would say the other issue too is there's kind of like a downward spiral slash self-fulfilling prophecy here in that when the bonds reach double digit yields, investors get concerned about the ability for core weave to fund. - It's cat-backed growth and then that contributes to additional selling as well. - For listeners who are less familiar with what core we're doing, could you please give us a quick wrap-up of the business and especially what the key lovers are for the credit? - Sure, yep, absolutely. So it's actually a fairly straightforward simple business. They're pure play, NeoCloud. Most of their data centers are at least, they have to buy the GPU servers, which are very expensive to put into those data centers. And then what they do is they enter into these off-take agreements or compute rental agreements with customers that are either hyperscalers or AI labs or AI startups, some enterprises. And those are generally take-or-pay contracts that are four years on average, five years in duration. In terms of the key drivers here, it's gonna be the pricing on the contracts as well as the cost of those GPUs and the financing terms for those GPUs as well. - I think that contracted power went up by more than 900 megawatts in the quarter to over 4.6 gigawatts now. About 1.7 gigawatts are already running today. What is the difference between active power and contracted power? - Sure, so active power is essentially what's up and running today, right? And that's, as you mentioned, that's 1.7 gigawatts. Nearly all of that is gonna be chipped and at least out to customers. There's a small lag there when you get the active power, it takes a bit of time to put the GPU servers in and run them out, but nearly all of that is kinda up and running capacity. The contracted power 4.6 gigawatts this includes all of the leases that they have for those powered shells that have not yet commenced. So it includes both the active power as well as those leases not yet commenced. So there's essentially about the differences is 3 gigawatts that are signed and have not yet come online. And the amount that they've actually sold to customers or pre-sold to customers that shows up in RPO is gonna fall somewhere between those two figures. - Is that also one of the reasons why we seeing these very negative profits as a mismatch between, you already have to start up the facilities while you're not yet collecting revenue? - Absolutely, there's ramp up costs here. They kinda target like mid 20% operating margin on a contract, but it takes a couple months to kinda get to steady state production and go through like normal ramp up type costs. So as the business has done a decent job scaling but it's still like not yet at scale and that increase in active power has grown pretty dramatically. That's gonna have an outsized impact on margins. And we really saw that play out. I would say earlier this year, where margins were quite depressed, but they've rebounded quite nicely sequentially throughout the year and they're gonna be, you know, decently strong probably in that like 17% range for Q4. - One key concern, I would say what's also getting portrayed in the media quite heavily is delays with data centers. And I think there's also one concern with regards to Q4 with what could slow down bringing that contract that power really online are being able to harvest that expected revenue. Yeah, sure. There's all types of execution challenges that, and regulatory sort of challenges that, that, that neoclass and hyperscalers have to go through. There's, you know, in terms of procuring the labor and getting these things fully permitted. There's been a lot of the state and local pushback. I would say that the, what we've seen, like the blue wave outcome from the midterm elections, it's been a lot of pushback and moratoriums and things like that on, on data centers. And so all of these things are a challenge. I think the, the good news is the 2027 capacity is, um, for coral, but also for the industry is, for the most part, on their construction, for sure, right? And, and also, for the most part, fully permitted. So that's kind of the good news. We have to kind of wait and see what it does to future capacity beyond 2027. But I, despite all of the kind of pushback, data centers do bring in a lot of tax revenue. They do bring in construction jobs. And so we'll have to kind of wait and see what actually gets implemented in practice. Yeah, I think I also read somewhere that despite the negativity to getting data centers into your backyard, it brings also some really high qualified jobs into more rural areas of, of the US. So it's not all bad. Um, another key thing I want to judge about is on take or pay agreement. So my naive read of take or pay is that the customer pays for the capacity whether they use it or not. So it doesn't really matter for core weave, how much they actually use. Is that how the contract work? Or are there any other considerations? Yeah, absolutely. I mean, 90% of their, their business was take or pay, right? They essentially doing nothing or very little on demand. Um, and what that means is is, it is contractually committed that the customers like, um, can't back out. Now, there are certain like service level agreements and could be, you know, material breaches or whatnot. If there's a, if there's a delay, like a material delay and when that service comes online, or they're not, or the uptime, is it there? The customers could, uh, could walk. So there are, you know, exceptions, um, clearly written out, um, in those contracts. I think that in the, um, in the corner environment in the current market, if anyone, if any of their customers walked, they could re-contract that out, um, at a higher price. And so customers generally like aren't walking, um, we saw that a little bit, I think with Microsoft and some of the other new clouds, but generally speaking in this, in, in the current environment, um, customers are not walking, uh, away from our contracts. Um, on the, on the flip side, if the data center, um, developer has, you know, if that, if that development of data center is material delayed, we can also potentially walk from those obligations. So that, that is kind of like a mitigating, mitigating factor. Can, uh, customer, um, get out of contract, obviously, not super interesting for them at the moment, but let's move into 2028. Yeah, I mean, if it, if they've signed it and it's take our pay, they can't, they, they would only, uh, be able to get out of the contract if there was a, uh, like a breach on or a failure to deliver or a failure to provide availability, um, from, from core weave. What happens if core weave delivers late? Um, if it, if it is materially late, like beyond what's in the terms of their contract, um, the customer can walk. But if, if they did deliver late, um, it's probably, it's not going to be because of availability of chips, right? Because I, it's, it's going to be because of availability, uh, or, or, uh, inability to bring up, bring on that powered shell. And if that's the case, I think that core weave, um, the customer could walk, but core weave could also walk essentially. If you like this episode, subscribe to fixed and floating. You won't miss anything, and it really helps the show. And now back to the conversation. Could we classify because there's a breath of neo-clouds, you have the hyperscalers, um, who are bringing on data center capacity. Can we classify those players into who is more sophisticated, also who is more experienced in developing data centers? Um, because it just seems to me that they are all, everybody's just talking about this whole, um, combined mass of, of players while it seems to me there must be at least some differentiation. Yeah. No, they're absolutely is. Look, the hyperscalers are going to be highly sophisticated, right? They've been in this business a long time. Um, clearly Amazon and alphabet and Microsoft are going to be the most sophisticated in the market. Um, there's a mix of leasing and owning going on there are an alphabet, I think tends to own a bit more in some of the other players, you know, tend to lease a bit more, but they're all fairly sophisticated in terms of bringing on capacity and managing it. The neo-clouds, a lot of them are, you know, former Bitcoin miners, they've kind of, you know, core even included too, right? I mean, these were natural gas traders that pivot to Ethereum mining that are now, you know, doing AI cloud infrastructure. And so, um, I would say across the neo-cloud space, there's, there's generally, um, a bit less experience and less sophistication there. However, core values is for this so long among the neo-clouds, like they've already kind of delivered, they're delivering at scale, right? They're at 1.7 gigawatts of active power. And that continues to scale. Um, whereas some of the other neo-clouds have plans, you have lots of contracted capacity, but they don't necessarily have a lot of active capacity. So, I think that's the differentiation there. Is like, I would put Corleaf more into that sophisticated bucket. Um, but like for the broader neo-cloud space, um, there's a lot of trial and error, I would say going on right now. Um, on top of, um, development capabilities, it also seems to me that there's also some differentiation under software layer. Um, is that something which is differentiating for core weave, um, and does that help margins, or is that only something which helps on the margin? Yeah, I look a little bit. I'll give you kind of both perspectives here. And this is like, so Corleaf will say that, um, it's MFU, which is like model flops utilization, um, is, is higher than the industry. Um, and, and the reasons for that is they're pure play in AI infrastructure, they've kind of designed the software stack specifically for GPU as a service. And so they don't have as much, you know, technical debt per se, you know, relative relative to the high hyper scalar. So they've done things like remove the hypervisor layer, um, and other things on the software side to make it specifically tuned, um, to this AI cloud business. Um, they, the tend to get good pricing in margins. And they also have gotten like pretty good industry rankings as well. So like, for example, the semi analysts, um, cluster max rankings, they got like the platinum status here. And so there are some proof points. Um, however, kind of like broadly speaking, they're, they're, they are competing against very sophisticated players like we talked about, all the hyper scalars. Um, I don't think there's a sustainable advantage here. Um, I would argue that this is more of a commodity. And at the end of the day, um, this, the winners are going to be the ones that have large economies of scale and like an ability to execute. Um, they're doing a decent job on both fronts, but you know, I don't, I don't necessarily think there's like an inherent or dirt, durable advantage on the, on the software side for them. Microsoft, I think, is around more to thirds of revenue last year. And it's now, um, a bit more over one third, open AI and then tropical also large parts of the backlog. I think backlog rose to over 130 billion. Um, most of that come quite early in the quarter. So we haven't seen that many new deals. Um, how much of the backlog depends on customers that are not investment grade? Yeah, I would say kind of most of it is not investment grade, right? And so it's really kind of their investment grade customers or meta is, is, is the biggest customer, um, that they have. And Microsoft has, has, has, was when from being the biggest customer has now kind of like relatively, you know, smaller relative to meta, um, and, and, and, and open AI. And so I would say, you know, the majority of the, the customers are, are not investment grade. Um, they're starting starting to get more enterprise business, enterprise customers, but that is still fairly small. It's AI labs and AI startups and hyperscalers are kind of the big buckets there. Understood, and enterprise customers are also more spot buyers, I would say, or short-term buyers. Yeah, they tend to be doing the more like three-year contracts, so that's what the recent like the recent facility, they did the 5.5 facility, was specifically targeted towards enterprise customers, and they tend to buy in like two or three-year contracts, they don't think five years out. If you think about pricing, I think management said that four to five-year contracts now sign a 20 to 25 million per megawatt a year, and a short contract is up to 40 million. The existing book was closer to 10 to 12 million. Why is the pricing curve so inverted? Yeah, sure, so we're in such a tight market environment right now, and so I think inverted pricing is actually quite normal for tight markets, and so there's a shortage of AI compute, and that's really driven by the shortage of power shells, and so customers are willing to pay up for that capacity to meet customer demands, right, and to grow their revenue and satisfy customers. It's really about customer attention. So, even if it's hurtful to margins, even if it's un-economical, there's reasons to that customers want it. There's also kind of a view that like, hey, just because we're supply constraint today doesn't mean we're going to be supply constraint three, four years from now, right, and there is a lot of supply coming on, right, we had about 45 gigawatts of data center capacity in the US at the end of 2025, that's probably like tripling over the next several years, and so I don't think the current, we can extrapolate the current market dynamics that far in the future, and then just kind of generally speaking, the way that the cloud business works is the longer term committee you make, and the larger committee you make, you get kind of more favorable and better pricing for that. What do you think about the operating business model and think about one specific project, project? So, Corviv mostly doesn't own its own data center, at least the land, and the power cell comes then from developers like Core Scientific, and then it puts its own chips inside. Could you explain that relationship with the landlords? Yeah, sure, so the developer finances and pays for the data center, and for like a one gigawatt data center, that could be somewhere on the order of $15 to $20 billion. They lease that out to Corviv for a long term lease, that could be like 15 or 16 years. It's generally about $2 billion a year, roughly in rent, plus energy, you're at at least $500 million a year, so from a cash cost perspective, Corviv could be out, call it $2.5, $3 billion a year, between data center rent and energy, and maybe some other maintenance costs as well, but that's just the power cell, like Corviv is responsible for buying all the hardware and installing that into the data center, and that hardware, and that hardware has gone up over time, right, so for like Hopper chips, it was over $25 billion and Black Wall was over $30, and now Vera Rubin, it's over $35. So it's really like the tech hardware CapEx costs, it's like $35 to $40 billion on bleeding edge and a video chips for one gig of capacity, so it's way more expensive than the actual data center, despite being a short live asset. So we're talking around $60 billion for the whole package, what I've been personally appealing about the credit is that it's essentially project finance of a holding company on top, so each box holds a customer contract, chips and a loan secured on both. So could you take one data center and walk us through how it is financed from the customer contract down to the loan? Yeah, sure, so most of Corviv's funding is done through these delayed draw term loans, that's the asset level funding that you talked about. Those are collateralized by the GPUs, they're collateralized by importantly by the customer contracts, as well as the leases, and in most cases they also have a Corviv parent guarantee. And so Corviv generally can finance like 70 to 90% of the cost of that hardware using those facilities. Those are delayed draw term loans, which means the money doesn't get borrowed until Corviv needs it to buy the CapEx, and they're also advertising. So as Corviv collects the payments from its customers, it first gets used to pay down the loan, and then access amount gets dividend back up to the parent. For listeners who aren't familiar with the term, what is a delayed draw term loan? Yeah, so basically it's just the loan, Corviv enters into an agreement with the lender, but the borrowing doesn't happen immediately when the loan is signed, it's delayed draw, which means that Corviv bars the money when it needs it, it's basically when it has to outlay the cash to buy the GPU servers, they'll draw down the loan. Yeah, these loans are also advertising, so Corviv is paying down essentially principal and interest, they're using the customer payments to pay down the loan over time. And so if you think about it, most of their contracts are five years, and the loans are designed to kind of match that maturity profile of the contracts, and so when the customer contract ends, those loans will be fully paid off. How much of the delayed draw term loans have recourse to Corviv itself? I'll solve them, so the exception is the meta facility, this is their 4.0 facility, this is an investment grade customer, and they basically didn't need the parent guarantee from Corviv's, because you're contractually committed by meta, essentially. So and I think generally speaking, the most valuable part here is the collateral, right? That's the customer contracts, and it's the value of the GPUs, whereas the parent guarantee is sort of like the less valuable aspect, and so if you have a customer like meta, you don't need that parent guarantee, and so that facility is non-recourse. Corviv does have another IG facility coming, so it's very possible that one looks like the meta one as well, and the meta one I think is 8.5 billion dollars, so it's a rather large, you know, low in size, once that's fully drawn upon. What happens if, let's say, a non-investment grade client is unable to fulfill its obligations, and essentially defaults on the deal it has with Corviv. The ability and the topic goes passed in two years, and they can't pay for the compute. What happens as an afterthought? Yeah, so there's a few different types of collateral here, right? There's the collateral from the, it's backed by the customer contracts, but it's also backed by the GPUs, and it's backed by the at, by the leases too, so there's like the right to use the leases as an asset. So there's lots of different ways for creditors to get their money back. They could, they could re-contract that out to another customer. They could sell the GPUs, they could monetize the leases, and they also have a potential claim on Anthropic as well, through a hypothetical default type scenario. But if that doesn't work, what happens, what is their claim, their recourse towards core weave? Correct. So for the facilities which are mostly recourse, they have a parent guarantee, and then core weave would have to cover the shortfall. So core weave is on the hook, you know, outside of all, if the critters can't get made whole, like from all that collateral they have in those different avenues that we talked about, they would look to the parent guarantee to obtain, to get like additional funds to help cover the creditors in that box. This year core weave collected prepayments on 70% of its deals, some of them up to 100% of the contract value. So they got essentially all the contract value in advance, which is obviously a big cash inflow now, and it expects about 9 billion in prepayments and deposits in the next three months, so another big cash inflow. What do the prepayments do to the economics of a data center? Yes, so I think the market conditions are extremely tight right now, and so I think that a lot of meal clouds and hyperscalers included, like Oracle, are seeking out prepayments and demanding prepayments, and I think a lot of customers are happy to do it. And so I think that it could change the economics somewhat, right, a customer that's making a prepayment, obviously wants something in return for that, but I think first and foremost in this extremely tight market environment, customers are just desperate to get the capacity and making a prepayment is essentially the cost of doing business right now. So most of the clients, even the hyperscalers are just struggling for compute, and they are very happy to secure the core weave potentially, doesn't need to give up that much in pricing and margin power at the moment. Correct. Correct. In the current environment, that's right. If something goes wrong, does the customer who prepaid rank ahead of the bond holders? So this is kind of a tricky question, a customer prepayment is like an unsecured claim, but I think it really depends on where that prepayment was made, and is it a prepayment to customer parent is the core weave parent, or is it a prepayment to the other link box that the wholly owned subsidiary, right, that the customer has the actual contract with? So I think I'm not entirely sure where that would rank. If you're looking for more credit-related content, check out fixandfloating.com, and now back to the show. The investment grade facility in March, I think, priced at Silver Plus 225, there was one in August at Silver Plus 550, so materially higher. The next one should be investment grade, as you mentioned, yielding about 7 to 7.5%, financing more than 90% of the CapEx, so at the top of the range you gave us earlier, how much does core weave actually make on a contract once it is financed like that? So what's really the equity return on that investment? Yeah, so the kind of the unit economics that they've laid out is over the lifetime of a contract, it's kind of like 24% operating margin, and 16%, kind of a pre-tax margin, after all those interest payments for the financing. So that's kind of the unit economics. I think those have maybe a bit of rosy assumptions in them, including the cost of financing. And so maybe it's a bit lower than that, but the offset here is that the pricing, recent pricing has been really strong, they've gone through a number of price increases. On the recent call, they talked about, I think you mentioned it earlier, like the 20 to 25 million dollars per megawatt on a new 4 or 5-year contract, and so I would say those two things kind of offset, and that's in the current environment maybe the eating economics are holding. You mentioned the 16%, is that a return on asset or a certain return on equity, so only at a 10% they are chipping in? That's kind of the sort of like a pre-tax margin, is the way that I think about that. On the whole asset. Over the duration of the 5-year contract, essentially. Now there's potential oxide or downside to that, so depending on, because they own the chips once those are fully paid off, and so if they could re-contract out that capacity at reasonable rates, they could see upside to that figure, but if they can't find a customer to renew that capacity, then there's obviously a lot of downside. But for the initial 5-year, that's kind of where it lands. That makes totally sense. I think you just mentioned being able to put a new contract on the existing chips. One thing we haven't talked about is that they're still under hook for another probably 10 years on the Powered Shell. Is that an issue that they need to pay another 10 years to billion a year, which kind of is very deluded for the equity? Yeah, no, it's kind of a major issue, right, that this is a pretty material medium term risk for correlative that we could potentially overbuild data center capacity, particularly in the United States. It's just given how many of these like Gigawatt plus projects that are coming online over the next several years. I think one of the issues is that some of their customers are the hyperscalers, and they like Meta and Microsoft, they own their own data centers, and they have long-term leases themselves on data centers. There's too much capacity in the industry, why would they renew with correlative? I think there's a material medium term risk that there's just unruly so there's too much capacity out there. If you talk about cost of capital and especially the cost of capital driven by delayed term loan rates and in the past, unsecured financing at the parent level, is there a point where a new contract doesn't make sense? Yeah, I know it's a good question. I would say there's kind of a few factors besides the cost of financing. It's also like the cost of the chips and the pricing on new contracts, so depending on where all those things are, at any given time, I don't think they've kind of reached that point where it doesn't make sense. I think we've seen them pull back from the unsecured market right in double digit yields, and so I think that kind of gives you an indication of they want to be at like 10% or lower in terms of cost of debt financing. In the corn and market, there's maybe a bit more wiggle room, but I think that as long as they kind of stay under that 10% level, it makes sense to continue to move forward and given the demand that we're currently seeing. We touched upon operating margins before. There were about 10% disc water, net loss was still about 700 million USD. Could you explain for listeners why the company still loses money? Yeah, so there's a few issues here. One is the ramp up costs, right, because it's, you know, it takes a few months to kind of get profitable for each new equipment that you install. There's takes, takes time to kind of when you start paying the lease to get it installed and start, you know, collecting from the customer and then wrapping up it. efficient matter. And since they're not like a huge business yet, they're just basically a pure play. And they're kind of gaining scale, as we speak, there's sufficient ramp up costs. I think the other issue, when you think about a free cash flow perspective, there's a huge timing related impact here where, currently even the whole business, you have upfront catbacks and ratable revenue. And so I think historically, it's probably taken close to like four years I would estimate, to get to that free cash flow cumulative break even. With the recent pricing where it is, that's obviously a price, maybe it's like two to three years. My numbers are more conservative, right than the numbers you're probably hearing elsewhere, including overhead and things like that, and not talking about extrapolating short-term contracts at 40 million a megawatt, but that's kind of how I think about it. So there's the ramp up impact on margins, but there's also this timing, really painful timing related for cash flow burn. That doesn't mean it's not profitable or cash generating for the life of the business, but it's a pretty major headwinds as you're buying that catbacks. Despite that timing related impact, they now guide to a positive net, a trusted net income next year. What has to go right? - I think they need to execute. They need to continue to scale in terms of their active power. So if they see material delays to that, that's going to be an issue. They need to kind of continue to assign new customer contracts at profitable pricing and margins. - On Tuesday, management said that they won't come to the bond market because it's current yields, it's an economic for them. What does it mean for bondholders, especially as more and more funding is coming now ahead of them? - Yes, so most of it, they're primed right. There's a lot of, they're layered essentially behind a lot of secured debt and that secured debt is going to continue to increase. And the unsecured debt is going to probably be more flatish because they've said that they're not coming to unsecures. They're basically shut out of that market with 15% yields anyway. They've already done so much in terms of comfort, so I'm not sure how much more they could do there. So I think most of the debt going forward is going to be at that secured level. So it's not great to be layered behind all this debt, but keep in mind, all of that secured debt basically sits at those boxes. They are self-amortizing as the customers pay their monthly bills. Those are used to kind of pay down the debt at those boxes. So it's not like the secured debt is sitting at the parent level. - If you think about debt in general, I think it increased by about 12 billion in a quarter to around 47 billion and gross leverage is now around eight times EBDA. On consensus numbers, the company still burns about 60 billion of cash flow from 27 to 29. Could he give us a split of all the financing forms where that money should come from? - Yeah, so we talked about like that 78 and 90% long to cost for the delayed draw term loans, so that's going to have to be the bulk of it. Instead of doing on secured debt, they're going to lead more heavily into customer prepayments. The way about prepayment typically works is let's say you enter into a five-year contract, maybe the customer pays you up front for your five. So you could kind of get what you need between a combination, if your loan to cost is 80%, and then you're getting like a 20% prepayment that kind of gets you most of the way there. Now they do have overhead and other costs and things like that and like not every contract has prepayments. And so they do need additional sources of funding. I think the other area that they're going to lean heavily into, I would say is the OEM financing, which has started to grow, but I think they could do more there. Specifically, Dell is a major supplier of AI servers and to Corey even, and the whole industry, and they've got a really strong balance sheet, Dell's leverage is well below its own target. So there's a lot of room there to extend more financing to customers and help continue to fuel their own growth. And so I think, you know, Corey Lee is going to take advantage of that, in my opinion. - I think that's a really strong point. I think that the general issue as well is that the two biggest uses of compute and dropping an OEM AI, they don't have a strong balance sheet, right? So to facilitate AI build out, you need these other very strong balance sheets to facilitate that 130,000 GPUs came of contract in 2027 and another 70K in 2028. Most of the data center leases run much longer. What happens once such a contract runs out and Corey is still on the hook of the lease? - Yeah, so they need to, there's a few different options, right? They could re-contract that out to the existing customer, they could re-contract that out to a new customer, they could kind of serve it up as demands, you know, on demands, there's a variety of options that they could do once those contracts expire. The good news for them in the current environment, the stuff that is expiring is actually being re-contracted out at pricing that's higher than their original rates, which is kind of remarkable. That's like not a normal situation, right? Like these things depreciate over time and as the technology advances. - So even for older chips? - Yeah, for older chips, they're re-contracting out at higher rates than they had originally. And I think what's really was a, I think of a powerful kind of comment that that management made on the Q3 call was they talked about the 9 billion of prepayments that they're getting in Q4. Three billion of those prepayments are for re-contracting out existing capacity. So customers are kind of so desperate for capacity right now that they're willing to make a prepayment on an old generation of a chip. That there's no query doesn't even have any CAPEX associated with it. There's like no, they already own the CAPEX. The debt is fully paid off on that. They own the GPUs and customers are willing to prepay even for their older generations chips. - Codech has about 90 billion of future lease payments for listeners who are not familiar with leases. What exactly are future lease payments and why use an undiscounted number? - Okay, and the number has actually increased a bit. It's about 110 billion as of Q3 from most recent, 10 queue filing. And so the way leases work is when the leases actually commence and the power shell is made available to core we use, it comes on the balance sheet and it comes on the balance sheet at present value of those future lease payments. And that's for both operating leases and financing leases. Now in addition to what's on the balance sheet, there's also leases that have not yet commenced. Those are off balance sheet and they're typically disclosed in the case and accused. Those are not yet discounted to present value. Those are undiscounted or future value of the payments. What I like to do is get kind of a total undiscounted number to keep things simple. What's the total amount of future rent that they have to pay? And then we could use that to estimate future rent payments. We could compare that with the RPO, which is 133 billion. So I think it's a lot more simple to kind of think about the undiscounted total amount of future lease payments. - Does the circular financing worry you? - Yeah, for sure, I mean the circular financing, this whole industry, this whole AI build out has been kind of perpetuated by circle. are financing, think back to the original deal between Microsoft and OpenAI, where they invest in OpenAI and OpenAI, essentially agrees to use them exclusively. It's not exclusive anymore, but that was kind of the first circular deal and almost every deal since then has had some circular component. Clearly, you have NVIDIA driving a lot of these circular type investments. They're an equity owner, they've provided equity for CoreWeave, and they've also provided it backstop on JPUs, because there's a certain amount of capacity that if CoreWeave doesn't sell to other customers, they'll step in and buy it, but it's only like $6 or $7 billion. It's not that meaningful relative to the $133 billion of RPO that they have from customers, and the substantial amount of more RPO that they're going to need to fill up all of their contracted capacity from all their data centers. The other kind of thing I think about as it relates to this is CoreWeave knows it has this backstop, which means it just kind of emboldens them to be a bit more aggressive, so I'm not actually sure that it provides much safety in the way of creditors. You just mentioned that CoreWeave actually needs to sign a lot more deals to essentially fill up all the contracted compute. You should have a bit more light on how we can think about the sales activity of CoreWeave in the future, and how that makes it really a sustainable structure, I would say. Yeah, so we talked about contracted capacity at like 4.6 gigawatts. They don't disclose how much of that is sold to customers, but we kind of like do some financial gymnastics to try and estimate that figure. We do have the backlog of $133 billion, so we can estimate what's the average duration on that and try and determine how much is actually contracted. I've estimated probably there's two gigawatts or so that's not yet sold to customers. I think the comments on the Q3 call kind of support that, because the company said that they haven't released start to sell 2020 capacity yet, and they're waiting for that to season further, which kind of makes sense, because pricing is still going up. They are also trying to kind of time when they sell the capacity, so it doesn't make sense for them to sell everything now because you still have this inverted curve, but a pricing is still pretty high. Yeah, the pricing is still kind of moving higher on new contracts. So that's one reason, and that's certainly the reason that they gave. The other kind of thing to think about is each new year, there's like a new generation of chips, and we don't necessarily know the pricing of those chips yet, and we don't know the efficiency yet of those chips, and so it's easier to kind of sell to customers the existing architectures, because that's how these things are priced, they're priced on what type of GPU, and then the duration, essentially, and the amount. So it's just, I think, a bit more challenging to sell future generations when we don't know the price or the efficiency or the energy efficiency as well. Is bringing your own chips becoming a thing there? So we have, for example, unthrobbing coming with its own chips, only using the Shell Corvieve has leased, and they're essentially only providing the software layer and essentially managing the facility. Yeah, so I don't think Corvieve has done any of these sort of bring growing hardware type arrangements, you know, Oracle is certainly doing some of them, and I think it really depends on the chip architecture as well. And so if you have like, if you have an open AI has its own custom ASICs like jalapeno, and maybe they're getting funding there from an industry partner, similarly with TPUs. I'm not sure how much of this kind of bring your own chips that we're seeing on the video side of things. In some ways, Corvieve looks pretty close to a read, you've long leased on one side and contracted rent on the other. How do you value a company like that? Yeah, so this is a, I don't know, I don't necessarily, I don't know if I look at a read because they're kind of like on the other end of the lease where they are, they're essentially the tenant, and they're basically look like a cloud provider, just a very specific type of cloud provider. You know, it's hard to value a company like this because it's so extremely capital intensive, and I think that certainly could create some boom bus, particularly on the data center or leasing side of things. And so depreciation is just such a high part of their costs. Is EBTA the right measure when chips are replaced pretty much every few years? Yeah, look, at the end of the day, I don't think it matters so much what metric you're using as long as you're hand the capping it appropriately. So if you do use EBTA, I would kind of make sure it's somewhat modest multiple, given how capital intensive it is, and how much depreciation there is in the business. You could look at an EVD operating profit, or you could kind of look at free cashflow on a normalize basis, like over the duration of a customer contract, and value it that way as well. What are the agencies looking at? I mean, bonds are still rated double B minus. Bonds are clearly showing a lot more stress than where the ratings are. I think the market is looking at some of these risks, the agencies are, I think are taking kind of a weight and sea approach. So they haven't, they're not going to necessarily downgrade just because the market is panicking. I think that they're waiting to see if fundamentals change, if there are any kind of real signs of stress here, for core weave, if they can't get the funding they need, or fundamental standards, things like that, but clearly there's a pretty large disconnect here between the ratings and where the bonds are traded. How do the bonds compare with other AI credit, like for example, or actually, or the other Neo Clouds? Yes, so core weave is the only one that has non-convertible debts, bonds among the Neo Clouds. I would say there's like signs of stress throughout the industry or call, you know, trading around 300 basis points, that's a triple B name that's trading at the whole way, right? So not that kind of different from from core weave, which is like, you know, a B-rated credit that's trading kind of like a, like a triple C or somewhere in between. So, but like, you know, Oracle is 300 right and core weave is like 950 or something like that. They're just, you know, a very, both showing signs of stress, both for very similar reasons, but Oracle has outside of this like essentially Neo Cloud business that Oracle has, they're broadly diversified business, they already have like, there's no really incremental overhead, the corporate overhead to support it and they've got database, they've got applications, you know, they have traditional cloud infrastructure. So they just, you know, it's an substantially lower leverage. So just both showing signs of stress, but like different, you know, one is IG and the other one is high yield slash distress. rip up what does the market get wrong about Corvie right now? Yeah, I think the market is just very swept up in the technical pressure and the negative sentiment around AI. I think that the fundamentals are strong and they're actually improving. I think Corvie has done a nice job executing well among all the various neo clouds. There are medium-term risks, right? You know, one, I think we talked about before, like they might have too much capacity when the cycle turns and then two, there's customer renewal risk as well. But I don't think those play out in 2007. I think that's more of like a 2028 to 2030 issue, right? So it's going to take a few years for that to happen. And so I think that, you know, fundamentals remain quite strong and are actually improve in the near future. And then also I would say that I felt and continue to feel that they'll probably get the funding that they need in the near term just because it's mostly based on the collateral and the customer contracts. It's less reliant on that parent guarantee. Jordan, that was great. Where can people find you and learn more about the work you do? Thank you. Appreciate that Joseph. You know, all of my research reports are published on credit sites.com. It's a platform for credit investors. I cover, you know, all the hyperscalers as well as Corvieve, Hardware OEM, software companies. Listeners can also find me on LinkedIn. Jordan, thank you very much for coming on the show. If you enjoyed this one on Corvieve, please follow and subscribe wherever you get your podcasts and see you next time.

Podcast Summary

Key Points:

  1. CoreWeave’s bond yields have surged to 15%, driven by market sentiment and technical factors, despite strong fundamentals.
  2. The company’s business model relies on long-term take-or-pay contracts with hyperscalers and AI labs, generating stable, high-margin revenue.
  3. Strong pricing power and a significant increase in prepayments—especially on older chip contracts—reflect extreme demand for AI compute capacity.
  4. While medium-term risks like capacity oversupply and customer renewal exist, they are not expected to materialize until 2028–2030.
  5. CoreWeave’s financing is largely secured by customer contracts and GPU assets, reducing reliance on parent guarantees and enabling strong credit quality.
  6. Despite short-term operating losses due to ramp-up costs and timing mismatches, the business is on track to achieve positive net income in the next fiscal year.
  7. The market is mispricing CoreWeave due to technical noise and sentiment, ignoring its improving fundamentals and resilient revenue structure.
  8. The company’s ability to re-contract existing capacity at higher prices signals strong customer demand and pricing resilience.

Summary:

CoreWeave, a pure-play AI infrastructure provider, operates through off-take agreements with hyperscalers and AI labs, delivering GPU-powered compute capacity in data centers. Despite short-term operating losses due to ramp-up costs and timing mismatches, the company is achieving strong revenue and margin growth, driven by high demand and aggressive pricing. A surge in prepayments—up to 100% of contract value—reflects customers’ desperation for compute, with 9 billion in prepayments expected in Q4, including 3 billion from re-contracting existing capacity.

The business is financed via secured delayed-draw term loans collateralized by customer contracts and GPU assets, minimizing exposure to parent guarantees. While medium-term risks like overcapacity and customer renewal exist, they are not expected to impact operations until 2028–2030. Credit investors are mispricing the company due to technical volatility and negative sentiment, ignoring robust fundamentals, strong pricing power, and favorable unit economics (16% pre-tax margin over contract life).

CoreWeave stands out among neo-clouds for its non-convertible bonds and solid financial structure, with credit spreads compressed from 1,200 to 450 basis points due to improved fundamentals and funding confidence. 6 GW of contracted capacity, signals strong growth and resilience in the AI compute cycle.

FAQs

High yields reflect investor concerns about future capacity oversupply and customer renewal risks, not a reflection of weak fundamentals. The market is driven by technical sentiment rather than actual financial stress.

Customers are making prepayments on existing contracts due to intense demand for AI compute capacity. This provides immediate cash flow and helps CoreWeave fund chip purchases and operations.

CoreWeave uses delayed draw term loans, collateralized by customer contracts and GPU assets. These loans are structured to match the 5-year duration of customer contracts and are primarily non-recourse.

Potential overcapacity in the data center industry and customer renewal risk are medium-term concerns, but the speaker believes these won’t materialize until 2028–2030.

Strong pricing on new contracts and significant revenue beats have boosted margins. The company is now in a more stable phase after initial ramp-up costs.

This represents total leased capacity, including 1.7 gigawatts of active power. The remaining 3 gigawatts are not yet operational, indicating a significant pipeline of future revenue.

Chat with AI

Loading...

Pro features

Go deeper with this episode

Unlock creator-grade tools that turn any transcript into show notes and subtitle files.