The Company at the Heart of the AI Bubble  ## Metadata - Author: [[Decoder with Nilay Patel]] - Full Title: The Company at the Heart of the AI Bubble - Category: #podcasts - URL: https://share.snipd.com/episode/0cfc309c-3907-43fb-ad92-262951cab43f ## Document Notes ## Highlights - **Crypto Pivot Fueled A NeoCloud Rise** - CoreWeave pivoted from Ethereum mining to leasing NVIDIA GPUs as AI demand rose. - It built a NeoCloud business that rents specialized GPU compute to large AI customers. Transcript: Nilay Patel CoreWeave was founded by former commodities traders as an Ethereum mining firm out in New Jersey. And crypto. Like so much of the tech industry, when crypto sputtered, CoreWeave pivoted to AI and started to use the same GPUs it was stockpiling to mine Ethereum to build its own data center operation. CoreWeave's big innovation is creative financing. It raises huge amounts of money to build data centers full of NVIDIA GPUs, and then leases those GPUs back to big AI companies to feed the industry's seemingly insatiable demand for Compute. That's turned CoreWeave into a pillar of the obscenely expensive AI infrastructure bill that's turned markets inside out over the past year. CoreWeave itself went public this past March, and it's worth almost $50 billion today. And just this year alone, CoreWeave has signed multiple contracts worth tens of billions of dollars, with companies like Meta and OpenAI for access to its data centers. But what Liz has found is that the story of CoreWeave is a lot more complicated than just a few Wall Street types making some smart bets and getting filthy rich. In fact, the narrative around CoreWeave, that it's just making the picks and shovels of the AI gold rush, is built around some very complex financial maneuvering, as well as a unique Relationship to NVIDIA itself. ([Time 0:01:38](https://share.snipd.com/snip/7b3cbc28-4e16-4395-b098-fd2790493119)) - **Engineering And Energy Are Its Moat** - CoreWeave's advantages are connectivity expertise and access to energy rather than unique software stacks. - Those advantages let it charge premiums but may not be durable long-term. Transcript: Liz Lopatto A little more sophisticated than that, but not a lot more sophisticated than that. Because it turns out the way that you connect the chips really matters in terms of being able to get good performance and good uptime out of them. And it also turns out that it's really hard to get energy. And that is a major supply constraint on AI right now. CoreWeave has a pretty good track record of getting energy, and they also have relatively good engineering and can charge a premium for renting out their chips, essentially. But it's not a huge mode. It's not necessarily a durable mode either, because if they lose the engineering advantage they've got, then they are basically in a situation where they're competing on price, and That's not a situation where they can win. Nilay Patel That's really the heart of it ([Time 0:06:08](https://share.snipd.com/snip/5ba79736-c2c2-4211-8fe9-4bbbff8704ed)) - **Crypto Roots And Perfect Timing** - CoreWeave originated in crypto and timed a pivot after Ethereum mining profits collapsed. - Many crypto miners similarly shifted into AI when model demand surged post-ChatGPT. Transcript: Liz Lopatto Is crypto. Yeah. Of course, it's crypto. It's crypto. So these guys pivoted basically after the crypto boom in 2022 away from crypto, because in 2021, Corwe have made all of its money by mining Ethereum. So they basically timed things correctly to really ride the AI boom. And, you know, one of the things that you see in the AI space that strikes me as a flag is that there are a lot of people who pivoted from crypto into AI because they thought it was going to Be more profitable. And you may remember one of the reasons why NVIDIA was starting to have this incredible run was because all of these crypto companies were buying its chips. And NVIDIA was like not the most enthused about that because... Not at all, actually. No. I think they even nerfed one of their chips. But they didn't love crypto. And when AI started to really get frothy, especially after ChatGPT was released, I think that was something that felt better to NVIDIA and to a lot of other people than, yeah, we're going To mine Bitcoin. And so, like, Corweave was in the right place at the right time to really ride that. Nilay Patel I understand that abstractly, right? If you're NVIDIA, you don't want to be in a race to do literally a single calculation to mine Bitcoin or to participate in the Ethereum blockchain you'll get beaten by ASICs or other specialized Compute over time. And maybe you just don't want to participate in that. And you certainly don't want to reduce your ability to supply other markets that you think are more valuable long-term. If you're CoreWeave, maybe you do believe that that is the future and you've gotten really good at buying and connecting GPUs to build data centers for specific tasks. AI comes along, NVIDIA says, this is the thing. This is the multi-purpose compute solution that our chips are made for that can, if you believe it, can accelerate the economy in various ways. This is where the growth is going to be. And then you look at CoreWeave and you're like, well, you're already pretty good at building data centers full of our GPUs. Now do it for the thing that we want you to do it for. And then you just arrive at CoreWeave exists. It's ([Time 0:08:51](https://share.snipd.com/snip/22247b84-a47c-4f9b-b307-0534de2becee)) - **NVIDIA Is Both Supplier And Backstop** - CoreWeave depends heavily on NVIDIA for chips, capital support, and demand guarantees. - NVIDIA backstops revenue and acted to buy shares during CoreWeave's weak IPO moments. Transcript: Liz Lopatto Just not possible for CoreWeave to exist without NVIDIA. And we can get more into that in a minute. But just like thinking about how this company works, if you look at the S1 filing, they say that they own more than 250,000 NVIDIA chips, which is big. And that's, you know, infrastructure that's necessary for running AI models. And in fact, it only has NVIDIA chips. It doesn't have any others. So on top of that, you have NVIDIA, which is an investor in CoreWeave. When, you know, the IPO faltered in March, NVIDIA swooped in and bought shares. NVIDIA has backstopped CoreWeave in a number of ways, including saying that they will buy any excess compute that isn't sold to customers. So they're essentially providing a guarantee of income, which is remarkable. It's like having essentially, you know, a knight, right? Like NVIDIA is the king and CoreWeave is their knight. And, you know, so NVIDIA is protecting CoreWeave in many ways, and CoreWeave in many ways seems to serve NVIDIA. ([Time 0:11:02](https://share.snipd.com/snip/6522649a-d670-47fd-b681-6df708ddc3f2)) - **CoreWeave Carries The Buildout Risk** - CoreWeave absorbs build and obsolescence risk so hyperscalers can use overflow compute without owning assets. - That risk-transferral is central to its business model and appeal to big tech customers. Transcript: Liz Lopatto It's taking on some of the risk of building data centers. And, like, that is an undertaking. Construction projects, as I think everybody knows, often don't happen on schedule for reasons that have nothing to do with how good people are at building things. Sometimes if the weather sucks, you just can't build. That's just a thing that happens. And so there is inherently risk in building stuff. And then on top of that, if you think about chips, they frequently and very quickly become outdated. And so if you are doing a big spend on this generation of chips, and then NVIDIA comes out with a new, better generation of chips next year, you're doing another big spend. That is also something of a risk, right? Being in this business, there are these assets that depreciate. And so CoreWeave is taking on those risks and Microsoft is just renting their compute. It doesn't have to worry about, you know, is this data center going to be something that I need to continue to operate forever? Are these chips going to be outdated next year? None of that matters. It's just, oh, this is a place where I can put excess business. ([Time 0:15:27](https://share.snipd.com/snip/6f712bc8-d015-45b8-a2ac-688811af44c7)) - **Debt Fuels Rapid Buildouts — With Risk** - CoreWeave and peers rely heavily on debt to accelerate GPU buildouts. - Fast buildouts maximize chip useful life but increase bankruptcy risk if AI revenues lag. Transcript: Liz Lopatto Impulse is no, but I mean, we're going to find out. I could be wrong. It wouldn't be the first time I've been wrong, but I look at CoreWeave and I see how much debt they've taken on and it really makes me nervous. There's been sort of this ongoing theme, I would say, in the last couple of months where we've seen a lot of debt being taken out in the markets for this AI out. And like, there's good reason why you would use debt rather than profits, mainly that you can get a bunch of money really quick. And that lets you do the build out really quick. And then potentially, you know, that matters for the lifetime of the chips. Because if you do the build out really quick, then you get the maximum lifetime out of these chips. And maybe you make your profits faster. Like, that's the hope, right? Nilay Patel So the chips piece of this is the other thing that just constantly breaks my brain whenever I talk to you about it or I read anything about it. CoreWeave is taking on tons of debt. It's taking tons of loans. It's collateralizing those loans with the chips it's bought from NVIDIA. The payoff of those loans is that the chips will do extraordinarily valuable workloads. But the chips wear out, and AI as an industry has not yet proven that it can do a bunch of really, really valuable workloads. There's a chronology problem here that suggests to me that AI won't pay off before all of the NVIDIA chips literally wear out, that OpenAI will just do unprofitable Sora video generation At such rate that the chips will wear out before someone invents an AI workload that doubles the profits of Citibank or something. I don't see how those timelines actually align such that this business can succeed. Is there a better thesis out there than that? Liz Lopatto Well, so this is the thing that I'm really curious about. And obviously, drop me a line if you know, because the chip depreciation cycle really matters here. ([Time 0:18:34](https://share.snipd.com/snip/a705caf2-a7c4-4f6c-9976-b3e5badfb1bd)) - **GPU Depreciation Is A Key Unknown** - Lenders accept GPUs as collateral, but depreciation and heavy use create uncertainty about their value. - The unknown lifespan of GPUs is a critical unknown for loan structures and the sector's stability. Transcript: Nilay Patel CoreWeave is taking on tons of debt. It's taking tons of loans. It's collateralizing those loans with the chips it's bought from NVIDIA. The payoff of those loans is that the chips will do extraordinarily valuable workloads. But the chips wear out, and AI as an industry has not yet proven that it can do a bunch of really, really valuable workloads. There's a chronology problem here that suggests to me that AI won't pay off before all of the NVIDIA chips literally wear out, that OpenAI will just do unprofitable Sora video generation At such rate that the chips will wear out before someone invents an AI workload that doubles the profits of Citibank or something. I don't see how those timelines actually align such that this business can succeed. Is there a better thesis out there than that? Liz Lopatto Well, so this is the thing that I'm really curious about. And obviously, drop me a line if you know, because the chip depreciation cycle really matters here. And there seems to be a fair amount of disagreement about what it actually is. For those of you who follow Michael Burry, who you may recall from The Big Short as the guy who nailed the timing on the financial crisis, he's arguing that the depreciation schedules That we're looking at right now are, in fact, too rosy. That may be true. You know, I've barely scratched the surface here, but the thing that I have found is that a lot of things vary, right? So where the location of the chips are, what kind of infrastructure is around them, how they're being serviced, all of these other things, how heavy the use is, there seems like there Are a lot of variables involved in how quickly these chips wear out. And so that, to me, is one of the central questions about whether this works out or not is, is this good collateral? There's one way of looking at this where these chips are effectively the same as a truck. And we have vehicles that get loans all the time and trucks depreciate. We have a pretty good schedule for the depreciation of the truck for your business. And so you can get a loan with the truck as collateral, and that's not really an issue. But because this particular kind of financial product is so new, it's sort of difficult to figure out what's going on with the chips. And so to me, that is one of the central questions here about, will this succeed? How good are the chips? How long did they last? ([Time 0:19:20](https://share.snipd.com/snip/05d614a0-b491-40ee-8698-55b6e9f838d6)) - **Timing Mismatch Between Loans And Adoption** - CoreWeave must hit fast AI adoption to service large debt piles; slower, normal tech adoption risks failure. - Historical tech rollouts suggest AI's full commercial adoption may be much slower than current financial bets assume. Transcript: Liz Lopatto The way. This is why the debt is important because the chip loans are only one kind of debt that Corweave has taken out. There's a revolving credit line with JP Morgan. There's some vendor financing involved. There are secured loans with investors. All of these types of lending have different structures. And depending on when the loans come due and how this is structured, that's going to affect Corweave's timeline for when they have to have the money. Maybe as long as they have access to credit markets, they can extend that some, right? Like they can take out a loan to pay off the interest on the other loan and give themselves a little more runway. But eventually you start to look at the company and you think, given the size of these loans, we really need to have a very fast takeoff for AI. That is built into the assumptions here. And so if you imagine AI is a normal technology, like a big and important technology, but a normal one, that takeoff isn't going to happen that way. ([Time 0:22:30](https://share.snipd.com/snip/2a96adc8-59b8-4299-bb64-d782624abeee)) - **Circular Financing Masks True Demand** - Circular financing (money flowing between investors, customers, and vendors) inflates apparent demand and worries investors. - Examples include Microsoft–OpenAI and stock-for-capacity deals that move value in loops. Transcript: Nilay Patel Explain what circular financing is. Liz Lopatto One way of understanding circular financing is that essentially a company, let's call it OpenAI, is receiving billions of dollars from a company before sending that money back to Pay for computing or other services, right? So if we're thinking about OpenAI and Microsoft, Microsoft has invested in OpenAI. They have a partnership and OpenAI is using Microsoft servers and I assume paying for them. So that has made investors really nervous. And sort of a lot of the focus here has been on OpenAI specifically, where Microsoft a certain amount of funding and then OpenAI sends back a certain amount of revenue. And I think that figure is $13 billion in both directions, which is a clever way of financing is one way of looking at it. Another way of looking at it is maybe more cynically, you're passing around the same $5, like when you and your friends go out to eat. And, you know, if you think about OpenAI's arrangements with CoreWeave, they have, I think, a $22 billion contract at this point. And CoreWeave has given OpenAI $350 million of company stock. That is circular. There's some amount of money that is just going the whole way around. These are not the only deals where you're seeing this. I'm picking on OpenAI because they are the most, I think, important company in this space for a lot of people. ([Time 0:28:37](https://share.snipd.com/snip/674c4397-9b5f-4724-b825-e199e2b84e76)) - **Track Demand And Hyperscaler Capacity Closely** - Watch demand for inference and cheaper model breakthroughs as signals that compute demand will shift. - Monitor hyperscalers' capacity expansions because they can absorb excess supply and harm NeoClouds. Transcript: Nilay Patel I think one of the big ones is demand. Liz Lopatto Right now, most of the demand for compute comes from training models and not necessarily from inference or like people using the models. If the demand for inference doesn't tick up quickly, that becomes an issue. Also, let's say we have another deep seek moment where there's a relatively light model that doesn't demand a lot of compute that is really powerful. That changes things, right? It's really difficult to figure out what demand is going to look like next year, let alone five years into the future or however long you need to know in order to justify these data center Buildouts. And if you are a big company, like let's say Amazon, one of the things you can do is you can say, look, we're betting part of the company on this. We're going to build out these data centers. And if this AI bet doesn't work out the way we want, it turns out we have other parts of the business. We can use the chips to, I don't know, be more AWS. They might be overpowered for that, but it's fine. We'll just stop our CapEx for a couple years and eat it. They can do that. Meta has an advertising business. Google has an advertising business. All of these companies that are much larger than CoreWeave that are participating in this build out are in some ways insulated from the risks of overbuilding. Because if they overbuild, they have other businesses that can just stick in those data centers and sooner or later they'll be fine. ([Time 0:31:57](https://share.snipd.com/snip/680f2413-6761-4563-a046-c39f7a7c74b0)) - **NeoClouds Extend NVIDIA's Market Power** - NeoClouds functionally act like special purpose vehicles that amplify NVIDIA's market control. - They lock customers into NVIDIA chips while giving NVIDIA leverage over hyperscalers. Transcript: Liz Lopatto A lady of a certain age, the first time I ever heard of a special purpose vehicle was Enron. For those of you who are maybe less familiar with the Enron scandal, there were a bunch of things that Enron did that were fraud. One of them was they created special purpose vehicles and used them to hide their debt. The thing about fraud is that the part that's illegal is the lying, right? It may not actually be illegal to create a special purpose vehicle and load it up with debt. As long as you tell everybody that's what you're doing. Looking at Corweave and the other Neo clouds and sort of unfocusing my eyes a little, I kind of wondered if one might think of them as an exercise as being essentially like special purpose Vehicles for NVIDIA. Like, obviously, they have other investors. But NVIDIA is putting an amount of money in that company. We know more about CoreWeave because it's public, but it's not just CoreWeave. It's turning around and using that to secure a loan, leveraging up, and then buying more NVIDIA chips. And for NVIDIA, this is actually a pretty great arrangement, not just because somebody's buying a bunch of their chips, but also because it gives them leverage over their biggest customers. If you think about your Microsofts and your Amazons, they've been talking about creating their own chips to help lower their costs that they would then preferentially use in their Data centers. So not only are you creating more competition for your own chips, you know, jacking up the price, you're also effectively creating a market where your chips are the default. So if Microsoft has its own chip, but everything is meant to run on an NVIDIA chip, that chip is not going to be as useful. So that's part of what I see here. ([Time 0:33:44](https://share.snipd.com/snip/0aa6de66-f98f-40b9-9e1d-9401fcda4741))