AI companies are hiding more debt than you think | Ed Zitron

I'm just a slight rant here, which I know is atypical for this show. The natural logical end point of what I'm saying is that this is one of the greatest capital misallocations in history, to quote just Dario, great analyst there. And also we are building hundreds of billions of dollars of risk directly into our retirement systems, and I don't just mean private retirements. CalPERS, the California pension system, is in Blue Owl. Teachers' pensions, cops' pensions, firemen's pensions, Dutch pension funds, they are in AI data centers. The risk is everywhere. And like I said, I have tried to disprove this. I've really tried because I don't want to just be a vacuous doomsayer. Otherwise, I'd be no better than the AI boosters. But I've really looked, and the best people can give me is they say, 'Well, I found a document that says that there's going to be all of this demand in 20.' It's just people dreaming. People dreaming so they don't have to face the nightmare of reality.

Greetings and Oracle. And it may explain why those reporting earnings this week are actually sort of kind of being punished in the stocks despite growth.

It's actually a little bit worse than that. So the Nik story is really interesting because you can read it one of two ways. It suggests it's either $1.65 trillion of total debt with 300 billion off balance sheet, or it's double, it's just 3 trillion. And what's great is we don't know because it's off balance sheet. So the way it works is that Meta is a good example. Their Hyperion data center, there's a thing called an SPV, special purpose vehicle. It's basically just a little entity they set up. And they, in this case, have a kind of SPV called a variable interest entity. Very complex way of just saying a weird little entity that they don't have control over. Now, to be clear, we're referring to Meta's data center, a 27 billion one that they will be the only customer of. They're paying for it. They will pay the lease. Pretty sure they're paying the GPUs, though I think that might come out of the debt. But because Blue Owl bought 80% of this venture and Meta bought 20%, it is considered not Meta's controlling stake, and thus it's off the balance sheet. Now, I must be clear, there was a Wall Street Journal story that came out that said that Ernst & Young had noted this was a red flag. Now, don't worry though, Ernst & Young still said, 'Look, sounds good to me. Moving on.' Just what is the point of an auditor at this point? But it is much worse because most data center deals are SPVs. Most of them are VIEs. So, think of it like this. When you do a data center and you raise debt, the SPV itself is what keeps the money. The money flows into the SPV, this little company that buys the GPUs that has the debt. And when the customer, when the data center is theoretically built and all of the expenses have come out of the SPV, the revenue from the customer that may or may not exist will flow into the SPV and then pay out to investors. The problem with this is if the revenue doesn't arrive, and the other problem is that in many cases these are non-recourse loans, which means that you can sue the SPV, you can take all the SPV stuff, you can't hit the holding company at all. Now, there are exceptions to this. Meta, Hyperion, they are on the hook for the payments. They promised to make bondholders whole. Core, technically, their SPVs are non-recourse, but you can sue Core eventually. But the SPV puts this barrier between the company and the investor. So investors could sue, investors could liquidate. But what happens? Let's say, I don't know, random idea, if you have thousands of these SPVs, multi-billion dollar data centers, and let's say half of them go tits up, and suddenly they all have to dump those GPUs, cuz they will. They're going to have to liquidate the assets. What do you think happens with supply and demand when a massive wave of supply goes? I don't think it's good for prices. So we just have this kind of private debt bubble. I think it's the equivalent of the subprime mortgage crisis, but for data centers. You've got, I think, what, $178 billion worth of these deals last year. Hundreds of billions, I think half a trillion of debt is out there according to Bloomberg now, just sitting around. And all of these SPVs, they rely on customer revenue. They must be, like, investors will not make a penny if customers do not pay. Most of these are also pretty poorly underwritten, too. Blue Owl, apparently, when discussing investing in Stargate Abilene, they took 10 minutes to deliberate. I would not be surprised if that was one of the more deep due diligence missions compared with everything I've seen, because all of these data centers are being funded under the idea that they're stable infrastructure, that they're these reliable infrastructure projects like a power plant or a factory, when in fact you are buying a warehouse, theoretically if it ever gets built, full of these quickly depreciating GPUs. Best case scenario, we're talking six years. And also, the only clients you have, the only real customers of AI compute, are AI startups that were all unprofitable, or hyperscalers selling compute to them. So yeah, not great is probably how I'd say this is going. It's not brilliant. I mean, the debt being used to build data centers has a duration of, I think, on the very, very low end, five, but on the top end, sort of 20 years. But then the assets being purchased need to be replaced. I mean, being generous, every five years, but every few years.

Well, I don't think anyone thought about that. That's the, I think the answer is that I just don't think anyone thought about that at all. I don't think anyone said, 'Oh, what if that happened?' Because you make a good point, which is if you're investing in one of these project financing loans, you are getting paid back through the revenue of the plan. You're not getting a full payback in six years. And in six years, what will be past, what will be a Verubam, LMFO, Sleen, Dion, I don't know what Jensen Huang's going to call the next ones, but we'll be multiple generations worth of GPUs in theory. AI bubble will be burst by then, or even just what if there's just, if the client doesn't exist in four years, like what if the client doesn't, like that is the most reasonable question here. And while in some cases like Meta's, they have a kind of a backstop, like Meta will backstop it, what happens in five years? Because those backstops are going to go away. Meta also could just not take ownership of the data center. But the point is these data center debt deals are, it's kind of like a subprime loan in the sense that after a teaser rate, what, six months in, the rates increased and then everything became much harder to pay. It's going to be in this case a year or so in, or just a year in, and construction finishes when they realize, oh, customers can't pay. Oh, our customers found a better deal elsewhere because it's taken two or three years to build this Blackwell data center, or just, I don't know, we have to upgrade it in six years. With what money? It's just, are investors going to have to put in more cash? Is the, is the, is the data center going to have to, no one's really, and I've looked a lot, it's actually really difficult to find any documentation of these loans by design. I don't think they want people thinking about them too hard, but it is, though, it's a slightly different financial vehicle. It's the same thing as the subprime mortgages. It's people getting loans that don't really have any reason, like they don't really, they shouldn't, let me rephrase that. It's people getting loans that they can't afford, or at least they can only afford based on magical thinking. And to be clear, subprime mortgages were not just poor people. They were also richer people who could not afford their houses. Subprime referred to anyone who just had bad rates. They were just referred to anyone with a bad rate, really. And I mean, most data center loans are junk, junk, junk. They're crap. They're high-interest yields. But that's great for the investors in theory. In theory, because there's so much money in AI, unless you look at how much money there is, it's all going to go poorly.