Let's say it's a trillion or so of off-balance-sheet debt. That means that there's a trillion dollars in loans that we just don't — we actually cannot really quantify outside of media reporting. This is a corporate scandal. This is a huge scandal. This should be front-page news everywhere. This should be a shareholder riot. But because we live, as you put it well on my show, in this cult of worshiping the wealthy, we just kind of put it to the side. This is a significant risk to these companies, though. Like Google — we're talking just as Google's earnings came out, they narrowly missed on search revenue. Their other businesses are slowing down. It's happening across the board. So their businesses are going to slow down just as these massive debts creep up.
Welcome to Prof G Markets. I'm Ed Elson. It is July 23rd. Let's check in on yesterday's market vitals. The S&P 500 and the Nasdaq declined. The Dow was flat. Crude climbed above $95 per barrel for the first time in nearly six weeks as both the US and Iran escalated their attacks. The yield on 10-year Treasuries increased. And finally, Google and Tesla shares fell after both companies reported earnings. We will get into those reports later.
Okay, what else is happening?
The big AI spenders are increasingly turning to debt, and a lot of it turns out to be hidden. A Nikkei Asia investigation found that five tech giants — Alphabet, Microsoft, Amazon, Meta, and Oracle — are carrying roughly $1.65 trillion in debt that doesn't show up on their balance sheets. That's more than the $1.35 trillion they actually report. Meaning the debt we can't see is now bigger than the debt we can. Meta's off-balance-sheet debt is roughly three times its reported debt, and Oracle's has ballooned about 30-fold in four years. It's all legal, but it raises one big question: what happens if AI demand isn't as strong as investors are betting? Joining us to discuss this question, we're speaking with Ed Zitron, author of the Where's Your Ed newsletter and host of the Better Offline podcast. Ed, thank you for joining us.
I just want to give you some context. Yesterday I was talking about Oracle on this show, and I was talking about how they're increasingly relying on debt to finance their data centers and how it's becoming kind of borderline unmanageable, and that's why their credit rating is getting downgraded and their borrowing costs are exploding. They're kind of being sent into this downward spiral. Up until this point, though, it has been my understanding that the debt in the AI ecosystem has been relatively contained to a handful of companies like Oracle, but that it hadn't infected the bigger names, which is why I was so alarmed to see this reporting from Nikkei, which says there's $1.7 trillion of hidden debt that we haven't been seeing, that we haven't been looking at. You've just written a huge piece on this. You investigated these numbers. You did a sweeping analysis of all of this. What do we know about debt in AI, and how is it that so much of it appears to be hidden?
So there are several factors here with the hyperscalers. The reason they're able to do that is because they found this interesting accounting treatment that Ernst & Young claim is a red flag in Meta's case, where when you build a data center, it's not like you as the company say, 'I'm going to buy all the stuff, I'm going to get the debt for this, and here we go.' They make a special purpose vehicle or a variable interest entity, which is basically an SPV where you don't own as much. So, for example, with Hyperion, I think investors like Pimco and Blue Owl own 80% of it, and Meta only owns 20% of it, despite Meta being the only client, the person that's going to fill it full of GPUs as well. I think they're moving some across their assets despite them being the obvious — despite Meta announcing it and saying this is our data center, because they don't have full ownership of it. It's not counted as something they put on their balance sheet. They will have it, I think, within operating leases when it starts paying. But the thing is, this is the entire AI data center industry. Every one of them are these SPVs that people pull money into. The SPV owns the debt. The SPV often owns a lot of the risk, though, through non-recourse loans, which means that technically you have to go after the assets of the SPV before you go after the company. But nevertheless, these things own the risk, they own the GPUs, they pay investors not out of anything other than the revenues from the data centers, which leads to the important point of what happens if the revenue for the data center isn't there. How do investors get made whole? And the answer is they do not. And the only reason this is not considered a problem yet is it hasn't actually happened — like the data centers have not been built at the scale that would have to happen for the revenue not to flow in.
So just to back up to an SPV — what actually is an SPV? And this is important because you say that these data center SPVs are the AI bubble's equivalent of CDOs, which of course were the financial instruments that basically sparked the Great Financial Crisis. You say that that is our equivalent today. So what actually is an SPV, and why are they so popular?
So to be clear, a CDO is slightly different. It's holding mortgage bonds and stuff, but they are the same kind of financial instrument that will cause genuine harm to the markets in the world. So an SPV is basically a company. It's a company. It has people, people that technically are on the board of directors. You can do that, but really it's just a holding entity for money and stuff. And so when they build a data center, that SPV raises the debt. That SPV holds the debt. That SPV buys the GPUs, it builds the data center, and when the money flows into it, the SPV is the one that pays out to contractors, that pays the OPEX costs. And there's a cash waterfall even where it goes paying OPEX, paying investors, and then paying the people that own it in theory if there's more money than there should be. That is not going to happen. But nevertheless, these things are so dangerous because they are being sold as stable infrastructure like commercial real estate or residential, and just saying, 'Oh, it's physical land. It's good. It's great.' But the problem is, and the critical difference is, that data centers are nothing like a regular mortgage and they're nothing like commercial real estate. These are complex financial operations both in the construction and the ongoing costs. And then there's the important thing of you need revenue to actually get paid out of them. And because they're all project financing, which just means they exist to finance a project, there is no other money for them to pay people out unless, like, the company on the other end — a CoreWeave and Oracle or what have you — isn't on the hook for those payments. Basically, the client is. And if the client is, say, I don't know, OpenAI and can't afford to pay it, then investors who have invested in these SPVs, well, they're kind of out of luck. And the problem is when I say investors, I mean anyone who's involved in private credit right now. And private credit is well funded, as I'm sure we'll get into, by pensions, retirement funds, insurance funds, teachers' pensions, cops' pensions, all this different stuff. We are all involved in private credit whether we want to or not.
I just want to read you a quote from Amanda Yakob of Bloomberg, and you quoted this in your piece. She made this comparison to Enron. She said, quote, 'Enron exploited US accounting rules to hide from investors and lenders hundreds of millions in debt it had bundled into off-balance-sheet entities, obligations that contributed to one of the biggest corporate collapses in US history.' And what you are describing right now is that there is a trend among the big tech companies — Meta, Microsoft, Amazon, Google, Oracle — they are creating these shell entities, these SPVs, seemingly to hide literally hundreds of billions of dollars of debt that they are issuing in order to finance their data centers. In other words, that they're showing us in their reporting and in their earnings, like, here's everything we've built, but then they're hiding all of the borrowing that they took on to build all those things that they're kind of bragging about in their reporting. That does seem to be a very, very similar link. I guess, could you expand on that, on how this has come to be and potentially how this might unravel?
So I want to be clear, though, other than Oracle, I don't think these companies die. But I will say it's kind of hard — that Nikkei piece, it wasn't clear if they were saying there was $300 billion or $1.6 trillion of hidden debt. It was a very weirdly written piece. But let's say it's a trillion or so of off-balance-sheet debt. That means that there's a trillion dollars in loans that we just don't — we actually cannot really quantify outside of media reporting. But I must be clear, it's not just hyperscalers doing it. It's CoreWeave, it's Iron, it's Nebius, it's Nvidia. In some cases, it's whoever is building a data center is using these shell corporations as a means of obfuscating ownership and also obfuscating risk, as in basically handing off the risk to investors. And even though there are some that are recourse loans, even if there's a — up the chain, you could eventually sue the company. At that point, you would have sold off all the GPUs, so you'd be screwed either way. But the larger point is, yeah, this is a corporate scandal. This is a huge scandal. This should be front-page news everywhere. This should be a shareholder riot. But because we live, as you put it well on my show, in this cult of worshiping the wealthy, we just kind of put it to the side. This is a significant risk to these companies, though. This is a real — like Google, we're talking just as Google's earnings came out, they narrowly missed on search revenue. Their other businesses are slowing down. It's happening across the board. So their businesses are going to slow down just as these massive debts creep up. And let's say they have ways of getting out of these things. The investors will still be screwed. The investors involved will still be screwed. And if they cancel these projects, I imagine they'll pay off investors to a point. But anyone involved is going to take a loss, and then this spreads out. Like these are the — this is the good news. The fact that the hyperscalers have these — these are the good SPVs. These are the ones that I think might survive. The problem is every single data center is through an SPV, and every single investor who's invested in a data center has invested not in the data center itself but in the potential revenue it brings in. And the problem there is that is dependent on there being 15 or more times the AI compute demand than currently exists. It's genuinely terrifying when you start quantifying it. And people are way too flippant about this stuff. They say, 'Oh, it's not as bad. Oh, it's not as bad as the Great Financial Crisis.' It's still a Great Financial Crisis that this happens. It's still horrifying, and it's everywhere.
You write, quote, 'Much like a subprime mortgage, AI data center debt is being poorly underwritten, virtually uncollateralized, and issued to projects that have extremely low likelihoods of repayment, all based on flimsy information and hyperdriven mania.' End quote. Sounds like a provocative statement, but I actually think it's pretty much factual and correct. I think one thing that isn't in that statement, when we make the comparison to the Great Financial Crisis, is the fact that a lot of that poorly underwritten debt, that virtually uncollateralized debt, was obfuscated and hidden away to the point where ratings agencies and investors and traders didn't even know what was happening because it had been so well hidden in the markets. It had been so well kind of complexified. I guess my question to you: do you think that this is the same situation where investors, ratings agencies — I mean, we know that S&P just downgraded Oracle's debt, so clearly they have some semblance of an understanding of what might be going wrong here. But is it your view that this is not being priced in, that investors literally don't know what is happening in terms of the debt that is being issued among many of these companies?
So let me tell you the tale of private credit. So private credit in the last few years has gone on a tear, merging with or acquiring insurance companies and retirement funds. Apollo bought Athene — well, they merged with them. Blue Owl bought KKR — I forget who the others are, but nevertheless, Blackstone's infrastructure fund is funded by retirement funds. The problem is, is a lot of the SPVs are funded by private credit. That 'private' part is the lethal part. S&P doesn't rate private debt. No one does other than the private credit fund. So there was a story in The Information a few months ago that said that Blue Owl decided to invest, I think, up to $10 billion in Stargate Abene after 10 minutes. How much due diligence do you think we can get done in 10 minutes? Probably not very much. This is standard for the SPVs. So while people will say, 'Oh, well, the ratings agencies failed here,' actually, no. The SEC failed here. The insanity of private credit, which is basically a multi-trillion-dollar shadow banking system, is what's propping this up for the most part. The debt raised for these SPVs is coming from private credit. You look at CoreWeave, it's coming from private credit. When it comes from the banks, the banks themselves — SMBC and MUFG out of Japan — they're just funneling the debt straight in. So much of this is coming through institutional investors that are funding money through private credit that the ratings agencies aren't even involved. And the ratings agencies don't — the only reason they did it with Oracle is Oracle has predominantly raised its money via bonds, except for the fact that a lot of their private projects are SPVs, and the SPV raises the debt. And it's this situation where in theory, like I think the Michigan data center they're building is a recourse loan, so you can go directly after Oracle. But that's the thing — even in that situation, there are these legal barriers that if every data center is exploding will make it hard to litigate at scale, or at least hard to pull down who's actually responsible. And we don't — like, if you are anything touching an Athene or a KKR or any other insurance company, it is worth checking where the money is going. I think the Lions actually invested in the CyrusOne bond. And the thing is, these bonds get rated somehow, they get rated as junk, but because insurance companies are now effectively run by private credit funds, they don't give a crap. It's like, sure, because of this giant lie. And this giant lie is that data centers are the equivalent of investing in power plants and factories, that they're AI factories, when in fact what you're investing in is a very large, like, town-sized building or campus that is full of depreciating GPUs for an industry that has not proven it has the demand. And I mean, my estimate for actual AI compute demand is about $120 billion a year. I did the math, and it's something like if they build all 130 gigawatts of IT load — so the actual operational GPUs that they say is in planning based on Sightline Climate — we will need $1.68 trillion of annual compute revenue. We'll need that just to pay for them all. And they'll have to pay consistently, because if you stop paying consistently, the loan covenants with the SPVs break. And the crazy part about this is the global software industry is less than $800 billion. So we're just going to magic up another software industry. And to be clear, what I am saying sounds radical. I think you kind of said this already. It sounds like a scary thing. 'Oh, it's alarmist.' This is just very basic math. This is — Sightline Climate said 190 gigawatts of capacity was in planning or under construction. PUE of 1.35, which is just the energy efficiency, 130 gigawatts, $12 million a megawatt. It really is simple math that anyone could do. And the thing is, even if it's just half of that, we still don't have enough demand. And so we're in this weird situation where this risk is now spread everywhere to the point that because it's private, we actually do not know how far. And it's within the insurance companies. It's different to how AIG collapsed. It's just smaller. It's smaller little bits. But the other problem is that AI data centers are so expensive. They cost $500 million minimum, probably several billion. So instead of having millions of or hundreds of thousands of subprime mortgages that collapse in a kind of slow boil over time, it's going to be a $500 million, a $6 billion, a $2 billion, a $1 billion collapse. And each time one of these happens, that's a massive markdown with a private credit fund, but also a bunch of investors who have lost money and have no recourse beyond, I don't know, selling all the GPUs in a market that will become saturated with them.
Yeah, it's the recourse point that is getting us into such dangerous territory here. I mean, it's one thing to sell equity on a speculative bet about the future, in which case, you know, the people know what they're buying into. We know what the risks are. It's another thing to finance an extremely speculative business through debt and then to hide that debt, because I mean, if this doesn't work out, we're not just talking about equity going down here. We're talking about bankruptcies. We're talking about defaults, which —
Also, one other problem, which is the reason that insurance funds and retirement funds invest in private credit is because they need yield. They need ongoing yields. So if these data centers do not pay out, we have retirement funds and insurance premiums that cannot get paid. And I'm sure they have some buffer. I'm sure they have other assets, but I mean, it doesn't have to be an AIG commercial paper-level collapse for this to be systemically damaging. And after this, there is no more yield to be found. I guess other than Treasury bonds, which are going up now. Yay. But when they go down, I don't know. But so it really is, and what sucks is most people have no idea about this. Most people don't realize that, like, the California pension fund is invested in Blue Owl. Like, these are — it's everywhere. And it's not just in America. It's across the board. You've got, I think, the Dutch pension funds in data centers. CDPQ, which is the Quebec pension fund, they invested, I think, in a Core data center. Like, this is everywhere. And I'm furious at the fact that it's so rarely discussed, but also people are so quick to being like, 'Oh, it's not as bad. Oh, it's not as bad.' Not as bad is still bad.
All right. Well, we're going to have to continue this conversation another time. I'm going to have to let you go. But it is fascinating stuff, and I do encourage our listeners to go read your article. It's extremely rigorous and gets at a lot of the issues that we're talking about. Ed Zitron is the author of the Where's Your Ed newsletter and host of the Better Offline podcast. Ed, always appreciate your time. Thank you.
We'll be right back. And if you're enjoying the show, be sure to subscribe to the Prof G Markets YouTube channel at the link below.
We're back with Prof G Markets. Tesla just reported earnings, and Wall Street was disappointed. Ahead of earnings, Tesla had reported impressive delivery numbers for the quarter, up 25% year-over-year. That gave investors the impression that Tesla's worst quarters were behind it. This time last year, the company was reporting a 13% decrease in deliveries. But the company's second-quarter earnings told a different story. While revenue beat analyst expectations, up about 26% year-over-year, profits fell 5% over the same period. And the quarter's free cash flow came in about $1.1 billion in the red. The stock fell more than 3% after hours and is now down roughly 25% from its peak. So here to break down Tesla's quarter, we are speaking with Karim Bousta, co-founder and managing partner at DVX Ventures and former vice president at both Tesla and Lyft. Karim, thank you for joining us. The interesting story here: you got decent delivery numbers, especially compared to last year. Revenue was up, but profits down. What went wrong here on the bottom line?
Yeah, thanks for having me and for the great question. I think what we're seeing with these published results should not come as a surprise for anyone who's been understanding and following what's been going on with Tesla for the past, I'd say, 18 to 24 months. Essentially, what's been happening: we've seen declining sales for several quarters in a row, but more importantly, we've seen declining market share. And even in this quarter, where for the first time in a few quarters Tesla has been able to report increasing sales numbers, you have to look at this in the context of a growing EV market where the market overall has been growing faster than the sales numbers that Tesla has been reporting. So that's one factor which confirms what we've been seeing for the past few years, which is Tesla has a fundamental issue on the automotive side, which is essentially the vehicle lineup, the product lineup is becoming old. It has not been able to renew it. There's been a refresh of the Model Y last year, but it's still based on the same platform that launched almost 10 years ago now with the Model 3 and then the Model Y. So there hasn't been real innovation, both from a product side and technology side, that would have allowed the company to catch up on the market share losses that they've been seeing, and at the same time continue improving the margins. It's actually been the opposite. The margins have been eroding. We're seeing it in the results that have been reported this quarter. So all this comes down to the lack of innovation and new product and fundamental platform launches that we haven't seen for the past few years, and that's in the context of competition that has been increasing with all competitors catching up to where Tesla was even six or seven years ago and creating great products with better margins, better production systems than what Tesla has been able to accomplish. So all this to say, I'm not surprised by these results. It's all coming down to what we've been seeing in the past few quarters, and everything is coming into place.
