“Banks Are Herd Animals” | Ed Zitron on the AI Debt Slow Boil

Everyone expects correctly that there's some big explosion, but generally what happens — and it's happened across the railway bubble, happened in the dot-com bubble, happened in crypto bubbles, it's happened everywhere — it's always a slow temperature increase over time. It's always this sense that things are just getting worse and worse and worse, and then things start popping off. And even in the worst kinds of debt, which AI debt somehow isn't in yet, the distressed market — so the CCC, the worst of the worst debt — that's showing the most stress it's seen since 2023, the last bank crisis with Silicon Valley Bank. When you start hearing the distressed debt screaming, that's when stuff is getting nasty. So, all of this is to say is people look for explosions and they should. That's fair. What you should also look for is the frog in the pot, the slow boil, because that is what is going to bring this to its knees.

Good to be here.

So the closest I've got to a compelling counterargument is that it won't be as bad as I think — that there will be some kind of smaller AI industry after this ends, and that OpenAI and Anthropic could plod along cutting their training costs and so on and so forth. This would be a lot more compelling if it wasn't for the fact they had about $1.3 trillion worth of compute commitments and we have, what, 200, 250 gigawatts of supposedly planned AI data center capacity. If this was the argument I was brought at the beginning of last year, I'd actually buy it. I'd have been like, 'All right, that's somewhat reasonable.' But these companies are not being sold as piddly 300, 500 billion dollar companies, which sounds like a lot, but not when you've got Anthropic saying in their prospectus that they're going to have a total addressable market of $30 trillion, which is more than the US economy, about $10 trillion less than the US debt. So, a lot of the arguments against my work now are basically it will be all right on some level. And I could buy some version of that, but not for OpenAI and Anthropic. I just, I still don't think that's what happens to the AI industry in general, because running inference, so standing up the services, it's kind of like running an airline. I think I've mentioned this before, that you rent the GPUs, you get your 10,000 GPUs, and you say, 'I'm going to have this many customers.' You get too few, you're losing tons of money. You get too many, terrible service, you're losing money. So right now, the only reason that these services are getting funded is because we're in a hype cycle. We have people saying, 'Okay, I will pony up 300 million, 500 million in funding so that you can blow it all on cloud compute. Brilliant. Go ahead.' How is that going to work when the bubble bursts? And no one really has a compelling argument other than 'nah'. I'm serious. That is about the prevailing anti-Zitron argument is 'nuh-uh, actually it's amazing.' And it's frustrating because 'nuh-uh' is not going to pay the bills. And I think that even when I — and I spend most of my time trying to disprove myself. My premium newsletter this week, everyone should subscribe, is about junk bonds and junk debt. And I was really trying to see, okay, is there some way that these companies can finagle their way out of it. What I've actually found is a troubling amount of leveraged loans, so the worst kind of debt, that's piling up in the data center industry slowly but surely. It seems that every time I go looking for a counterargument, I find some new argument to support mine. This isn't even me being arrogant. I don't think anyone spends as much time looking over the numbers as me. And it's funny, actually, a mate of mine recently brought this up. It's like, is there someone you read who is kind of counter to you that you read and you see as an interesting challenge? No. And I don't say that out of any arrogance or anything. It's literally I would actually genuinely love to read a pro-AI person who puts aside however they feel about me insulting their favorite companies and says, 'Okay, how do I economically argue this?' Because at some point within the pro-AI argument, you eventually have to say, 'Actually, it's good they lose tens of billions of dollars. Actually, that's great.' And then you have to make this leap of logic between here and there that says, 'Well, at some point the value is going to get dramatically better, people are going to pay dramatically more, or the cost of running this is going to get dramatically cheaper.' People will say, 'Oh, well the chips will get cheaper or the chips will get more efficient.' Here's the thing. You can make a chip. You can make your jalapeno chip as Broadcom has with OpenAI, and you can say this is X faster. You'll notice though in all of the press around that, at no point have they said how much cheaper it is, how much money they'll be saving. If they had done that, they'd be telling you. Even with Nvidia talking about Vera Rubin, the next generation of Nvidia chips, it's, oh, it's 10x, 20x, 30x more efficient per megawatt when measured using DeepSeek R1, an open-source model released in March 2025. So it's very difficult to get a compelling argument against my work because it's really hard to argue in favor of this at this point on a rational basis. If your argument is something big is going to happen that changes everything, you're just kind of waiting for Father Christmas at that point.

I think they're getting a little more anxious. So, this OpenAI story, this annualized run rate thing that we can go into.

It's funny. They're nervous, but nervous in a very funny way where they're like, 'Well, actually, it's gross net revenue. It's actually good that this is happening. Actually, OpenAI's revenue is 70 billion annualized. If you look, if you squint real hard, it's actually great.' The nerves are there. The nerves are there as well with this Oracle force majeure story, because with OpenAI and Anthropic, it's always this wibbly wobbly stuff about revenues. Everyone just assumes they have the mandate of heaven. But everyone knows that if the data centers don't get built, this all falls apart. And now Oracle is saying they're going to truck the gas, the natural gas, liquefied natural gas, I believe, to the data center instead of needing that pipeline in New Mexico, which is extremely dangerous and also extremely inefficient. People also don't like that. I would say the data center stuff is where people are getting the most nervous.

So, when you think of how lending is done, the 10-year Treasury, so the American 10-year bond that the government sells every couple of weeks or a month — someone's going to kill me in the comments for not knowing that one — it is kind of the benchmark rate for everything because it's not super long, and it's also, even though mortgages are 30 years, most people refinance within 7 to 10 years. So 10-year is good for that, and it's basically the yardstick of most debt. So if the 10-year, which is currently about 5.269, just looking at it right now, if it's set, that means that is pretty much the lowest rate you're going to borrow on. So anyone borrowing is always benchmark plus a certain amount of basis points, with 100 basis points equals 1%. So the worse a company you are, the more you're going to pay above that benchmark. But if the floor of lending is going higher and higher, and it's doing that because of the war in Iran, war in Ukraine, kind of hyperscaler debt, but people aren't really sure about that. It's more towards the general sense that inflation is a problem in America. So the government is going to raise interest rates, but it means that the benchmark rate is going up and up and up when everyone has to borrow. And this is at a time when AI data center debt, people are starting to get a little bit worried about it. Starting to worry about whether all of this is actually going to work out. And so you're in a situation where you've got a high interest rate climate and it's going to get higher. The Fed's going to hike at least one more time this year. You've got an industry where people are already concerned. And if you're a CoreWeave or an Oracle, your debt is already trading. And so when a bond sells off, the effective yield goes up because let's say you're paying $100, just an example, for a 5% bond, and then someone sells it and now it's 98 cents on the dollar. That means the effective yield is higher. It's like 5.0-something. So Oracle — so you are, as a borrower, let's say you're Oracle, and your debt's trading, I think it's like 7.5% on average. That is the floor of where you're borrowing today. So, you would probably be borrowing above that because no one's going to be like, 'Oh, yeah, I'll pay you exactly what the market's giving you.' No, I want a little vig above that because if I wanted that rate, I'd just buy your debt off the market. So, this is a real problem because Oracle is now allegedly borrowing tens of billions more in some sort of vulgar SPV situation to buy more Nvidia chips. And the thing is, these AI data centers already take years to build. They already take a bunch of specialist talent and materials to build. They're already taking forever to build, but even when you build them, in theory, they are still 30 to 40% gross margin businesses. They're not that great. So, another percent or even a half percent of interest on top of that debt eats into the chance of recovery, especially because when you've got something like a data center or any project financing, which is just exactly what it sounds like, you are paying interest-only payments on the debt until it's ready to go. The lenders are fair in that you just make sure it's not growing. You're just paying the top off. But let's say you're paying more interest on that debt, and on top of it, you're paying it for longer because the data center is taking longer. We're talking hundreds of millions of dollars of extra costs on top for no value. It's not like that hundreds of millions of dollars is going into buying more chips, or it's going into paying people to finish it, or even paying off debt. It's literally just to keep your head above water. So everything's just getting slowly more expensive. In bubbles, everyone expects correctly that there's some big explosion. But generally what happens — and it's happened across the railway bubble, happened in the dot-com bubble, happened in crypto bubbles, it's happened everywhere — it's always a slow temperature increase over time. It's always this sense that things are just getting worse and worse and worse, and then things start popping off. And even in the worst kinds of debt, which AI debt somehow isn't in yet, the distressed market, so the CCC, the worst of the worst debt, that's showing the most stress it's seen since 2023, the last bank crisis, Silicon Valley Bank. That's usually — I just had Rob Wigglesworth from the FT, who wrote a book on bonds, to talk about this. When you start hearing the distressed debt screaming, that's when stuff is getting nasty. So, all of this is to say is people look for explosions, and they should. That's fair. What you should also look for is the frog in the pot, the slow boil, because that is what is going to bring this to its knees.

So there is some debate over this with the various bond writers around the world. You've got some people that will say no, it's not. It's just entirely that the US debt is too high and inflation is high, so everyone expects interest rates to go up, so we're asking for more from the government for their debt. But the argument also is that tech is issuing so much debt, and when I say tech I mean big, stable, big strong boys like Microsoft, less so, but Amazon, Google and Meta. These are companies that have a very, very, very low chance of default. So people are investing in them like they would a sovereign debt from a country, like US Treasury bonds. The question of whether that's actually crowding out is kind of up for debate, but it is definitely adding competition. The other thing is that there is this Treasury increase that's there, but it's also the reputational increase. Big tech's never taken on this much debt. There's never been anything like this in history. People try — even I just made the mistake of even making the railway comparison — but there's never been anything like this because this is not just big companies are raising lots of money. It's all sorts of smaller companies are raising lots of money. It's these project financing SPVs are popping up with data center holding corporation LLC and borrowing at these lower rates because they have a hyperscaler who signed a document, which I know you're going to get to. It's this very complex, multifaceted problem that means that tech bonds, even with these big stable companies, are going to start seeing nastier interest rates because the market is starting to say, 'All right.' I mean, of course Google will never die, but you're no longer — think of it like this: a year ago you were lending money to Google as an investor with their bonds based on the fact that Google prints money and runs a surplus of cash. Same with Amazon. Same with Meta. These are companies that just print money. So, you're just lending them debt — no, lending them money knowing that they're printing their own. Now, if you lend to Meta, Google, and Amazon, you know that Amazon and Google are cash flow negative. They're going to be fine. They're not going to die, but still, this is no longer a case of you are funding an extension of their business, but they'll be totally fine with their payments. They can no longer pay their debts out of cash flow. They have to borrow more debt. Again, this does not mean there is a debt crisis within them yet. It would take a lot for them to enter one, but it means that the credit profile has materially changed. They're not getting downgraded, at least not anytime soon, but it's now these are fundamentally different companies to lend to than they were a year ago even.

So to explain, yes, what's happening is, so when a bond sells off, its effective yield goes up. Important detail. So, Oracle's debt is selling off. People are a little nervous about Mr. Ellison's big party, and right now the market, I think the average of their debt is about 7.5%. The average on the high yield index, just an index of high yield bonds that is the benchmark, it's about 8%. So, they're close. That's because Oracle has a few low-interest bonds from a few years ago that are kind of pushing the number up. Yes, effectively the market is saying if you want to buy Oracle debt or if you want us to buy it off of you, we're going to be paying you like it's junk. This means that if Oracle raises any further debt, it will be priced like junk. And Oracle is in real danger. Depending — Morgan Stanley put out a warning yesterday saying that the delays to its data center could pose a serious credit risk. If Oracle gets downgraded, all of the debt held by investment banks and pension funds, all of Oracle's debt will get sold off because they can't legally hold junk debt. That's only if they get downgraded, though. But nevertheless, even before then, Oracle is effectively paying junk rates. And this is a very serious thing. Like I mentioned earlier, an extra half percent or even a percent is quite horrible. It means that you are paying — in the case of that debt I mentioned, the September 2025 bonds, that was $18 billion of debt between what I think it was 5 to 6, 12%, I forget exactly what it was. If you repriced that today, if they did the exact same issuance, they'd be paying 6 to 7 to 8 billion more in interest across the life alone. Again, this is not useful money. This is not money that's going into GPUs, and seriously, 6 to 7 billion, that's about a data center's worth, and a small data center. So Oracle is not done with the market, and the market is nervous about this. The market does not know what to do with Mr. Ellison.

First of all, I think it's the first time I've seen anyone declare a force majeure with relation to a data center. An act of God. I was at an investor conference when it happened and the noise I made was hilarious. So, woo. And then I had to explain the noise to people. But a force majeure notice is generally something, and I've mentioned this on the show before, that you do if you're like a pipeline out in the Middle East. A bomb goes off, hits the pipeline, a missile hits it, or there's a massive weather event that means, oh, we can't deliver the natural gas. And these are expected — if not great, they're expected within climates of war. When it comes to data centers, it's not generally something you do. And the argument that Oracle is making is that because of delays of getting the pipeline, which is now basically on the rocks in New Mexico to power the data center, they want to delay making their lease payments even though they have something called a hell or high water lease. That means they got to pay no matter what. This is a dodgy legal technique. And Oracle's Twitter account — and by the way, if you ever see Oracle tweet about something and they're like, 'Actually, it's not a problem,' it's the biggest problem in the world. It is the absolute worst thing. And naturally Oracle immediately was like, 'Actually, a force majeure, no, this is completely normal.' And then Bloom Energy popped up, who's the — they're the battery company they're working with, or the battery or some sort, I forget exactly how the energy storage there is. Look up Sam Coppelman. He's brilliant. He's written a lot about this. Nevertheless, Bloom Energy was like, 'Actually, Oracle is definitely going to pay us.' This is the stuff that happens when a company is deeply uncomfortable and they realize that there's a problem. So this force majeure notice is extremely weird and bad. But like all of these things, it's not bad enough to make the credit agencies freak out yet. The market kind of freaked out. The real thing is, and I think it's probably going to be first half of this year at this point — next year even at this point — Oracle is just going to cancel one of these things or say we are going to relocate it or so, which would be insane. But they're going to have to come up with some weird notice. And I bet they try and blame one of their counterparties and say, oh, one of the contractors, like every dodgy building project, just, oh, one of the contractors muffed it up. So it's just strange. It's without precedent. This whole — we are in unprecedented times and we have been for what feels like 20 years, but in particular with AI, tech has never done infrastructure projects like this. This is not like the dot-com bubble. It's nothing like it. Also that was telecommunications companies rather than hyperscalers. The actual software companies and hardware companies really weren't the ones, other than a few here and there. So, and also the debt wasn't as bad. It's just so strange. It's just a very strange era.