AI Bubble: ‘Something profound has changed’ | Ed Zitron

I have the books of a major data center company and they're losing hundreds of millions of dollars. Hundreds of millions. This is most of the AI industry. The usual ability that the AI industry has to put the toothpaste back in the tube — it's not working anymore. And people are starting to realize that the maths does not work. The more truth in numbers that comes out, the quicker this ends.

Thanks for having me.

Yeah.

So eager beavers who've been watching and listening to me for a while will notice that I've spent the last few months getting into the debt markets, and that's because the stock markets are manipulated. They have been for a long time. They are emotional beasts. They go up, they go down, they go side to side, but the debt markets less so. They tend to — sure, they can be emotional. Sure, there are dumb investors, but they tend to boil. And when they go over the edge, it's kind of hard to get the thing back in the cup. That's a terrible metaphor, but you get my point. So this week, CoreWeave's debt — so when bonds sell off, the expected yield goes up for their debt. Long story short, their debt touched 16%. One of their fixed-term loans touched 16%. So we are — and I just checked just before coming on, it was all of them are above 14. One of them's above 15.3. Point is, these are personal savings rates. They're personal loan rates at these points, which makes just borrowing for CoreWeave and by extension most Neoclouds and most data center SPVs just a ton more expensive. So the stock market is puking for reasons we'll get into, but it feels like suddenly everyone started doing the maths I did a year and a half ago.

Year and a half ago. Gartner, bloody Gartner, who they've been fans of the metaverse, whatever. So, October 2025, Gartner said, 'Neoclouds are reshaping AI infrastructure and that tech service leaders who fail to integrate them into their portfolio risk higher cost, slower innovation, and diminished competitive advantages in the AI services race.' So, Gartner said yesterday that by 2030, 80% of Neoclouds will be dead. So what's happened is suddenly everyone went, 'Can CoreWeave pay its debts? Could they raise tens of billions? Can they pay that if something goes wrong?' And suddenly everyone's saying, 'No. Nope. They're not sure they can.' They're not fully accepting that CoreWeave — they've not pledged themselves to the graveyard smash quite yet, but they have definitely seen something ugly. And I think it's just this: there is something profound that's changed in the last few weeks where you've got all sorts of economists going — Goldman Sachs came out with a report, I think it was last week or the week before, it was like how much money would hyperscalers need to make just to cover the capex that they're spending in 2026 to 2027. The answer is, to just break even, I think they need like $360 billion of brand spanking new AI revenue at a time when they've got maybe a buck 40, buck 50, buck 70 if they're lucky. Now it's $425 billion if they want to have a 10% return on invested capital. So suddenly everyone's doing these things and you've got a bunch of people saying, 'Oh, we need trillions of dollars of AI revenue to make capex worthwhile.' If only someone had said something. I like — if only someone had said. Now,

it could be anyone, but Gartner also added one other very — and this is Gartner. These are the hypiest hypesters. They are an analyst group, but they are always analyzing to the benefit of the tech industry. But they said through 2030, AI data center owners would need to generate a cumulative $8.5 trillion in revenue to cover the announced capital spending, the financing cost, and a 12% return on invested capital. Now these are companies with like 20%, 30% ROIC on their investments in general. So this is what we would call bad in the financial industry. So the stock market right now — we'll get into the Anthropic stuff in a bit. Stock market's going down. Everyone's worried. The Korean stock exchange, the KOSPI, has dumped its puke two days straight. Could go back up because Jensen Huang goes out and said, 'Actually, the numbers will be even bigger next year.' And people go, 'Yay.' But something has shifted where the usual ability that the AI industry has to put the toothpaste back in the tube — it's not working anymore. It's not working. Confidence is not being restored. And then there's this Anthropic stuff that I know you want to get into, but there is something that has happened and I don't know where it is, but there was an Information story saying that — I don't know if you remember I've brought up MUFG and SMBC, the two Japanese banks. Do you remember I brought them up in the past?

There's an Information story about a week and a half ago that said they've become more selective. More selective. More selective now on how many data centers they're going to invest in. Now, these are the loosest lenders other than Blue Owl. They're actually funding Blue Owl as well in the data center financing. So, it's the little things I've been warning about the last few years. They're all starting to come together. And the problem with the AI boom, bubble, whatever you call it, is that this is not something that's being substantiated by revenues or returns. It's hype. It's interpreting numbers and saying, 'I've read the tea leaves and the run rates and the big numbers in the analyses, and I've decided that AI will be the biggest, most hugest and wettest sun ever.' Problem is, is that the sun is dry. This boy's never touched the water. And people are starting to realize that the maths does not work. And the problem is, is that Anthropic was the last key to be jangling. It was the last thing they could show that would improve things. And well, the numbers came out and they weren't what people needed. The sell-off hit CoreWeave the hardest as far as I can tell so far anyway. And so obviously I had to take a little look at the share price going back a year and year-to-date, which is one that really surprised me, is that they are down 18% from the date of recording a year ago, and they are only about 20% above their debut IPO price as well, which is also not good for a year and a few to several months later.

Yeah.

I think we're — I think we're there, but a tipping point is more complex than just 'thing bad.' So, let me explain. I've kind of mentioned this with Will, but I want to go over it again. So the way you should look at debt in general is whatever the last time the debt of the company in question priced at — which means when CoreWeave has a loan or a bond, pardon me, that sells off on the market, when it sells off, it's cheaper, thus the effective yield, because you're paying less on the dollar for it, is higher. Whatever Core's debt currently trades at is roughly where the debt will land, if not higher, if they borrow from someone. This means that the average cost of borrowing for CoreWeave now — it's not going to be below 10, might be 11, 12, 13, ridiculous rates. This is also — CoreWeave is arguably one of the best capitalized, best connected, best funded Neoclouds out there, which means anyone raising data center debt right now is both in the toilet and being flushed at the same time. It's just — it's going to be nasty. Vast, which is this brand new Neocloud, barely a few months old, apparently raising a leveraged loan right now, real toilet loan there, at over 11%. And the banks in question, I think it's JP Morgan and Goldman, I think it is, pardon me if I'm wrong on that one, they are not sticking around. They're trying to sell off more of the debt. They're not keeping it. Now, generally, when the bank thinks it's going to get paid, the bank keeps the debt. They keep it because they want all that money. So, why wouldn't they want to keep it? It's almost as — And yes, there are complexities to the debt markets. I know people sell things, they buy things, whatever, but they're not holding quite as much. Why? It's probably because they don't think they're going to get paid long term. And to quote the CEO of the company from Margin Call, they're selling to willing buyers. But the thing is, it's not so — CoreWeave and the AI bubble writ large. One of the mistakes I've made in my analyses, the early ones I would say, is that I've been looking for an explosion when you're looking for, to quote my mate Casey, a dead body. It's a slow boil. You find corpses here and there. Things die, things don't happen. Contracts don't get signed. Debt doesn't get raised. So, it's less that CoreWeave suddenly explodes and more CoreWeave misses payments and CoreWeave cannot raise more debt. There was an analyst I found that said that CoreWeave needs to raise another 100 billion or so in debt by 2030. So across the next three years, at what — like you're talking 10, 11, 12, 13% rates. At some point, even if you think AI is the most important thing ever, there is a point at which this debt is unsustainable against the potential return. Because even if this company is cash flow positive when you remove all the costs, when it's just on a pure operating expenses basis, the problem is depreciation, which is an upfront cost amortized over time, is still going to eat into their profits forever at this point. Like, just most of the chips they bought haven't been put in service. Most of the data centers that Core Scientific is building for them haven't been put in service. So the depreciation hasn't even started for most of this stuff. So Core's just — it's more like a disease-ridden company than it is going to explode. Now, crazy things could happen. Nvidia CEO Jensen Huang just said, 'We don't need to backstop CoreWeave. We don't need to backstop them.' Which is funny because they've already backstopped CoreWeave. There was a story in Business Insider back in February saying CoreWeave says that Nvidia has its back on leases. Backstop. Like, come on, man. But the thing is, why would I possibly rely on the financial press to push back on this? Why would I possibly imagine the height of the financial press would say, 'Hi, Jensen. Hi. Did you know that you already did that? Why did you say that if you already did it?' But no, why would we need object permanence? This is only the literal underpinning of the entire bloody stock market. I don't know — I'm very black pill when it comes to financial journalism right now. It drives me up the goddamn wall because it's like even I — should I talk about Anthropic already? I'm eager to —

Okay. So, we should launch into the Anthropic. So the reason that the markets are puking today, which is Thursday, so this will be running on Friday, is because Anthropic's Q3 numbers came out in a leak from an investor meeting leading up to their IPO. Now, their Q3 numbers were — they had $16.8 billion in revenue. Sounds great, right? Well, they had 11.5 billion in Q2. This means their growth rate is 46% quarter for quarter, which is huge. Wow. Except it's not enough. It's not enough. They had 4.8 billion in Q1. It went from doubling to not even going up 50%. Dario, that doesn't sound so good. But then when you dug into it more, oh, we have an operating margin of 11%. Adjusted operating margin. You know, we'd be profitable if it wasn't for all these costs. And then they had a gross margin of 85 — no, sorry, adjusted gross margin of 85%. I wish I could do this with my — like, I want to buy a $50 billion house. Yeah, if you just remove all my costs and if you extrapolate into the future, I'm going to be a trillionaire. Give me a mansion. Give me — build me a palace in Las Vegas. I want one. But the other problem is, is that you remember run rates, right? You remember my complaints about run rates? So annualized run rates take a period of time of revenue, multiply by a number. No one knows what the revenue number or the period of time means, but you get a magic number out of it. Latest one from Anthropic was $65 billion in July. Wow, what a huge number. And then Anthropic neglected to make up a new one. They didn't include one. So AI boosters are all bringing the Manila — like Bug Dwire style, bringing the Manila envelope up to the podium today because they've all kind of realized that despite this being genuinely impressive revenue — that's a huge amount of money — it isn't enough. Anthropic has $413 billion worth of non-cancellable commitments per Reuters. They also revealed that 5 to 7% of their revenue came from distillation, as in Chinese AI model companies using Anthropic's outputs to train their own models. And Anthropic claims they've killed it.

I also — the stealing machine gets stolen from. But the funny thing is, people are like, 'Oh, that's fine. It's fine. It's not a big deal.' Would you say that if Microsoft said that 5 to 7% of their revenue went away? Would anyone care? No one would. They'd be like, 'Whoa, what are you talking about? 5 to 7% of your revenue. That's bloody huge.' But then there's the other problem that revenue is decelerating at a time when it has to accelerate. Anthropic has to be — they were meant to, per investors with rebar in their skulls, meant to have a $100 billion annualized run rate. Who gives them? They were meant to have this by the end of the year. It's obviously not going to happen, but you can see the AI boosters, they're all starting their cars in the garage a bit because they're all realizing that they had — the boosters, the AI companies had made these huge promises. They'd said we are going to grow eternally. Reuters reported a few months ago that Anthropic's IPO, their valuation hinges upon getting to 180 to $200 billion in revenue by 2028. It's just the numbers are not there. And the problem is that it's not just, oh, Anthropic can't afford its bills. It's that Google has taken on hundreds of billions — sorry, $800-something billion of off-balance-sheet agreements to provide compute to Anthropic. It's Broadcom has taken out a $35 billion private — it's a private credit-backed thing with Apollo and Blackstone where $30 billion of the debt is backstopped by Broadcom, as in if Anthropic doesn't pay, Broadcom will have to pay either way. You've got Hut 8, Fluid Stack, all of these connected counterparties that must have Anthropic grow because if Anthropic does not grow, they do not get paid the money. But then there's the other problem, which is even if the distillation wasn't taken out — which it's gone, they've said, Anthropic, the story ran three days ago that said, 'Oh, we've reduced the attacks by 90 to 99%,' just meaningless numbers. Again, even then, Anthropic will have to just to match its current growth rate, they will have to make $24.5 billion in the next quarter. And the quarter after that was — I can't even do the maths off the top of my head. I just did it beforehand, I'm going to admit. Is this to keep up with the growth trajectory it has set itself? Anthropic needs to be cooking at like 35 to 40 billion dollars worth of quarterly revenue in the next year without fail because otherwise you're just not going to make it. And this is the funny thing, I think much like CoreWeave, the numbers have always been there. The numbers — they've always been floating in the ether. Everyone's kind of — I've been saying this for like a year and a half. People have said, 'Ed, Ed, you're wrong because it hasn't happened yet.' And by the way, I was early. Of course I was. I know.

Here's the thing. The things that people are criticizing Anthropic and CoreWeave for are the same things I've been saying, which is to do all the things you are promising. To make all the things you have said will happen happen will require you to grow at such an aggressive rate that everything is priced for perfection. And this is not a perfect company because even with this rapid revenue growth per Ramp, 80% of Anthropic and OpenAI's enterprise revenues are from the 1% of their customers, mostly AI startups, who are selling services where you can burn $20 a month, $100 a month worth of tokens from Anthropic and OpenAI. And even with significant discounts, that money don't make sense. For 200 bucks a month on Cognition, you can burn $1,200 worth of tokens from OpenAI and Anthropic. That don't make no sense. But that is the revenue underpinning this growth story. Which means that Anthropic can't really sell its way out of this story. This isn't one where they can go to — it's not like Anthropic's customer base is mostly big businesses, as in a JP Morgan or a Goldman Sachs, and say, 'Okay, we need you to pay more. Oh, we're upping the cost.' No, it's mostly venture-backed startups who will gladly let their users burn as much as possible. Except even they are starting to feel the credit crunch. And so the way this all falls apart is a burn, a boil rather than an explosion. It's going to be a decay of these underlying numbers because at some point Anthropic is going to have to do a prepayment for all of this compute capacity that it's promised at some point. And we haven't even mentioned the largest, wettest sun in the valley, OpenAI. But OpenAI faces every single problem I'm talking about plus plus because Clammy Sam Altman has made even more promises. Ananya Varma at the Wall Street Journal reported a few months ago $750 billion worth of planned commitments, compute commitments through 2030. Where's the money coming from, Sammy? The answer is nowhere. He's not — both of these companies. So a few weeks ago, I did an analysis, it was a week ago even, where I said I think that both of these will get rated triple C. So pretty much as bad as it gets credit rating before they are literally just saying this company's in default. They'll never pay their debts. This isn't one where they can, like SpaceX, trick their way out of it and say, 'Oh, actually we're going to be — we're actually going to — we're a bit profitable. We just — if we pull back on the capex.' These companies don't have capex. OpenAI and Anthropic do not buy chips. Their sugar daddies do. The Microsofts, the Googles, the Amazons, the CoreWeaves, the Nvidias in some cases. But fundamentally, the mathematical mismatch is becoming too hard to really bear. And while these companies might not run out of money immediately, the problem is that the counterparties, these various data center companies, run on debt. Tons of debt. And the debt — even though the interest rate of the 10-year Treasury bond has dropped a bit, even then they're now being priced both on that, which is still elevated, still above 5.08, I believe, percent. They're also paying the data center premium. And the data center premium, just to be completely upfront with what that means, the data center premium is, 'Hey, are you going to exist in three years?' And that is no longer a foregone conclusion. Much like Anthropic growing to this meteoric size of a business that's bigger than Microsoft and Google is no longer a foregone conclusion. And it had to be. The whole AI trade hinges upon the idea that nothing bad will ever happen.

Well, the opaqueness. Yes.