AI Subprime Crisis: ‘Wall Street is run by babies’ | Ed Zitron

50% of data centers aren't getting built.

That should create a market panic. It

doesn't because I think it's just a confidence game at this point. They will

start having to unwind trades. They will

have to also look past the AI bubble and say, "What's next?" And there isn't anything. And the longer that people

anything. And the longer that people cling to mythology, and the longer that people use these flimsy, halfass ways of defending the AI AI bubble, the worse

things get when it explodes.

On the tech report with me today is the writer of Where's Your Edat and the host of the better offline podcast, Editron?

Thanks for coming on again.

Thanks for having me. People have

compared the 2008 subprime crash to the AI bulb before and we've spoken about it as well, but focused on some of the differences in particular that the unbalanced overinvestment that we're

seeing now as compared to the sort of irresponsibly sold and overvalued mortgages that were that were going around in 2008. But that said, in your newsletter this week, you focus on the

similarities and you spell out what you call the subprime AI crisis. If you

could explain that for us. If you

remember the big short, there's the stripper with the five mortgages. This

is kind of what the AI bubble is like.

So, if you think of it like this, the housing bubble as people know it was only was not real. It was houses were bought and sold and all that, but the actual valuation of housing at that time

was massively inflated by the easy access to money. It was the fact that anyone could get a mortgage. You had

Washington Mutual that was eventually acquired I think by JP Morgan Chase uh that they were just giving out mortgage.

There was a famous story of a mariachi singer who claimed to have six figures and he managed to get his mortgage by showing a picture of himself as a mariachi singer in front of a house. So

in the same way that housing was massively overinflated and indeed the housing boom was only possible with lax mortgage investment. Well, it's the same

mortgage investment. Well, it's the same thing with AI. AI is not a real industry. If it was a regular industry

industry. If it was a regular industry that was held to the same standards as I don't know a regular business like hey you need to have a path to profitability. Hey you can't just lose

profitability. Hey you can't just lose five times what you make they it wouldn't happen. These companies

wouldn't happen. These companies wouldn't have the valuations they have.

They wouldn't have had the scale they have. None of this would be possible.

have. None of this would be possible.

And the subprime AI crisis is what I so if you remember back in the 2008 crash and the leadup to it people had these adjustable rate mortgages. So they

started out these low teaser rates that were very affordable and people sign them up signed up to these mortgages even without really understanding what they meant. In the same way AI startups

they meant. In the same way AI startups are building on top of AI models that are going to get more expensive that are only affordable when they have unlimited VC subsidy. And you have people

VC subsidy. And you have people subscribing to AI services like Claude who don't realize that they're building their workflows on top of something that has to get worse, that has to get more

expensive because right now they're heavily subsidized. This entire thing,

heavily subsidized. This entire thing, the subprime AI crisis I call it is what happens when the costs start becoming real and when the companies in question

anthropic and open AI start increasing their costs as a way or rate limiting in some cases as a way of starting to become more efficient though I don't think this actually makes them more

profitable. You you mentioned the rate

profitable. You you mentioned the rate limiting there and Anthropic has given us a perfect example very recently of raising their their usage limits I think as part of a promotion and then bringing

them down leading to lots of complaints about it. How are AI companies or if we

about it. How are AI companies or if we take anthropic kind of specifically how are they ever going to climb down from giving people £5,000 essentially $5,000

essentially in exchange for $200? So

that's the thing. That's the subprime AI crisis. Misspoke in the last answer. It

crisis. Misspoke in the last answer. It

might make them more profitable, but I don't think it's going to make them profitable. Now, these rate limits are

profitable. Now, these rate limits are particularly grizzly because what Anthropic did was in the middle of March, they started a 2x off- peak plan.

So off- peak rate limit increase. And

then a day before it was due to end, they then announced peak rate limit lowering. So during peak hours, which is

lowering. So during peak hours, which is like 5:00 a.m. to 11:00 p.m. Pacific, so

just like all all waking hours, I guess, uh rate limits would be much much smaller. And if you go on Twitter or X

smaller. And if you go on Twitter or X the everything app or whatever they're calling it these days and search Claude limit, you will see people screaming like they're being stung to death by

bees because the argument has always been in the AI bubble. Well, this is what companies do. they just charge a little bit and then then when they get the the attention and they get the

users, they just raise the prices.

Except when you look at how people are reacting to rate limits, rate limits so that they can no longer burn thousands of dollars worth of tokens on a $200 a month subscription or $130 on a $20 a

month subscription. They are freaking

month subscription. They are freaking out. These aren't people that are going

out. These aren't people that are going to respond well to the idea that they're going to pay more or be able to use it less. And that's kind of the thing. And

less. And that's kind of the thing. And

that is the subprime AI crisis. If you

It's really horrible. By the way, the people who lost their houses are victims of what I don't think is legally a crime, but should be. I think the way that these mortgages were offered were disgusting. You have people talking

disgusting. You have people talking about, "Oh, I had a negative amatization mortgage and I'm going to lose my house." That's because people built

house." That's because people built their lives on top of this unrealistic dream that they could actually afford these things. And they believed instead

these things. And they believed instead of saying, "Wait, crap. I won't be able to afford this in 3 months or 6 months."

when the rate changes, they said, "Well, the housing market's only ever going up and I'll be able to refinance." In the AI bubble, this is there's a version of this where people say they're profitable

on inference. They're not. There's no

on inference. They're not. There's no

proof of that. There's none. Dario Amade

said it was a stylized fact and uh Sam Alman said it in a dinner once, but there's no proof otherwise. They say,

"Oh, well, they're getting more efficient. Oh, it's less dollars per

efficient. Oh, it's less dollars per flop or what have you." No one actually has any tangible proof, but people have this mythology that they use so they don't have to think about reality. And

that's the thing with these rate limits with Claude. I think it's a scarless way

with Claude. I think it's a scarless way to get people used to the new. It's so

that people get used to lower rate limits. But the thing is, what happens

limits. But the thing is, what happens to the people that built their workflows on top of it? I'm not saying it's useful. I'm not saying it's good. I'm

useful. I'm not saying it's good. I'm

just saying that there are people that use it to write code who can now write less code anthropics idea maybe that people pay the API rate pay for the

models directly but I don't think people are going to do that. I haven't found one person and goodness have I looked who has said well we were getting a good deal I'm going to just pay the rate.

Nobody says that that's not how people work. People don't work in people are

work. People don't work in people are not going to go, "Well, I was paying $200 a month for thousands of dollars of tokens, but I'll happily take $400." But

then there's an empathetic way of looking at it. If you're used to being able to burn that many tokens, what the hell are you meant to do? And that is the rugpull that Anthropic is doing.

Open AAI, by the way, immediately got to hand it to Clammy Sam Orman.

Immediately, they were like, "Yeah, we've just reset rate limits." just just like it really is just like jumping off the Titanic getting on the Anola Gay at this point.

Definitely didn't learn his lesson from uh the Pentagon deal. I think

Oh god. Remember that.

I do. Yeah.

That was like that was like a few weeks ago. Feels like a year ago.

ago. Feels like a year ago.

It feels much longer. What What maths are investors doing to rationalize that disparity between the the costs and the the price it's being sold at? Is it that

they've drunk the Kool-Aid of the Uber and the Amazon web service comparison or is it just that kind of well I've already put some money in I should probably spend more to try and get it back out again.

So I think there are a lot of them that convince themselves that something will change. I think there I think there's a

change. I think there I think there's a lot of that. I think there's probably a few cynical ones who just know. I do

think it's the Amazon Web Services one which is untrue by the way. $51 billion

normalized for inflation in 10 years to make the single most useful compute thing we've done in ever like the backbone of the internet and open AAI

has now raised 120 billion anyway putting all that aside I think that there is a lot of group think and there is just this assumption that costs will come down I don't think many of them have done the maths my favorite one

though is that you have a few venture capitalists just claiming incorrectly that anthropic is has positive gross margins on serving their models. There's

no proof of this and indeed they have 45% gross margins but Eric newcomer of newcomer blog that's that's the name um revealed in that blog that actually it's

45% gross margins on compute for for anthropic but that's non-GAAP so not generally accepted accounting principles so it's just their moon math it's just

like oh yeah you know I think what anthropic does and I base this based on the thing that newcomer posted and indeed uh Dario Day's own statements. He

I think what he's doing is he's saying that the model costs this much to build.

So Claude Sonet 35 cost X number to build and we've made an this much money because of this and thus the model is profitable. Thus we have good good gross

profitable. Thus we have good good gross margins. Now I don't know about you, I'm

margins. Now I don't know about you, I'm I'm crazy. Gross margins to me are like

I'm crazy. Gross margins to me are like like uh revenue minus COGS cost of goods sold and then uh the margins that. But

this is the AI bubble. We don't we don't do that kind of messy margin math where we could just make stuff up. And I really think that's it.

stuff up. And I really think that's it.

Because at this point, if you're a venture capitalist that he that's heavily invested in this, if you have to think like me, you're going to start your car in the garage at this point. If

you if you have to accept what I'm saying, you have to just be like, "Oh, right. This probably won't work." I

right. This probably won't work." I

think a few of them are clinging to the idea that the next generation of Nvidia GPUs is 10x more efficient. By the way, that's what they said about Blackwell.

But that's the thing. What happens to all the Blackwells? They've sold what over a hundred billion dollars worth of these things now. What happens to the Blackwells? If they're not profitable,

Blackwells? If they're not profitable, what are we doing here? And that's the thing. No one really has a good answer.

thing. No one really has a good answer.

To this day, no one has come come to me with a good answer other than, "Well, Daario said it. Daario said it's profitable. Why Why don't you Why don't

profitable. Why Why don't you Why don't you trust Daario?"

Looking at some of the the news we've seen lately in your in your newsletter, you said one of the pale white horses of the AI bubble is if data centers deals

start getting delayed or cancelled. And

in Bloomberg, they reported yesterday that almost half of the US data centers planned for this year are expected to be delayed or cancelled. And in fact, in the FT today, there was a story about a

2 gawatt data center deal between the AI startup poolside and and Coreweave is collapsed and and over what they call I think it was strategic and timing

reasons, which I always enjoy how vague they are with the reasons that things fall through.

Strategic they weren't going to do it time. there isn't enough time to do it.

time. there isn't enough time to do it.

Like yeah, I'm just I'm just wondering why why doesn't Wall Street seem to care about this that that data centers are getting cancelled and delayed.

Well, that Bloomberg piece, by the way, is based off of the same research from my piece last week. So, I guess if people want to be a week behind, they can read Bloomberg. Oo, Um, but

nevertheless, um, the cool the call we've won as well reported as an exclusive in the Wall Street Journal in October 2025. The reason that Wall

October 2025. The reason that Wall Street isn't reacting is that Wall Street is run by babies. I assume people with the who don't actually care about reality and have their own imaginary

things because if they did, they'd start freaking out. That poolside deal was for

freaking out. That poolside deal was for a 2 gawatt data center for a relatively unknown startup for a startup poolside that and that was like, "Yeah, we're going to make models, I guess, and we're

also going to do AI infrastructure. Wall

Street Journal write that crap up." Just

yep, that sounds good enough to me. dead

less than six months later. And what was great was it strategic and timing reasons. They then tried to set sell it

reasons. They then tried to set sell it to Google. Google went no. And then

to Google. Google went no. And then

their funding round fell apart. The

thing is, it gets back to what I was previously saying. If you have to start

previously saying. If you have to start truly digesting what I'm saying, the world looks a bit scary and apocalyptic if you're heavily levered in this or indeed even if you were just

particularly bullish because this isn't something like the.com bubble for example where you could look back and say well people were using the internet and the internet did have this and like

people were already transacting. It's

not the same because when they used a website, it didn't sometimes send them to a different website or it didn't cost randomly $50 to access the website because your internet connection varies

in price. The capex for that fiber lasts

in price. The capex for that fiber lasts 15 20 years. GPUs 3 to six.

But if you stop thinking in a way where you rationalize everything using the past, everything seems crazy. 50% of

data centers aren't getting built.

That should create a market panic. It

doesn't because I think it's just a confidence game at this point. I think

they all realize that if they start truly accepting my gospel, my things I'm saying, they will they will start having to unwind trades, they will have to also

look past the AI bubble and say, "What's next?" And there isn't anything other

next?" And there isn't anything other than the private credit bubble. That's

going to be nasty. on that. What do you make of OpenAI hitting uh its 825 billion valuation with the combined news

of it also apparently raising $122 billion in funding, which for clarity I assume is just the same one that we were talking about the other month with uh

with the conditional funding reliant on achieving AGI or surviving an IPO.

It's the same thing. And that 82550, I forget which it is. That's the post money valuation. So something plus 120.

money valuation. So something plus 120.

Not going to do the maths. Um I think that that round is so weird because other than the fact it's barely half actually funded, Nvidia already said it's the last time

they're investing in OpenAI. And it's

this strange thing where every people don't seem that excited around it.

They're all kind of limply going, "Yay, it's the largest funding round of all time." Like, no one seems It used to be

time." Like, no one seems It used to be like a When OpenAI announced their fund, last funding round in 2024. It's like a six billion, I think it would. Forgive

me for not remembering. People were

doing back flips. People like, "Wow, it's a big day for Silicon Valley." Here

they raised the biggest round of all time and people like, "Oo, good." And I think it's because people realize this stuff is overvalued. There is currently a problem where secondary sellers, so

open AAI private companies, startups generally private companies. You can't

sell on the public market, but there are these secondary markets where you can sell your shares. So there's a story that's come out saying that there's about $600 million worth of OpenAI shares people are trying to sell.

Nobody's buying. Everyone wants to buy into Anthropic because Anthropic is cheaper. Anthropic is at 380. Open AAI

cheaper. Anthropic is at 380. Open AAI

is at 8 something. So people are like, "Ah, I don't want to buy in at that high valuation." But here's the thing, people

valuation." But here's the thing, people I feel like if people thought OpenAI was going to be a 2, three, 4 trillion company, now would be a great time to buy in. So people don't want to, which

buy in. So people don't want to, which makes me think that people are also sort of realizing that everything's everything's a bit weird because even with all that money, and this is really

insane, even with all that money, Gun said, I don't think Sam Orton could tell you how they go profitable. I don't

actually think he could tell you how this business lasts 10 years other than we will make more money. Yes. And I

think that that is something people don't want to say but are thinking about very very hard and I don't think anyone has a good a good idea or an answer.

Same with anthropic. The coach decks the one that Eric Newcomer published about anthropic from end of December last year. That one was interesting because

year. That one was interesting because they believe it will be a multi-t trillion dollar company in a few years, but also that the cost of training is not increasing at the speed of revenue.

Still not actually not actually true based on anything I've read, but also that's not the same as becoming more profitable. That just means one thing is

profitable. That just means one thing is not quite going as fast as another. They

everything is based on just these almost mimemetic ideas, these little concepts they come up with to sleep better at night. Bedtime stories to make the AI

night. Bedtime stories to make the AI bubble talk not so scary. And I think that I really want these companies to go public. I want to look at I want to get

public. I want to look at I want to get in the guts there. I want to sit down with a big cigar and read the S1's because once they actually file an S1, those things are audited. And I can't

wait to see the audited financials of either of these companies because they're gonna look like a dog's They're these aren't well-run companies.

And I mean, they're every AI startup is a terrible business. But the worst of them is anthropic and an open AI. One

one other story I saw in the news which I kind of want to get your reaction to was the the it kind of goes back to that data center stuff we were talking about earlier as well is that Oracle laying

off tens of thousands of staff with a 6 a.m. email to presumably fund building

a.m. email to presumably fund building data centers which presumably now half of which won't be being built this year.

And I just wanted to get your reaction to that to be honest.

Well, I mean that's not totally fair.

Sam Alton posted the other day that over in I think their Michigan data center, they have several steel beams. I'm not kidding. That was that was quite

kidding. That was that was quite literally what he said. The thing is, it's not just the ones getting cancelled and the ones getting delayed.

These are not delays. These are

realistic timelines becoming reality.

These things are always going to take years and years and years. The concept

of a hypers scale data center is very new. The we don't have a gigawatt data

new. The we don't have a gigawatt data center yet. And to be clear, gigawatt

center yet. And to be clear, gigawatt data center is actually a campus of several data centers, daisy chain, but we don't have one yet. And Stargate

Abene, the one that I talk about, the the big Oracle OpenAI data center, my favorite subject, they are meant to open that by the end of the year. They do not

have the third building ready yet. There

are eight of them.

It's times like this I feel a little crazy because you talk to people about this tech journalist investors and this and they go, "Yeah, but they'll work it out. Yeah, yeah, they'll work it out.

out. Yeah, yeah, they'll work it out.

That is the best they've got." And it makes you feel a little nuts because you're like, "Yeah, but if you look at what's happening, nothing actually matches up." And they go, "Yeah, but it's fine. It's fine."

And the funny thing is when we compare to I'm doing a big premium piece this week about how AI isn't too big to fail.

partly just it isn't even close. I can

get into that if you want. But the thing is with that is when you look back at what it was that people used to justify subprime mortgage crisis. There was an insane piece in the New York Times in 2007 that said uh we shouldn't stop

subprime lending because it's going to stop poor people getting mortgages. And

just to be clear, the the prime the subprime mortgage crisis in the near prime mortgage crisis was not caused by poor people.

Everyone got in. It was actually lower income families had less access to housing during the time because the price of housing kept going up. It was

middle- inome, higher income people who did it too, who also couldn't afford their mortgages. Putting all that aside,

their mortgages. Putting all that aside, the rationale at the time was the housing market will keep going up. And

there's a Brookings Institute person who said, "Well, it's not a problem.

Delinquency rates aren't scary." And

actually, as long as um unemployment doesn't go up, we're fine. You'll never

guess what happened. You'll never guess what happened. And that's the thing. I

what happened. And that's the thing. I

think we're in the hype cycle now. But I

also don't think this is comparable to a previous era. At least not at the scale

previous era. At least not at the scale that people think it is. Like you can't like for like say this.com bubble because venture capital was much smaller back then. The markets were much smaller

back then. The markets were much smaller back then. It isn't like too big to fail

back then. It isn't like too big to fail because there's nothing within this that's too big to fail. too big to fail referred of course to the AIG and the destruction of commercial paper which

was how businesses and big well big banks and businesses like GE Capital funded their operations.

The lending in general would have broken without the bailouts. TARP existed also an important thing. TARP the uh troubled asset relief program that was used in

America that was not used to buy houses.

It was used to buy financial instruments that people had gambled with. It's why

everyone's so angry at the banks for getting bailed out because it wasn't like people's bad houses were bought. In

fact, the 40some billion dollars that was meant to do that just evaporated.

Didn't really help anyone. Um, that's

the thing. I think people are, and I said this before, clinging to the ghosts of the past. They want to dance with the ghosts because if you look back and you say this is like this, it's less scary.

Even if the thing in the future is bad, you can say, "Well, we were okay after the.com bubble. We were okay after the

the.com bubble. We were okay after the great financial crisis." No, we weren't.

The current state of things, Trump, our current economy, the AI bubble, all of this kind of comes out of the fact that derivatives were not. They were

regulated to an extent, but they weren't regulated hard enough. And after the great financial crisis in 2020, the Vulkar rule, which was a part of the DoddFrank Act, which let which stopped

speculating with customer money, the Vulkar rule was changed under the first Trump administration to allow banks to invest in venture capital and private credit. All of this comes back to a lack

credit. All of this comes back to a lack of regulation. But none of this is

of regulation. But none of this is something you can solve by looking backwards and justifying it, saying things worked out in the past in this way. This is nothing like that. And the

way. This is nothing like that. And the

longer that people cling to mythology and the longer that people use these flimsy, half-fast ways of defending the AI AI bubble, the worse things get when it explodes. I I don't care. I don't

it explodes. I I don't care. I don't

care if people agree with me or disagree with me. I love doing my work, but it's

with me. I love doing my work, but it's frustrating because consumers are still being conned into investing in these stocks. People are still being conned

stocks. People are still being conned into believing that AI will be here. the

subprime AI crisis I talk about, the customers of these AI companies are using them believing these products will always work this way. But as Anthropic

has shown with these peak peak hours for rate limits and by the way this is just this isn't the first time they did something like this. Last year they added priority processing and service

tiers both open AAI and Anthropic.

They added these for enterprise customers. These things are only going

customers. These things are only going to keep happening and it's going to become blatantly obvious that the AI you use today just will not function in the same way even in a few months which will make your way of life under AI

untenable.

Well Isron, thanks for coming on.

My absolute pleasure.

If you enjoyed today's episode and you want to keep hearing more from the tech report, please consider liking and subscribing. Also, you can get episodes

subscribing. Also, you can get episodes of the tech report wherever you get your podcasts.