Transcription
Hello and welcome to Why is this happening with me, your host Chris Hayes. And welcome back to our ongoing series, Why is this happening? The mind-body problem, about all the implications of the AI boom.
Um, you have probably seen, well, I don't know if you have, but maybe you've read the book, uh, The Big Short by Michael Lewis, or seen the phenomenal movie by Adam McKay, which is like, I think genuinely a classic, The Big Short. And the tale there is a chronicle of a desperate group of, what you might call, our kind of financial dissidents, who in the sort of era of 2006, 2007, as the housing boom is going and everyone's making a ton of money, start to sniff out there's something very deeply amiss, deeply wrong. And the more information they get, the more convinced they become that everyone in the market, more or less, is wrong. That this entire money-making machine is about to collapse in on itself. Because they're people in finance, they take bets that it will collapse, and those bets prove to be accurate. It's a very satisfying tale and a satisfying um book and movie because it's a kind of David and Goliath, and you sort of get to follow the profit as the prophecies come true.
And the reason I bring all this up is, we're in a boom right now with AI. I mean, the amount of money that's being put into it is staggering. And I think the broad amount of the financial system and the press thinks that that it's going to pay off. Um, at least, you know, that's basically been what markets have priced things at. If you look at the stock of the big AI companies, and particularly Nvidia, that makes the chips, this all depends on it, keeps going up and up and up. It's been a little down recently, but there are just like there were at the housing bubble, there are dissenters, and I talked to one today. Now, the thing about being a dissenter, a dissident against this sort of conventional wisdom is, you could be wrong. You know, there were people that thought, for instance, um, the internet was never going to amount to anything, or that personal computers were a ridiculous technology. And in the long view of history, it's always a fine line between crank and prophet. Sort of depends on what happens in the end. But given the fact that there is so much money at stake, and because I think the the maximalist view that this is going to end in tears is one that people get the least amount of exposure to, I wanted to spend some time reckoning with that view today with my guest, Ed Zitron. Ed is the CEO of EasyPR. He's the host of the Better Offline podcast, and he writes the Where's Your Edat newsletter on Ghost. And I would say that he is one of the foremost um AI bears, or even AI haters, on the whole internet. Is that fair?
I think that's fair. But I really hate bad software, like that really. I love technology. I re I owe my life to technology, but the current state of the tech industry sucks.
You're a tech person. You were a gaming journalist for a little bit.
Yes, I was. I wrote about video games in in England for goodness about five, six years. Moved to America in 2008. Great year. Fantastic year to move to America. And yes, um, it's strange watching what's happening. And it feels like as I watch the pieces fall together, it's all been kind of working up to this because the AI bubble is a symptom of a larger problem with the software industry. The the hypergrowth era is ending. We have seen software as this thing will always grow exponentially. So every single bubble is looked at through the same lens. You say metaverse will grow exponentially. NFTs will become will take over all culture. We won't buy physical things anymore as far as collectibles go. We'll only have these digital versions. Everything seen through what I call the rot economy, the growth at all cost mindset. And I think we're coming to the close of it.
Let's stay there because I think one of the reasons that I hear skepticism about AI from people, and I think is irrational, is that we did just go through this huge hype cycle. Yeah. That was, I mean, it's a little bit wiped from memory, but before AI, during COVID, there was a huge amount of money sloshing around the economy um because of all of the money was being directly injected both through fiscal stimulus and the Fed, and people were spending all their time in front of their screens. Um, and there was this huge boom around the metaverse, the metaverse, blockchain, and like non-fungible tokens, NFTs. Yes. What do you think the takeaways from that sort of boom-bust cycle are?
So a little bit before the metaverse, there was another bubble that other people forgot about, and that was Clubhouse. So Clubhouse was this audio-only social network that, if you talk to venture capitalists at the time, oh my god, this was the biggest thing ever. It was going to be worth a bazillion dollars. It was radio too. Now, nobody went and looked up how much radio makes or what the revenues were, really anything. But they got celebrities on. They did everything they could, and they, it was very clear the venture capitalists were pushing this so that they could get a big acquisition. Now, never happened, and Clubhouse has kind of fallen into irrelevance. With NFTs, with the metaverse, people forget that Meta used to be called Facebook. We still call it Facebook. They changed their entire company name to Meta. One of the craziest things in history, and we just don't talk about it. It's insane that happened. You had people on CBS News being like, "Yep, the metaverse is here. We're all going to live in the metaverse. It's gonna, it's very real. It's going to happen." And the actual experience was a very bad virtual reality experience. But I feel like the tech industry has kind of been LARPing for the last 10, 15 years. They had LARPing, meaning live action role playing, pretending, going through the motions, because you had the era of smartphones and mobile apps, huge deal. You had the era of software as a service, SaaS. These were the big revenue drivers of the tech industry. It was a way to get more money out of people because you had a subscription service, or you had an app you could buy on your phone. Most of those all have monthly subscriptions now. Yeah. Convert to subscriptions. Exactly. So it was a way of getting people away from that troublesome thing where they only paid you once. Nevertheless, this worked for a while. And so that worked, and then, uh, apps worked, and then, um, nothing else really worked. We kind of started running out of kind of hypergrowth ideas. We haven't really had one since, like, I don't know, LinkedIn. Maybe there are probably some examples, but we just stopped having big things that worked. So the tech industry did what it did before: hire a lot of people, put a lot of money into things, buy a lot of things, buy little companies. Uh, the Activision Blizzard acquisition from Microsoft was claimed as a metaverse thing, which was very silly, because if the metaverse is all video games, that's just so much. But it's the tech industry doing what it thinks works. And I think we're getting to the end point of that.
So here's what I, I want to talk a little bit. I want to wait to talk about the tech and then, and just, and hive that off from the economics of it. But just, just stay on this for a second because to me, the big difference is, no one could ever really explain to me the use case of the metaverse. You know, people would sometimes be like, yeah, I'd be like, well, what's, what, what do I do with it? And people would say, you know, with crypto, it's like, well, can I pay for a cup of coffee with it? No, you can't really do that. No, it's, it's essentially, um, there was a little bit of talk of like, the blockchain is going to replace contracts. And I was like, but is anyone actually doing that? No. Yeah. And, and even the metaverse, you know, you kept sort of saying like, what's, what's the use case? What's it do? No one can ever answer. I don't feel like that way with AI. Like there are use cases. In fact, I've used it for use cases. And it seems to me like there's a much clearer, like, okay, um, do this doc review. Here's a thousand documents. Um, it would take a human to go through them. It can read and synthesize. Now, the question of whether it could like do it well or not is a different question. But to me, the difference is that you can at least articulate what it's for in certain circumstances or reasons or places it might be useful, in a way that for me, the metaverse never did.
So there are uses for large language models. If you remove all of the hype, there are things they can do. The problem is, is here's a little challenge for you: Go and talk to a bunch of AI boosters and ban them from speaking in the future tense. Don't allow them to say anything about "this will," "it might," or "it could." Say what it does today. Because when you do that, it's not really clear what's changed. So yes, you can use it to review documents. The results are not great, or maybe they are. You actually don't know because you didn't read the documents. A thing that is based on statistics did. The idea that you can rely on it is inherently broken because OpenAI's own research says hallucinations are a part of these things. They're never going away. You have people in the AI industry claiming hallucinations are going away. They're just wrong. OpenAI said it. You going to argue with them? There, there's a recent paper I was just looking at yesterday that says that even when you, because the, when I use AI, I do, you know, use it where you sort of like gate the sources. And there's a paper I saw yesterday that like, even when you gate the sources, right? When you're saying, "Just use these sources," you cannot purge a hallucination. No. And even with coding LLMs, because this is one of the most annoying debates ever, is the usefulness of LLMs within coding. And I think the big thing with the AI LLM coding debate is you're beginning to find out that there are, I don't want to say a lot, but there is a contingent of software engineers that might not know a lot about code, or might be getting by with not a ton of information. To them, this might seem magical. There's an amazing writer called Nick Suresh who did a piece, he did this amazing blog called "I Will Effing Pile Drive You If You Mention AI Again." And he made the point that this is something that has its uses for the little things, but the moment you start expanding it to building entire things for you and writing all this code for you, you are just kind of kicking the can. You're still going to have to read all this code to make sure it makes sense, or alternatively, you could not read it and just hope it works. Software function when the code is bad doesn't mean it's secure or stable or efficient, or indeed that someone else coming along in the future to read it can understand the intention because there was none because the large language model wrote it, right? The sort of quality control question, which you get in research, seems to be a big thing, right? But let's put that aside for a second, right? Is this, is that solvable? Right? Is the quality control question solvable? Is sort of, we'll put to the side. It seems to me that it's worthwhile to just, for, for the next part of this, to distinguish between, is a tech useful or even transformational, and is the current financial investment in it justified, as distinct questions.
Yes. And two examples come to mind. There was an enormous railroad boom that happened in the 19th century, right? In which an enormous percentage of the country's entire GDP went into railroads. It completely overbuilt, and it led to a huge crash. And that crash led to a great depression. It doesn't mean that railroads weren't useful. In fact, railroads are quite useful, right? But it also is the case that you can have a useful technology that leads to a boom and bust. The internet being another example, right? It isn't the case that the internet proved not to be useful. It's also the case that like, there was a huge boom in 1999 and 2000. Right. Right.
And I have this thing I've been saying, "The beginning of history." It's not fully connected to Foucault, but nevertheless, it's the, I don't think it's instructive. I understand why people do it. It's how human beings work. I don't think it's useful to look at these previous booms because when you put trains on the railroad, sometimes the train didn't just randomly go up or into the ground. I, with the original internet systems, uh, back in the middle of 2025, a guy called Jim Cavell from Goldman Sachs did a great piece along with some other analysts called "GenAI: Too Much Spend for Not Enough Gain." And I paraphrase the title there. And he made the point that when the, the internet did cost a lot of money, at the $64,000, some micro-system system service, nevertheless, that capital outlay was completely different. But there was also a very clear path to utility. It would be the dispersion of fiber optic cable. It would be the access points, the actual things being built, so that people could get to the internet, and high-speed internet on top of that. The same thing happened with smartphones. Note that in the early 2000s, there were clear road maps: smaller Bluetooth radios, smaller GPSs, smaller chips, smaller batteries. That would lead to smartphones. No such path exists for large language models. For that example to make sense, you would have to have a way in which the cost came down and the hallucinations went away. Neither of those appear to be happening. And indeed, the efficacy of these models, their actual outcomes, it's actually very difficult to measure them. The benchmarks are deliberately created for them, and all of the benchmarks for software engineering are focused on one programming language, Python, and very common GitHub issues. So to train for more things, they're having to create specialized data. They're going to have to do that forever. And even then, it isn't obvious if it's actually fixing things, right? I mean, this is this problem of, are they just basically, are they training on the test data, right? Are they, are you basically saying, "Here, take a look at all this this data, and then we're going to test you on it, and oh, lo and behold, your your performance is good." Fun fact about that, they actually found that one of the Anthropic models had just started going and looking for the solution. Wasn't trying to solve it, just went on GitHub and did it. Now, people mistake this for intelligence. No, you asked the thing to do a thing, and it did a thing, right? It's just, it's doing the functions it was told to do.
Okay, but that's a great example because, like, a year ago, it couldn't do that. I mean, it is doing something new, even if it's going to GitHub. Right. GitHub for people that don't know is a sort of this sort of open-source library where where people share and where you can host projects. Yes. Where you host projects and people share code. But like a year ago, it didn't do that. Right. It used the web search tool. It used the tool it's had for a while. Perhaps it did something new, I guess. But it, it's a relative, it's a lateral improvement. It's an improvement on the thing it's already doing. It's not making unique software. Even the coded code things you're seeing where people are spitting out websites. There are tens of thousands of website templates and open-source software projects that they're replicating. It did, this is a little bit, I won't get too in the weeds of it. They did something called a C++ compiler, and Anthropic said, "We made this, we did this." It was a clone of an open-source project, and it was less efficient. It was something like 10,000 times less efficient. Which is crazy. And it's these things don't make novel ideas because if you just look at what has already happened and say, "Well, based on this, this will happen," right? You'll never, you can't get out past the arithmetic statistical average, essentially.
Exactly. Yes. So let's talk about the, the scope of the money here. Basically, paint a picture of how big this bubble is that you say is a bubble, where the money's coming from, and how it's flowing.
So it's around a trillion dollars now, I think, by the end of the year, if you think about all of the venture capital funding, all of the money that's been put into data centers, all of the capital expenditures from Microsoft, Amazon, Google, Meta, and the money flowing through Taiwanese server companies like Hon Hai, so Foxconn, and Quanta, and all that. The money is coming from a few places. It's coming from venture capitalists, and I can get into the crisis there soon, private equity, and specifically private credit. And actually, a lot of the money is coming from Japan: Sumitomo, so SMBC, and Mitsubishi, MUFG. I swear I'm going somewhere with this. But the money is coming from private equity, private credit, venture capital, and in some cases, the hyperscalers themselves. But, and most of it's flowing to like three companies.
It's also, I mean, it seems to me that it's also, right? So when you're when you're talking about AI and Anthropic, right? They need to raise capital, right? Absolutely. But places like Google or Microsoft are spending, I mean, Google just throws off a ton of cash, right? So they've got this arguably the most profitable business in the history of human capitalism, and they can just sink that cash into more and more investment. The problem is that's slowly not becoming true. Amazon, I think, is raising tens of billions of dollars of bonds. Google already did the same thing. Microsoft. I probably will at some point. They're no, Microsoft, I think, is the only one out of them that is no longer that is, uh, not using debt. That's no longer just using cash flow to pay for this. I see. Because none of these businesses are profitable. Not a single one of them. What's really interesting is none of them talk about the AI revenue. None of them. Microsoft mentioned it in two quarters, last quarter of 2024, first quarter of 2025, and then stopped mentioning it entirely. IBM just stopped mentioning their AI revenue. It's, are they shy? And so when nobody wants to talk about the money, and nobody can really precisely describe the outcomes, that's when people should get a little concerned.
There's a B. So there's, there's, if we talk say a trillion dollars, right? And the idea is you're investing all this money, and what does the investment go to? Like, what is it, what needs to be built that all this money is sunk into?
So there's two things to look at. There are the AI companies. So the OpenAI's and Anthropic's of the world, and then the hyperscalers. And so let's talk software and hardware. Okay. So AI companies like Anthropic, it just came out, Krishna Rao, the Chief Financial Officer of Anthropic, in their case against the Department of Defense, just said that Anthropic, through March 2026, for its entire lifetime, made exceeding $5 billion. They've spent $10 billion in that period on training and inference. Inference is the creating of the output. Fancy word for that. Training is this word that's meant to conjure up in your head this idea of research and development. Training in large language models can mean everything from pre-training, so feeding a bunch of information, to post-training, which is everything from, "We're going to give you some stuff and test the outputs," to minor tweaks to stop something called model drift, which is just when a model that is trained on static information will eventually become irrelevant. So, you need to keep updating it to make sure when you feed it something, it understands it. And this, there's actual, you know, huge human intervention here, which is like, "No, that's wrong, that's wrong, that's wrong." Because you have to kind of train the model to to to learn. Exactly. You've got human human trainers who are training the models themselves, as in, model gives an output, and they go, "That's a good one, that's a bad one." Then you've got people literally creating training data. Now, where do they spend that money? So this is the top layer, the AI labs. Those ones are spending it renting GPUs from Nvidia, which are usually, in the case of Anthropic, held by Amazon or Google, or in the case of Amazon and Google, their own custom silicon, TPUs for Google, and, uh, Trainium and Inferentia for Amazon. If I'm Anthropic, I got, I got labor costs, right? I got employees. And then I have to to do all, all the stuff that I want to do, run these models. It's very, very computation intensive. Yes. And in order to do that computation, I need physical hardware. Yep. Um, the, the so-called GPUs, which are the chip that Nvidia and others make, which is this sort of, um, sort of frontier next-generation processing chip, right? Yes. And the way that it works is that Claude and AI rent, rent that that hardware. So it's crazy how much it costs as well, because you may think they want to say that inference is profitable. No one's actually proven this. It's actually quite expensive to provide a user a service. The other problem is coding models, especially, are incredibly computationally expensive. You've got one user who might be tying up six to 12 GPUs, each one costing $50,000 a piece or more. You've got, and the more, what's crazy, and what really makes this different to most software eras, is that your most excitable customers are the ones that cost you the most. And in all of the cases of the AI labs, they're subsidizing them. Claude Code, crazy fact, researcher called Shell found this. For every dollar that someone is spending on an Anthropic subscription, when they use Claude Code, they can spend anywhere from $8 to $13.5 worth of compute costs because Anthropic is subsidizing them, right?
So, let's, let's stay on Claude Code because this is important on the business model, right? So Claude Code is, people are have been, you know, crowing about it, and and every, almost every engineer I talk to is using it. Uh, you can do things where you're basically giving it plain language instructions and it's coding for you. The back end of what it's doing is extremely compute intensive. Yes. And the expense of that is renting the GPUs, the electricity, right, the storage, the server space, right? Those are the basic, usually you pay the company like Google or Amazon directly, but that's the cost. So that's the, the business relationship is, I'm, I'm Anthropic, and I'm paying some other company that's doing all that backend stuff, right? And the cost of that thing I'm paying them for, right, is like, can be like $13 for every $1 I'm getting in revenue. I mean, I think about this a lot. I use this example in another conversation just with, with Google and Gemini, where if you say, "What's a good Korean restaurant in Brooklyn?" Google will show me a Gemini response at the top, and then like there's a Reddit thread that's like, "Great Korean in Brooklyn," right? The computational, the like actual resource cost of just going to the Reddit thread is essentially zero, basically a tiny amount. But the Gemini cost was like, pretty significant to go generate all that computation in the back. And you've scrolled right past that and gone straight to Reddit because you trust a person way more than you're going to trust Gemini, right?
But, but, but so the point is, even if this thing is producing use, like in the Claude Code case, one of the things, one of the things, one of my understandings of your main argument here is that the current model is, they are wildly subsidizing because the compute is so expensive and intensive, in order to make it work, they have to wildly subsidize it on the on the consumer end. Yes. So, really simple explanation. Anthropic has two, and OpenAI has this as well. Two different kinds of customers. You've got a customer that pays you a monthly subscription, and you pay through an API. It just means connect the model to your thingy. Right? Now, when you use Claude Code, you're just paying a monthly subscription, $20, $100, or $200 a month. And then you have arbitrary limits that Anthropic doesn't really specifically say. But if you were paying on the API, so if you're paying for the tokens directly from Anthropic, you would be paying not $200 a month, but $2.5, $2,700 a month. Gotcha. Right. So, so these subscriptions are essentially massively marked down to get customers who subscribe. Right. But there doesn't appear to be a way that you can, I don't think anybody that's paying $200 a month is going to go, "Yeah, I'll pay three grand. That sounds great." I don't think that'll happen. And it, what it is, is an attempt to graft the previous business models and use the previous growth trick, which is the initification, you know, the cheap monthly fee that they can then rise and then they'll find ways of undercutting you.
Yeah, that was going to be my next, my next question, right? So this idea of, you subsidize users on the front end, you sort of lose money on every customer, you get enough market share that you could then get price power and increase. This is famously what Amazon used, which, you know, lost money on every customer and every book it sold for a shockingly long period of time, um, and achieved pricing power. And it's also, in Uber is another example, right? Where, you know, people remember this time when Uber came about, where you could take an Uber like five bucks. Yeah. I remember landing in cities when I was doing like business travel. I mean, because New York, it was always like, yeah, there was, it was relatively expensive, but, but still pretty cheap. But then sometimes you'd land somewhere and be like, a $6 ride from the airport to the hotel. You think to yourself, wait, how is this making any sense to me? This can't possibly be the case that anyone's making money out of this. But in comparison, it would be if like, every Uber driver cost Uber $50,000 a day. It's the economies of, and the economies are just completely different. When Uber was subsidized, I think between 2019 and 2022, when they became a kind of messy profitable, like, not a great one, it was maybe 32, 33 billion, which is a lot of money. Amazon Web Services, arguably one of the most, the single most important technological innovations ever, mostly done through just money and time. Though this isn't adjusted for inflation, in the 11 years from I think 2003 onwards, they spent $38 or $39 billion in capex. For some context, OpenAI raised $42 billion in 2025. So you're saying the scale of the subsidy here is just way bigger than those previous ones.
That's the point. Yes. And the underlying infrastructure, everything is more expensive, and it's not getting cheaper. Is so, if, okay. So if we talk about the, the, the front-end model makers, right? That they're they're subsidizing, you know, even in this filing, right? Anthropic, $5 billion of revenue, $10 billion in expenses, obviously that's not profitable. And that's just the compute. That's just the compute. Are so then there's the hyperscalers, right? Which are the, the physical owners of that are built. And these ones to be the ones that are building the data centers, right? In some cases, there's a lot of independent ones now that are building data centers in the hopes that AI demand arises that doesn't exist. And people like CoreWeave and Nimbix and such, who are just things called NeoClouds, they just are warehouses full of GPUs that are technically data centers. And do they, so let's say I'm one of those and I build a data center, do they then have a business relationship with one of the intermediaries like Amazon, or do they directly contract with Anthropic?
The answer is yes. So some of them do. Some of them, it gets even more complex. We don't need to go into it. There are people that rent the data centers who then sell the stuff. But nevertheless, that's actually kind of the problem. When you look at who's paying for AI compute, and you actually really go and look at like, who's paying the money, there are really only two kinds of customers: Anthropic and OpenAI, or hyperscalers. Meta, oh sorry, Nvidia. Nvidia has agreed to spend in the next five years, $26 billion in AI compute deals. And I don't think it's a good sign that the shovel seller is also paying for the digs.
Wait, no, wait. Take a second because it's going too fast. So you got, so this, this is why, so Nvidia buying compute is weird for this reason. I just want to walk people through this real soon. Yeah. Go on. They make the chip, right? The chip is the thing you sell to the person that's going to say, set up a data center, right? So in the ideal world, I'm in Nvidia. I sell a chip to the data center. The data center buys it from me because I make the useful thing. And then the data center sells its compute or rents it to one of the models. Right? If you're selling the chip, why would you want to be also buying the power of the data center? And Nvidia has made a deal where they're basically going to support the construction of a lot of data centers. Yep. Meaning they're going to be buying their own product essentially. Yes. They're feeding money to themselves, right? So here's some money for data centers so you can buy a bunch of our chips, which is a little bit like it seems a little bit like you're just paying yourself for something. So, I'll give you the, the really, the one that I think is going to blow up nasty. A company called CoreWeave, AI compute company. They're a public company. I think they they lose money hand over fist, and they have tens of billions of dollars of debt. And they're making, what are they doing? They just build, they have buildings, they fill them full of GPUs. Nvidia invested in them. Okay. Nvidia propped up their IPO. Nvidia bought $2 billion worth of stock recently, and Nvidia is also one of their largest customers. Yeah. Their other customers are OpenAI, Microsoft for OpenAI, and Google, unsurprisingly, for OpenAI. I'm not kidding you. Google is renting compute from CoreWeave to rent to OpenAI. Oh, wow. So, there are people that are doing compute middlemen where they, they rent, they rent, and then they rent it out to someone else. And I don't want to get too deep into it because we'll be here forever. But there are also colocation companies who build data centers to rent to CoreWeave to rent to someone else. It's, it's really bad when you actually look at the non-hyperscaler or OpenAI compute. There's less than a billion dollars of revenue on $178.5 billion of data center credit deals done in 2025.
Say that again. There's le, who, who has less than a billion dollars of revenue?
Everyone. As far as people paying to rent GPUs, when you remove all of the hyperscalers and OpenAI and Anthropic, right, it's less than a billion of revenue last year.
But doesn't that just mean that the big ones are driving all the business?
Yeah, but the big ones are also not talking about how much money they, and in fact, the big ones are losing all the money. And the one spending the most money, OpenAI is also burning so much money they need to constantly raise billions of dollars. Some of it coming from Amazon and Microsoft and Nvidia. At some point, you got to wonder if it's just the same billions being cycled again and again, right? And nobody making a profit other than Nvidia. Nvidia is just printing money. Okay, that's the one place. So, there is one place in this that that people are genuinely making a profit, which is, you know, I always use this example. I make a sandwich for $2 and I sell it to you for $4. Right? They, Nvidia makes a chip for X dollars and they sell it to someone for 2X or X plus Y. They are definitely making a lot of money. Yeah, the panini press guy, the panini press maker, they are making the money, but the sandwich costs a dollar and they're, it cost them $10 to make. It's really bad. And their only customers appear to be themselves, or a very small amount of AI companies, all of whom are terribly unprofitable. Right. So they're, they're definitely making a profit. And what you're identifying as the weakness is the people they're selling to are not making a profit. So Nvidia can make a, is definitely making a profit. They're booking profits. That's great gross margins as well. Inarguably true. Their stock is gone up hugely because they're doing that. What you're saying is the people they're selling to are not making a profit. And at a certain point, they can't keep buying if they're not making a profit. Yes. And then the people that are buying their chips, that they're selling their compute power to, which are the models, are also not making a profit. And so at a certain point, the music ends and people go diving for the chairs. Because if the models aren't profitable, then they don't need the data centers. And if the data centers aren't profitable, then no one needs the chips, and the whole thing collapses.
There's also one abstraction that makes things a little worse, which is when I say there isn't, there's less than a billion dollars last year of AI compute revenue outside of the hyperscalers, what I mean by that is it doesn't suggest there's actually much revenue potential in renting an AI data center. $178 billion, Bloomberg reported at the end of last year, $178.5 billion of data center credit deals. So debt were done in America alone last year. May even be more. That's a lot higher than a less than a billion. The other thing is all of these data center debt deals are basically done by new companies. So all of the debt's kind of crap, right? So these new companies, basically what's happening is a bunch of new entrants are saying, "Hey, I can find a warehouse, get a bunch of GPUs, find electricity source, make a data center." They're entering the market and they're floating debt to make these new data centers with the idea that when this all takes off, they're going to have a steady diet of customers they can sell the compute to, because compute demand is going to go up and up and up. But if the demand doesn't go up, then that collapses, and they won't be able to pay the debt that they raised from private credit that already has issues with, uh, people not paying their debts because their due diligence wasn't so good. Right. Right.
So the, so, so the way that this, your understanding of of the sort of vector that this gets into something that's a larger financial problem is of how much of this paper, you know, that there's there's a lot. Basically, your contention, your thesis is that there's a ton of bad debt floating around. Yes. And also, this is all happening in a historic downturn in venture capital and in private equity. Since 2018, PE, uh, venture capital has failed to on average. There are still some success stories, of course, to have a TVPI, total value put in, of higher of higher than 1.8 to 1.2. Sounds complex. It just means for every dollar you invest, you get somewhere between 80 cents and $1.20 back. That's not very good. The S&P 500 are beating the crap out of that. That's happening with venture capital. Private equity is also having the other problem, which is private equity is having trouble selling their companies. There was the massive rush in the kind of software era, the run-up there, where private equity bought an absolute crap ton of software companies. 30 to 40%, the co-president of Apollo said this recently, of private equity deals between 2018 and 2022 were for software companies, which means the private equity firm and the software company took on debt. And after that, of course, we had the, well, we had the 2021 era, the massive amounts of insane crazy deals. The metaverse era, ton of really bad companies got bought for 30, 40% higher than they should have been. So, you've got private equity and venture capital sitting at this time with a bunch of stuff they can't sell, which means they don't have liquidity, which means that they can't invest quite as much, and indeed they themselves might have debt they have to pay. This is happening at a time when technology and the infrastructure behind it, referring to AI, needs more money than it's ever needed ever. That's that's the thing.
So there's two parts of this I want to push on. So one is, if, if you think about this idea that, look, we're going to take on a lot of debt to build something out in the future that isn't profitable now, but will be. Okay, fine. That people do that all the time. That's like, that's kind of the risk of investment. That's the risk of investment. People do that all the time. That's the basic model here. So then the question is, okay, um, one is, can the co, right now it's very expensive and compute intensive to do this, but maybe it won't be in the future? And what I think is interesting about that question is that might be a really good thing for Claude or OpenAI, but if that were true, it's going to be a bad thing for all the data centers and the GPUs. Right? Like, the principle right now is you need a lot of computing power. The computing power is being populated with these huge physical infrastructures and enormous amounts of investment. But maybe we'll figure out a way, there's some evidence that, you know, one Chinese model has done this, that you don't need all that computer power and you can still get really, you can still get the same results. Even though that would seem like a great innovation at some level, if that were true, it means that all of that physical infrastructure is no longer needed or valuable. Right? So we can get back to the fact that it isn't getting cheaper.
What DeepSeek did was they trained cheaper, but the cost of inference is still going up because even if the model is, what do you mean by the cost of inference?
So the cost of inference is when you, the amount of money that it costs to create an output. So you will see that models, some models have got cheaper. People conflate that with it, with the companies themselves finding a cheaper way of doing this. They've never said that. They've just brought the price down. They can afford it when they can raise $5, $10, $30 billion at a time like Anthropic just did. What DeepSeek did was they were able to train a model for cheaper. Right. That's the Chinese company that sort of shocked people, and there was this big hit that happened to the market because of it. Yes. Well, like they, they were able to sort of shortcut this this process. Yes. Because they couldn't access the latest chips. But putting all that aside, the other problem is that pre-training, which when you shove all the data in, stopped having the same results. We kind of hit the diminishing returns point. So their only way to make these models do more was to burn more tokens. So even if a model cost comes down, you're using more tokens to do the same thing. You're spending more money as a user. We don't know what it costs them. They're all unprofitable. But to your point, you're completely right about these data centers. They also have another problem, which is, it takes about two years, three years to build an AI data center. Nvidia is selling new chips every year. This seems like a big problem, a depreciation problem, right?
Well, the depreciation problem is one in that they, they burn out in three to six years. We don't really know yet, but I've heard crazy failure rates like 10 to 20% within a year. But we truly don't know that. It's both the depreciation problem and the fact that let's take Blackwell, released kind of in 2024, but really in 2025. We still have data centers being built like Stargate, Abilene out in Texas for OpenAI, and Oracle that are using Blackwell GPUs that by the time that bloody thing's built, which will be 2027, they will be two to three years old, right? You will be have an entire data center full of obsolete GPUs. And all of the GPU data centers being built last year are going to be Blackwell. So you have just, Blackwell is what the, the, it's the current gen. The new gen is Ver Rubin. This is just the GPUs you've, Nvidia's. Yes. Yes. Sorry, I should have said that. So, you've got all these these data centers, and now you've got this flood of supply of an obsolete chip. I don't know. I, I ain't no economics knower or anything, but generally when the supply increases, they have to lower the price because everyone's got it. And you're already seeing the price of renting those GPUs come down, right? Because they're, because they're, they're older chips. And so they're going to, the same way that like, you know, a newer car sells for more than a than a used car. But also there are more and more of them coming online any every day. Right. Right. Right. Right. And also we don't know it. We don't, I don't even think they're profitable for the providers to run. Like, it's really, we don't, there is compelling evidence that no one's making a profit renting them, which is crazy. It's crazy we're all doing this and we don't know that for sure.
Wait. Meaning the, the folks that have the, the centers, the actual data centers.
I hear it's ruminant, so I can't confirm it. I heard of a data center out in North Dakota that was losing a million dollars a day. That's that's not a good business. And it's crazy because, so all right. So, so, so let's say, so one problem is the time scale for building the data centers is being outpaced by the new chips. You're building things that are obsolete. There's also the threat that happens of, um, that you actually come up with more, maybe you find more efficient ways in which you would, you have sort of stranded assets, right? That you have all these data centers, it turns out you don't need all this compute because we've come up with a, a more efficient way to do it. But again, the, the, the story that the AI people are telling, "Invest now, it's not profitable, keep building." And if we get to something that can, for instance, do what a first-year associate at a law firm does. Then you have a situation again, I'm just, this is the, this is the case, right? The case is, you got a situation, we hire first-year associates, they largely do things like doc review and they draft memos. And we're going to have a model, it's going to be trained on legal stuff, it's going to be a, you know, enterprise system that Claude charges $60,000 a year for. A huge amount of revenue would be like the most expensive software, basically. The business case here is that, it's you. You hire a first-year associate for $120,000. We charge you $60,000, right? We're making a ton of money. You're saving $60,000. And it's a bummer that the first-year law sort of gets out of a job. But if we could do that at scale, if there's millions and millions of these kinds of jobs that people are making high five figures to six figures that we can sell you software to replace, I mean, again, this is the contention of why it would be valuable. This is the core contention. Like if you, if you look into what these companies are saying.
I guess the question then becomes, is that a plausible outcome? Because I think if it is plausible, you could probably make the math work. And if it's not plausible, then you can't.
So I actually, the law firm example is great. The problem is, I don't think enough people know what people do at jobs. Law firm associates make law firms work. Law firm associates are doing the work that partners don't want to do. And if any partners are listening, you know I'm bloody right. So yeah, if you, if you what you're describing there would be AGI, just this conscious computer, which by the way, everyone's real excited to control a conscious creature that's just describing slavery. It's important to say what AGI is. It is describing slavery. And it's right. You're saying if you achieve what they call artificial general intelligence, we actually just had a conversation about this about with David Chalmers about consciousness. Um, that then you're actually, there's all sorts of moral implications of what that device is once you slave, right? Yes. It, it's literally. But anyway, back to back to the law slave. So this theoretical thing, yeah, if it could do literally what an associate did, sure. But an associate does much more than just doc review. They're doing the work.
of research. And it's not just I found a thing, right?
Look, it's drafting motions. If you get a motion wrong, a judge will sanction you and you will embarrass the partner. What you're paying for with employees in many cases is actually risk management and judgment and judgment and taste and culture and also risk management. You're handing the risk off to a human being that you can rely on and train.
Also, how are we going to make partners if we can't make associates? We're just going to hire a law student to become a partner. I mean, I don't know. I could sit around handing other people's work and talking. No, that's that's not true. Partners do all sorts of work, I'm sure. But nevertheless, yeah, in theory, if you could replace 10 $60,000, $150,000 in the case of a law student, you could replace 10 of them with $60,000. Sure. It isn't doing that.
And large language models are sold as the reason I mentioned the thing earlier with AI boosters. They need to be legally banned from saying in the future it will could, right? We need to talk about what's happening today. It isn't doing it. It isn't doing it. And in fact, every single example I hear of in specifically law, it large language models being used ends up with someone getting in trouble with a judge. I think they just had a DOJ person that this happened to as well.
Well, I don't think that's true that every example because the people are using AI all over the legal world, I can tell you. But there definitely have been um hallucinated citations that have been filed and I think in some cases even by government lawyers, the DOJ, uh that have been caught that where they're they're citing to a case that literally was invented by the AI. So, but to your point, yeah, if you could do the thing it doesn't do and has no proof of doing, yeah, sure, grandmother had wheels should be a bicycle and so on and so forth. A lot of this, in fact, all of this is really sold on the coulds and shoulds and wills.
Yeah. It's not sold, it's a bet about what its future capabilities are based on what I would describe as semiotic knowledge. Logic even. It's this idea that because things have worked this way in the past, it'll happen before. There was a time when the internet was slower, then everyone's internet connectivity went up. Not really the same thing because the technology was always there to get it faster. Fiber optic cable was there. The massive overbuild was there. This is not a problem that you solve by having more compute. It is not. I mean, they think it is, right? I mean, that is the I mean, just to be clear about what the distinct disagreement is. Their contention is that they have found a reliable and straightforward law of scaling, which is that the more compute you have, uh, the better it gets and the more that it starts to act in ways that are intelligent.
Except the scaling laws are broken. That diminishing returns I mentioned earlier. It's no longer getting the same kind of improvements just by pre-training them.
But isn't it get I mean, I just got to say like this is where my experience of use of the models is that they're getting much better at what specifically multi-step tasks in research. So I if if you use it for research so for here's a here's a great example. You go to Claude and you say I said this the other day. I'm trying to figure out the relative homicide rates in major American cities in the 1890s. I want to look at New Orleans, which is what I'm writing about, and compare it to New York and Philadelphia. A year ago, with previous models, you would have gotten essentially nonsense, or you would have gotten like, well, here's the Wikipedia. Here's a few things. In this case, it like went through it found like there's two like real sources on this. Like there's a book about southern homicides. There's another book about northeastern policing. I know this because I've actually done the research. Right. Right. Um, it goes through it basically does find in one of the books because it's in public domain what the New Orleans homicide rate is. It talks about what the data difficulties are in New York, Philadelphia. And it basically spits out an answer that I can check because it's citing it. That's basically correct. Okay. In New Orleans, it's 25 out of, uh, you know, I forget, 25 out of 100,000. And in New York and Philadelphia, it's five, something like that. A computer could not do that a year ago. Like, no, it just couldn't like it. Now, there's all sorts of ways in which I can check it because I have the expertise. But this was like a sophisticated multi-step thing that it had to go through and and and sort of use a bunch of powers that it just didn't have a year ago.
I mean, you had to check every step though, didn't you? You had to go and check all the data. I did have to check the citations. Yeah. So, at some point, I But so, what you're describing there is an improvement. They have found ways to connect them to web search tools, right? These things are able to drag stuff, but what you're ultimately describing is more sophisticated but less reliable search. It is an improvement because they're able to post-train it in that case. and say this result is bad, this result's good. I've used even the most sophisticated ones used by hedge funds, the searches. The problem is is that yeah, it will get some things right and it will find the occasional thing. Oh, you didn't see this in a 10K from 2 years ago. Problem is you have to check every single bloody thing. You can't rely on anything. You can't rely on a single thing. You perhaps it helped you get in the right direction. Is that worth this much money? Is the and is I mean, what you're describing you were seeing in models middle of 2025. I guess it's I guess we did something wow, we have better search, more sophisticated but less reliable search.
Well, or multi-step things. I mean, the thing the thing to me was that this is a fairly like compound task, right? So it has to do it's it's got to do a bunch of stuff and and the thing that I thought was striking was that it actually it did a good job of finding the right source, which that was sort of interesting to me like, oh, that is the book, you know, that is the book where this is contained. You didn't just like go to the Wikipedia page.
I think the thing that I I kind of come back to and this is the sort of horns of the dilemma and many people have sort of talked about this is that it seems to me that there's no way out of some kind of cataclysm for this reason. How do you mean? Either your case is correct, in which case it's just not going to be profitable and the whole thing is going to collapse in on itself. Yeah. Or you're wrong and they get a lot better and they are profitable and what being profitable means is that they can replace the labor of tens of millions of people, right? That's another cataclysm. Like the point is that like if they're right, if the thing that they're promising, which is like, oh yeah, we could start getting rid of all these people that do all these jobs and replacing it with AI, that's good for the profitability of these companies, but I think it's probably insanely destructive to America, the macroeconomy in American society. Sure. And I'm not afraid of that because just you don't think that's going to happen. You see no signs of it, right? I think that but it is the only way like the only way it would make sense. That's the point is the point is that like for the math to work out it has to be something pretty darn revolutionary and it has to be trillions of dollars. Like I worked it out mathematically by 2030 for any of this to make sense for Microsoft, Meta, Google, and Amazon, they need $2 trillion of new revenue, not enhanced revenue, I mean brand new, brand spec and new dollars in a software industry that has never been higher than $700 billion of yearly revenue in in an annual US GDP that's like $30 trillion, right? So you're talking about just 10% of the world, right? An enormous part of the entire economy and it needs to happen in the in the next 6 months.
There is one other thing though. Yeah. I think that there is a social contagion that will happen with this. Look around the world of bosses right now and the amount of them who are like, "Oh yeah, I can't wait to replace everyone. I'm going to replace all the actors in my movies. I'm going to replace all my workers. He's just going to give me money and then I'm gonna have all the money and the pieces of crap I sell things to as hogs praying for my slop. The excitement in it. But also, how many of them are just wrong? How many of them just say things that aren't true? We have people in newspapers saying things about AI that aren't true. We have bosses claiming things about AI that aren't true. We have people lying about it. It's truly obscene. And regular people know. Regular people like if you go and talk to like electricians, HVAC people, hairdressers, teachers, their reaction to this is horror. There are some who are using it to cut corners. Everyone wants to do that. Human beings do that. Well, there's also a lot of people that are like I mean, there's also a ton of people that have like crazily intense parasocial relationships, talk to it all the time. I think that there should be criminal tribunals for the companies that it's disgusting. Anyway, I think that we are going to see something happen before the economic stuff as an outcome of it actually where regular people have seen who how their bosses think of them and it's happened for years. You saw it with remote work where bosses were like, "Hey, you got to get back to the office, man. I you got to get back there. I got to be able to look at you every day. I got to be able to stomp around so you can feed off my mood. You got to go to the metaverse now cuz that's where I'm going to be. Have fun staying poor. You weren't in crypto. Also, I'm replacing you with AI." So you've got that and then you've got the other thing which is it needs to make all this money now now I'm next 6 months.
OpenAI even then raised $110 billion. Actually, they only raised $15 billion. $35 billion of the money from Amazon is due when they achieve AGI or go public and, uh, both the $30 billion from Nvidia and the $30 billion from SoftBank are being paid in $10 billion tranches like like Cler. Um, not literally though. And what's funny is SoftBank has to raise $40 billion in loans to pay for their part. Everything that's happening is a stress test of debt and equity. How much can venture capital spend? How much can hyperscalers afford? How much money is left in the coffin?
So then what's what is that out of your theory? What emerges as a prediction of the first place that you'll see a crack, a fault? Like what what would be the first sign? We're already seeing it with private credit. So, I kind of hinted at it earlier. Private credit, private equity, massively bought so many different software companies and when they bought them with these leverage buyouts, they bought them, pumped them full of debt, and then took on debt to buy them. I've heard something ridiculous like hedge, sorry, private equity firms are leveraged to four to six times the value of their assets. So, you're already seeing it. There was a stat that came out the other day. There's $42 billion of software loans just for software companies that are in distress status. So not likely to be paid, right? You have across the board, you can go and look inside there. They have to publish this private equity firms, private credit firms and BDCs, business development companies. Basically the same thing. They're suddenly starting to take payment in kind for loans as in you get stock, you get given stock, and then you just kind of put all the cost onto the end of the loan. They're not getting paid on these loans. These loans are going, they're starting to default.
So, what are the So, you you think you're going to start to see loan defaults rip like the private credit market is going to ripple first. We're already seeing it. And then I think and my real my three horsemen are you're going to see a data center project fall apart before it's complete. You're going to see an in-construction one collapse and then you're going to see a a fully constructed one that has to shut down cuz it runs out of money. Because remember, these things are debt. They they are heavily debt. They are full of debt. There's not a single one of them that is even close to profitable before the debt and then you add the debt on top.
I So that's that's interesting. So those three things to look for that in data centers. Exactly. And I think that because the AI bubble in my opinion is a symptom of the largest death of software as a growth model because the assumption was software eating the world, Mark Andre, uh, was that every industry could be software-tied, which is true, and that as a result, all of them could grow forever. Private equity, venture capital bought into this. They invested in all these companies except now nobody will buy the companies. M&A has died and it's harder to take them public. And if you can't take it public and you can't sell someone else, what do you do? Well, the answer is you sell them to another private equity firm. So, there's currently a game of hot potato called continuation funds or secondaries depending on who you ask. Eventually, no one's going to have the money to buy the thingy.
And well, the other thing is that at the core of all this, whatever financialization you're doing, M you have to make things that are profitable at the bottom of it. Yes, you do. I mean, that's that like whatever financialization happens, you can you know, you can have periods of where you buy an asset and you bloat it with debt and you sell it to someone else, right? And there are ways in some intermediary sense that people can make money off passing things hither and yan. But in the end, you got to make money. Things have to be useful and make money for everything to keep going. And the thing is all of these investments and Apollo's John ZTO, I think his name is, he said that the whole thing is is that these software companies were acquired or invested in based on the idea that they would grow like they did between 2005 and 2018. No, they grow like after 2018, there's less money and there's only so many software companies you can build, only so many people you can sell to.
The last thing to me that also seems possible just to end it here is that, you know, there's this sense that like sometimes they'll talk about, um, their own vision of artificial intelligence being like a utility like electricity or water. And what's interesting about that is that like utility companies aren't that profitable and they're not that sexy and they're not like it's a strange thing because it seems to me like maybe it is possible this becomes like a utility that like everyone sort of has it or has access to it. But if if that's the case, then it's basically just a commodity. Like it's not it's not actually like utility companies are not super profitable. In fact, they're regulated. And that's the thing, the utility thing doesn't really make sense because you water is water. An AI model is many different things run by many different companies that needs constant maintenance to avoid model drift because otherwise it's just a static object that cannot respond to new things. You don't need to continually make sure power is power before it turns into something else. Yes, there are power of regulation things. I know.
But one other very scary thing to add, I don't like scaring people, but some of the money that's going into data centers now more and more in fact from like Blackstone and so on and Harry's, I hear is from insurance funds and retirement funds because they've needed more liquidity than they had. So they've started investing in private credit loans, the data centers and the sales pitch is simple. This is the future. This is a good yield. You'll get paid more on your money. Without the worrisome thing of will they always be able to pay back their debt? And we mentioned Oracle earlier. Oracle is a mess, but Oracle has taken over a hundred billion of debt. They had negative cash flow of $24 billion and they are building 4.5 gigawatts of data centers, hundreds of billions of dollars, and they're building them for one company, OpenAI, who lost at least $8 billion, more like $10 or $15 billion last year and expects to burn $230 billion by 2030. Oracle will die if or if OpenAI dies. And this is not this is not a this is not me catastrophizing. Mathematically speaking, Oracle cannot pay the debt if OpenAI does not make more money than Nvidia does right now by 2030. Yes, that's the the bet and they'll say it themselves. Dario will say this and Sam, Dario Maday is that their their bet is on a revenue trajectory that is essentially unprecedented in human capitalism. They think they can achieve that and if they don't, it doesn't work.
Ed Zitron, host of the Better Offline podcast, writes the Where's Your Ed at newsletter on Ghost. Great to have you here. Thank you for having me.