S&P earnings dropped 40% from mid-2000 to mid-2001. People forget that, just how much, how fast everything collapsed as every single enterprise cut back and said, 'We don't need 10,000 routers, we need 1,500. We don't need 10,000 switches, we need 2,000.' That happened throughout the entire tech economy in the space of basically four quarters. To think that it can't happen in this boom, which is much, much bigger and it's accelerated much, much faster, you have to be kidding yourself. There is very big risk of financing in this market that will make all of this reflexive and will look like a crash where long-term secular demand could still be growing.
Welcome to the Risk Reversal podcast. I'm Guy Adami. This is my friend Jim Chanos. He's the founder of Chanos and Co. I think this is like maybe your 10th appearance on the podcast. Jim, you were also a fab professor, okay, at Yale University and at University of Wisconsin. You teach a class on financial fraud at both universities. Thanks for being back here, man.
The history, not how to commit it.
Hey, do you go to the... do you go to... We're going to get to you in a second. Do you go to the Harvard-Yale football game or no? Or just a lot of Wisco games. That's your deal.
I have been known to go to Harvard and Yale football games. Yes.
And I'm sure you're a great tailgater. All right. This is Gary Marcus. Now, this is one... this is really interesting how this kind of came together. Like, hundreds of thousands of people read Marcus on AI. That is your Substack. You're also a multiple-time author. Taming of Silicon Valley. You also testified next to Sam Altman in the Senate in 2023. That sounds almost cute, like a long time ago. We had different views of Sam Altman back in the day. And you're also an NYU professor emeritus. Gary, welcome to the podcast.
All right. So, this is kind of serendipitous how this came together. Jim was going to be on the podcast already. You graciously reached back out to us. You said, 'You're going to do the podcast. You want to do it in New York. You're going to be in New York.' And we were like, 'Okay, let's do this together.' Each of you have... I don't know, Jim, you've attacked this whole issue of the AI build from a financial standpoint. That's how you started with it. Gary, you are a technologist. Is that fair to say? You have a PhD from MIT, that sort of thing. You've been in the space for a long time. You wrote... your dissertation is this 30 years ago. Sorry to age you a little bit, on neural networks. And we all know, you know, this is... I don't know, like an early sort of... I don't know, fetus to what's going on right here. Does that sound kind of gross? I don't know.
Well, my dissertation was actually on two things. It was on how children learn language and it was on neural networks that are the ancestors of today's neural networks. So, I've always been interested in natural intelligence and artificial intelligence, you know, straight for a few decades, more than I care to admit. And when these neural networks came back in fashion in 2012, I had been mostly working on the cognitive science side of things, the human side. But I was like, I recognize all this stuff. This is what I worked on in my dissertation. And in fact, I wrote a book in 2001 called The Algebraic Mind, writing about the limits of the ways that people were then building neural networks. And I think I was actually the first person to point out that these things were going to hallucinate unless they changed how they built them. And so many of the problems that we see now, I saw quite a long time ago. And over and over again, I keep hearing this argument, 'Just give us more data and we'll solve them,' which hasn't really worked. I mean, it's solved a little bit, but not that much. So I've been observing the limits of these systems for quite a long time. And then I got into the economics of it from a technical understanding. So in August 2023, it dawned on me that basically everybody was building the same thing. And that meant that there was no technical moat. And so I wrote a pair of essays in my Substack in August '23 saying, I'm not sure the finances of all this is really going to work out. And I said, look, there's no real technical moat here. And what that's going to mean is price wars. Everything's going to become a commodity. OpenAI was riding high, but I said they're going to lose that lead. You know, you used to be in AI, lots of people took different approaches, but everybody converged on the same thing. And I said that is not going to work out economically. And that's what really brought me into the economics, was thinking about that.
Yeah. And so, and Jim, you know, the economics of the whole kind of buildout is something you were focused on like right out of the gate. And you know, a lot of our listeners and many folks who just know of you in the industry, you were the first to identify the financial fraud at Enron. And in many ways, you had to dig deep for that in ways that a lot of folks who covered that company and the industry very closely just maybe they just didn't want to know because it was obviously a huge engine within our economy. If you think about what was going on obviously in markets and performance, and there's probably a lot of similarities we can draw to what's going on now. If Gary was working or you're creating a thesis about this industry based on his academic pursuits and then how he was writing that sort of thing, how did you come to it? Because in many ways you look at a lot of different sectors. You're known to be like, you know, really focused on financials and the like, but this stuck out to you because of the economics of the industry.
Well, we had looked at the legacy data centers as REITs back in 2022 and got very bearish on the cloud data centers based on the absolutely horrible economics of building a data center to host, collocate, you know, other servers. And the returns were ridiculously low. They still are. They're low single-digit to mid-single-digit pre-tax returns on capital. And this isn't a booming business, right? The cloud was still growing rapidly, still is, and yet these were REITs. They were basically landlords. They were capital-intensive service businesses, so they weren't just flip a switch and collect a check and go to the bank every month, right? It was actually stuff broke, stuff went down, stuff needed to be replaced, redundancies needed to be there. And so they were and are being marketed as REITs so that you just don't include the depreciation. But data centers, whether legacy or AI, are very capital-intensive businesses and ongoing capital-intensive, meaning stuff needs replacing and upgrading all the time. And so the AI data centers struck us as a really interesting idea on the short side. As AI got going, it became very clear very, very rapidly that away from the world of the models, where the real money was going to be spent, was physically in the infrastructure and hosting the models. And I knew immediately, and we've seen strong evidence, that that is going to really be a commodity business at the end of the day. I can't opine on what the models derive, and I would defer to your other guest who knows that better than anyone. But I can analyze a real estate deal. And that's where I think people are really jumping off the cliffs right now because they're assuming spot pricing is long-term pricing. They're assuming all kinds of things. And what they don't realize is that the AI data center business is primarily an equipment leasing business. I'm buying chips from Nvidia and I'm renting them to you. And that's a financial construct.
It seems like investors, at least in the public markets, have gotten a little bit hip to that. We've seen the way that some of the NeoClouds have traded. And then, you know, the kind of mother of all NeoClouds is Oracle. And we're going to get to all of that. But Gary, when you started focusing on the here and now, and you said August of 2023, did you kind of approach it, well, maybe this could be great, maybe this could work out, and we'll talk about the technology, but really like the economics of it? Because again, there are a lot of folks like me who really weren't paying much attention to it, and all of a sudden, you know, ChatGPT gets started, passed around in early 2023, and people are like, this is magic, you know. And when you think about where that model is relative to the models now, did you want to be constructive on the AI industry, on the models and the pace in which they were evolving? I'm just curious, like you started a Substack back then. You were very early on discussing this on, you know, a fairly... you know, like you were talking about a lot at a time where a lot of folks weren't talking about a lot. I get an FT alert about 10 articles a day on, you probably too, on AI. I mean, it's from, you know, kind of reporting to opinion, that sort of thing. Back then it was not like that.
Well, the minute ChatGPT came out, it was like that. And that was actually December of '22 or maybe the last day of November of '22. So it was the end of '22. There had already been a lot of publicity around GPT-2 and I guess 3, if I'm remembering quickly.
Sorry to interrupt. In the public markets, in the public financial markets. So eventually...
I don't think... I don't think the financial markets cared. I think the public cared. So, you know, The Guardian had an op-ed that was allegedly written by GPT-3, but it was actually edited by humans as far back as 2020. And so it had already surfaced some to the public. And then when ChatGPT came out, which was November, December of '22, that's when it got really big. I immediately saw the limits on it. And we can talk about what's changed since then and what hasn't. But I wrote a piece on Christmas Day of 2022 in the Substack called 'What to Expect When You're Expecting GPT-4' because the crowd had gone nuts expecting that GPT-4 was basically going to be artificial general intelligence, AGI. And so I laid out seven predictions and I said, it's still going to hallucinate. You're still not going to be able to trust it for medicine. There's still going to be reliability problems. It's still going to have trouble understanding the physical world. And people thought I was an... basically, they, you know, they've called me a hater and whatever. But I said, if you look at how these things work, that's how it's going to go. And mostly it has gone that way. The exception is what I kept telling people to do is add symbolic models to make what I call neurosymbolic systems.
So a neurosymbolic system is a hybrid of neural networks, which are basically supposed to look a little bit like the brain. It's really just marketing; they're not that much like the brain. We can go into that if you want, but that's what most of the stuff around now today is, these neural networks. They're kind of statistical approximators. They try to approximate what it is that people are talking about in these giant databases. So you have those, and then there's classical AI, which just looks like computer programs. It's just, you know, if-then statements: if this equals that, then do this thing and otherwise do that other thing. And that's been around since... well, both of them have really been around since the 1950s. And there's been like a hatred between the two. Like in 1969, Marvin Minsky, who's one of the founders of AI, wrote a book with Seymour Papert called Perceptrons, arguing that neural networks weren't going anywhere. Basically, if you really look at the argument, it was structured very carefully, but nobody ever does, and they just take it as he killed the field, which is not entirely right. But anyway, there's been hostility back and forth on different sides. And what I've been saying is you need both. They actually have complementary advantages. So the neural networks are really good at statistical approximation. They're good at certain kinds of learning, but they're not good at abstractions and generalization and reasoning. You obviously need to bring these together. And as usual, people, you know, ignored me, dumped on me, or whatever. But late 2023, they actually realized there was some wisdom in this. They never admit it. And the reason for that, I think, is financial, and we can talk about it. But now they use these things like code interpreters. So they'll put in Python inside. They use things like harnesses that are if-then that cycle through and judge the responses of the large language model. So the way that this interfaces with economics is that the economic story has always been about scale. We will just make the large language model bigger, give us more money, it'll turn out great. And they don't want to admit that that's not actually what's driving progress anymore. That actually somewhat ran out by like 2024. What's actually making things better now, and some people will admit this, but many don't really want to talk about it, is these harnesses and things like that that are actually borrowed from the symbolic AI tradition. So you don't... I mean, you couldn't just scale things anymore. They kind of ran out of data for one thing. They used up all the data on the internet. They try to make synthetic data. We could talk about what that means, but people have gotten diminishing returns. So they're actually using these other techniques to some extent. And I, as a longtime advocate of this neurosymbolic AI, am pleased by that, but it's still not really enough. And basically, they have one trick now that works really well, but it's limited. Let me put something else historical, which is we for a long time knew how to build narrow AI. And the dream is general AI. And you can almost only understand that dream if you understand what the narrow AI is. And it's not hard. The dream was, or the history was, we had things like chess computers. They could do one thing well, but you couldn't trust them to do anything else. And ChatGPT at least gave the illusion of being general, that it could do anything. Now, the original ChatGPT could do anything, but it was really bad at being honest about it. Like, yes, it could write a funny poem about, you know, three guys doing a podcast in rhyme, and you're kind of impressed with that. But if accuracy mattered, you just couldn't trust it at all. So we had those kind of systems, couldn't really trust them. The trick that they're using now is basically a lot of what they do is they use the large language models to generate something and then they use the classical reasoning techniques to verify formally with symbols a solution that's right. This works great in math because we have a lot of tools. The major one that people use now is a computerized tool for symbolic reasoning called Lean. And so they take this thing called Lean and they verify a proof. The large language model proposes it and then they verify it formally, and this works great. It's made a lot of discoveries. Well, I mean, they're not great discoveries. Well, some of them are actually great. Some of them are great discoveries in math, maybe not world-changing, like a whole new idea, but at least like solving something that people couldn't solve. They're very impressive, let's say. They're maybe not world-changing, like Newton inventing calculus, but they're impressive. And they're doing this in domains where you can formally verify things, but it's still not really the dream of artificial general intelligence because most things in life aren't so formalizable. There are very few things in life where you can actually reduce them to axioms and prove it. So you can do that in geometry, but for example, if you want to decide like how we should handle the Taiwan war if it happens, like there's no formal verification that you can do there. And so in these more open-ended domains, we're still kind of stuck where we were. You know, where we were in AI in 1985 or something like that was you could do a few things well and then you'd have graduate students like fix all of the holes in the systems because they weren't really that great. They were actually kind of brittle. And now you have like Kenyans offshore or something like that fixing all the holes in the system because they're not really that great. I mean, like these companies have, I think, thousands of outside contractors trying to basically fix the errors in the systems, and you can't really know on any given day if you type something in if it's going to work. So aside from these math cases, and I see Jim wants to ask something, you generally don't know if it's going to work or not. Like you type something in and you might get an amazing answer. You're like, I can't believe... I mean, even I, who's kind of a skeptic, be like, it's amazing that it got this right. And then sometimes it's stupid. Like the only thing I'm really using them for now is proofreading. And sometimes it's like, 'Wow, it really... that's a really good point.' And sometimes it'll like tell me to fix something that isn't even there or, you know, it will just kind of repeat itself. And I'm like, 'You just told me that.' Like...
Do you think it knows it's you though and just wants to f*** with you a little bit or no? I mean, like maybe that... like, 'Gary Marcus, I'm dealing with Gary Marcus. You're trying to kill it.'
Most of the time it actually sucks up to me in the things that I've seen.
Gary, is it fair to say, for the lay person like me, that the difference is widest between hard sciences and social sciences? When you try to apply it to human behavior, it's not all there, whereas if there's a world of hard proofs, physics laws, you know, the laws of mathematical proofs...
I mean, there's lots of things I would count as hard sciences it's still not that great at, you know, understanding physiology or something like that where we just don't have the cut-and-dried fact. So it's like the hard edge of hard sciences that it's most appropriate to, where we really can map everything out. I don't think there's been any great discovery even in physics. You know, the closest is probably this thing about Navier-Stokes, but it turns out that the proof that they did doesn't apply to the real world at all. It's... that's almost like a perfect case of it. Like they did the formally mathematically provable piece of Navier-Stokes, which was legit because the Clay Prize was about it, but it doesn't really help physicists actually solve this in the real world. It's like a very special case that's mathematically defined. So it's like even in physics, it's only the, you know, very narrow special math cases and not real stuff.
You know, Jim, you just called yourself a lay person as it relates to technology. You're certainly not as it relates to financial analysis and investing in the like. And so everything that Gary just said about the models, and it seemed like, you know, it started as an academic thing and then, you know, there was a lot of folks who just really wanted to kind of reach these milestones, whatever they were. I always thought, and I'm also a lay person when it comes to technology, I was like, well, if all these labs are going for AGI, you know, they're not just going to lay down if, you know, one of them says, 'Hey, we reached AGI.' They're like, 'Well, what's your definition of AGI because we're doing that over here.' Like, how important was it? You came to this, as you just kind of detailed, about data centers in general. It didn't matter that it was, you know, hosting SaaS or this and that, whatever. What was like the impact like going back to the release of GPT? This is late '22. You know, a lot of folks, and this is where, you know, maybe Gary, I don't know if you agree with this or not, a lot of folks like focused on the public markets really weren't paying that close attention to it. You know what I mean? Like there was some focus, ah, what does this mean for Google's search business and this and that, whatever. If you're getting contextual answers on searches, that sort of thing, but the economics weren't a focus yet. And I'm just curious, at what point did you start focusing on the model... we started focusing on the AI buildout as opposed to legacy data center business?
Late '25. So last year, because the numbers were beginning to accelerate so dramatically. We started work on the hyperscalers and looking at returns and then more importantly returns on incremental invested capital, and those peaked in 2024 and have been declining ever since. And at a rapid rate, by the way. If it continues at the current curve of decline, the hyperscalers, who are the best positioned to invest in these various initiatives, will have returns below their weighted average cost of capital sometime in mid-'27, which I don't think the market is prepared for. But the numbers became so big relative to the financial markets and then the dominance of the AI-related narratives in terms of their size in the marketplace. They're now, depending on how you define it, 45 to 55% of the S&P is somehow AI-related. So you can't ignore it as a market participant. And as someone who's long the market as a hedge versus our shorts, which are idiosyncratic, you know, I have to pay attention and I have to understand that I'm long a lot of this stuff just by definition, as is every investor who's invested in index funds. It is now the economic tail wagging the technological tail wagging the economic dog, certainly a financial dog. And so I think that you ignore it at your own peril because it has gotten so big so fast.
You guys are kind of two sides of the same AI coin in many ways. If anyone, any listeners here are on, you know, Twitter or Substack or whatever, you know, you guys are not shy about like offering your views. And it's something like, like you just said, if 45% of the S&P 500 is somehow exposed to this, you're going to get a lot of hate if you're trying to pick it apart from a technological standpoint or an economic standpoint or as I think both of you guys are starting to do in the middle here. And so I'm just curious, like, you know, you're in the minority because so many people are getting rich, whether it's in the private markets, right, or if it's in the public markets because we know it's been driving a lot of that performance. So Gary, I'll just start with you, I guess. You know, how do you think of that being like the kind of pied piper in a way? I know Jim, you've gotten pretty used to that and what comes with that. Is that a little stressful? Like what does it mean? Like I don't know if you're married. What does your wife think about that? I've been doing TV for 17 years and, you know, Sarah, she's not watched one minute because she thinks everything that comes out of my mouth, not that she's an LLM and can figure out what the next word is, but she gets nervous about it.
I don't think my wife is particularly nervous about it. I have kind of faced hostility since I was a kid in graduate school. I started grad school when I was 19 and my first paper was with Steve Pinker and we submitted it around the time that I was able to legally drink, which is 21 in this country. And we got a, I think it was a 30-page single-spaced review basically, you know, accusing us of making stuff up or whatever. The guy who wrote the review later apologized and, you know, it turned out that we were attacking the work of his graduate student. He was manically depressive and one of his typewriters was broken and he excused himself in this review. And I had to make a decision at that point. This was my first paper submitted in graduate school. Was I going to stand up for what I thought was right or was I going to walk away? And, you know, coming from where I came from, I stood up and didn't walk away. And, you know, nothing after that really fazed me. So, like, there's intellectual combat on Twitter every day, but, you know, it's just part of doing business and I try to make teachable moments of it and I don't let it bother me.
Yeah. I mean, my many fans always seem to see their heads explode when I actually try to patiently explain to them that I'm probably net long the AI trade because I am. Our shorts do not equal the weightings in the indices. But teachable moments is exactly the point. Yesterday, people were sitting around just how cheaply the neoclouds are financing themselves and they were putting up all these charts of the interest rates, blended interest rates at the various different neoclouds, and 2%, 4%, 8%. And of course these neoclouds are financing themselves mostly with convertible debt where the coupons are 0% or 2%. So yes, on the surface they're paying very small amounts of interest, but that's not how you look at it. The cost of the capital is actually really high when you back out the implied option value of the convertible bond. And these companies know this, but what bothers me a lot is that we've now gotten to four years into the cycle where we are beginning to really throw a lot of capital at charlatans and promoters and people who are really trying to make a fast buck on retail investors' infatuation with the concept of AI growth. And you can have AI growth. It doesn't have to be profitable, as we know from the internet cycle. And that's where I think the teachable moments lie. And while things are going up, no one wants to hear that. And I get that. I've dealt with that for 40 years.
All right. So Jim, you just used the term charlatan. So who are these charlatans? Okay. Because yeah, I mean, I just said like not everything's going up anymore, but I would assume...
I don't want to name any charlatans, but can I just give an example?
Without naming names, there's a certain person who just said that by the end of 2028, we will have artificial general intelligence that can do anything a person did. And that same person said the same thing in 2024. And I offered him a million-dollar bet, which he did not take. And at that time, he said it would be by the end of 2025. So he just moved his prediction back three years, made no note of it. His fans on the internet got mad when I pointed out that he had said this before, and they thought I made it up. I'm like, here's the receipt. Here's the million-dollar bet and the quote that he said. But it is a charlatan thing to do, to keep moving back your predictions, you know, every year.
Where did confidence come from on your part? Because again, these guys think they're the masters of the universe and they now are deploying, you know, tens, hundreds of billions of dollars to kind of put behind, you know, you talk about receipts. Yeah, it was out there. But these folks, like, they really are believing, don't they, for the most part, believe they're right now because they're able to raise so much money and they're basically putting this magic out in the world.
It's very hard to tell what people actually believe versus what they say. One of the most important things in this whole industry is that VCs get paid a commission on the money that they invest. And so as long as they can tell the story with a straight face, they make a lot of money. You get 2% on a million-dollar investment, that's nothing. But you get 2% on a billion-dollar investment, you don't even care that much if it comes through in the end because you're doing pretty well straight up front.
So, well, part of that model is that like if what, 7% of their investments hit, you know what I mean? You can have 90% go to zero and you still...
have a fund when you're making 2% on, you know, handling 12 billion. You can have none of them do well. You could be, you know...
Well, you'll be out of business. You won't be able to raise another fund. You have to have...
But you have made so much money already.
Well, I'll just tell you that right now in the VC space, if you had not invested, let's call it in Anthropic, OpenAI, SpaceX, or xAI also, you know what I mean, and Anduril and a few others, you can't raise money right now. I mean, it's like really hard to do that. Now, there's some big outcomes. There's OpenRouter, which I think was founded, you know, just a couple years ago, and it gets taken out, I think, by Stripe for 13 billion dollars or something like that. That's a big return. People are not paying attention to some of that stuff. So there is some M&A going on, and, you know, Hugging Face, you know, that was 25 years ago too.
Yeah. All right. So, like, talk to me about the charlatans because you're more focused in the public markets to some degree and I think those are the ones that folks are like, 'Oh, Jensen's been around for 25 years.' You know what I mean? Like, how he's not a charlatan. That sort of thing. So, like, I don't know. You want to name some names here?
Well, I just say, just as a group, take a look at the guys who told you and raised billions of dollars, told you that raising money to buy servers to mine Bitcoin was going to be the next big thing. And now they've all transitioned seamlessly into becoming AI data center companies. There's stock promoters, as I said, tend to follow the financial cycle with a lag, and we're now seeing lots and lots of people try to do anything to affix AI to their business and get a halo effect that started, you know, a year and a half ago, two years ago. But I think that Gary's right in terms of the VC incentives. You're not going to hear a negative word because it's about raising capital and putting it to work. But the financial markets will do their job at some point. And everyone also forgets there's a reflexivity function in financial markets that then reverberates down through technology and whatever. And if you can't finance it, suddenly things change.
It seems to me like the bond market has kind of caught on and the, you know, the prices of loaning money to the data centers is going up, whereas the public market or the stock markets, I guess would be the word, have not caught on. And there seems to be, I don't know if it's an arbitrage opportunity, but just a decoupling between the two. Can you...
Yeah, which is why I mentioned convertible debt a little earlier. It fits with your question in that the equity markets are still very sanguine as it relates to a lot of this stuff. Slowly but surely, last three to four or five months, the credit markets have begun to tighten. And that's because if you lend money to someone, you're only getting your money back plus interest, right? There's no optionality on the upside. And so increasingly now the debt component costs are going up and you're going to see more and more equity, I suspect, equity-like structures like converts or pure equity. Because increasingly these companies don't want to show large interest expense and because they're not covering it. And so with the exception of the hyperscalers, pretty much everybody else is losing money on a, you know, real economic basis. And so I'm very anxious to see the Anthropic and OpenAI S-1s because I want to see how they define their profitability. But it's not happening in the data center business. It's dropping rapidly at the hyperscalers and they have their core businesses to fall back on, of course. But as I mentioned, next year should be a very interesting moment of truth as it relates to their incremental returns. And yet Goldman Sachs today raised their estimate of total spending on data centers by 50 gigawatts, which is two and a half to three trillion dollars over the next five years. Billion, that was just the increase. So they're saying, they're saying over the next five years the total spend will be 10 to 12 trillion, which is about 6% of GDP, which is, depending on how you want to define it, 1x to 2x the US personal savings rate. Right, so just the personal savings rate is going to go into AI data centers for the next five years, according to the street, before we get to funding the government deficit, everything else that the economy needs. So you're starting to get really silly numbers.
Yeah. Well, let's talk about this because this was out today and unconfirmed, but this was Oracle citing force majeure to shield itself.
Very common. It's very common, Dan. They came out later and said force majeure is a very commonplace...
Well, well, so you'll explain it, but I just want to kind of give you the background because you're talking about the credit markets, right? And by the way, the equity markets have been punishing Oracle, right? Going back a year or so, it had that huge, huge... it was up 35% on one day on announcement.
I wrote an essay that day called 'Peak Bubble'.
I said they might go up, but we will remember this as the peak moment of insanity that they went up this much on this.
Well, by the way, I was actually at the Robin Hood summit that day. I was on Fast Money. The Oracle report had just come out and we were saying similar in the moment because, you know, we'll go to the RPOs and the LTAs and all this sort of stuff, you know, the backlogs or whatever. And, you know, it was really interesting. And this is where I'm going to get to the Oracle thing also. You remember, like, days after the inauguration, we had Masa, charlatan maybe, is that in that camp? We had Masa, Sam Altman, and we had Larry Ellison in the Oval Office and they announced Stargate. And at the time, you know, it was a half a trillion dollars, okay. And do you remember what your boy Elon tweeted out almost immediately? No, my boy. He said, I was kidding. This...
That was one of his better calls. He said, 'Where's the money?'
He said, 'They don't have the money.' Okay, so let's flash forward to this right now.
Yeah, force majeure is act of God. So...
Yeah, I know what the meaning is, but what did they... they're claiming unforeseen delays. I don't know if it's a Iran war or what they're... they're... on one of the data centers they're...
So this is in New Mexico. It's Project Jupiter. It's part of Stargate. But here's the interesting thing, you know, Blue Owl, right? They own the developer, okay? And you know, they put equity in there. Um, you know, the banks put up the credit. So, it's an end-run-like structure. It's an SPV.
So, instead of 'my dog ate my homework,' it's 'my god ate my homework.'
Yeah. Well, all right. But so we're seeing, so there's an $18 billion loan on this and it's Goldman and it's BNP and it's SMBC and all the rest of it, and all these guys, and let's just call them what they are, you know, they're second-tier, third-tier banks. They all want in on this sort of stuff, right? So if Oracle is saying for whatever reason there are delays, right? And listen, there's a whole host of stuff. There's regulatory, there's NIMBY, there's moratoriums in the data centers. I mean, power, you know, access to power, that sort of thing.
Political. I mean, it's not just the populace, it's both parties. Now, I'm in a swing county in Pennsylvania, Bucks County, and so we get inundated with ads, and both Republican and Democratic candidates are both claiming the other side is being soft on data centers.
Yeah. All right. So, but let's just break this down for one sec. Okay. So, to your point about peak bubble going back a year, you know, Oracle is down 60% from those levels. All right. The stock on this headline is down 7% today. The CDS on Oracle debt is at new highs. I think it's trading at like 225 or something like that. So Jim, you know, Blue Owl saying, and this is the second time they've had to do this. There was a data center, I think, up in Wisconsin, wasn't it? Okay. That basically, you know, Blue Owl I think walked away from...
Yeah. And so Oracle is now, they're claiming force majeure. This seems like a mess. Um, so credit markets are paying attention, or those who are looking to insure credit. The stock market is paying attention. How can this end? Well, because if Oracle is supposedly saying this, not confirming it, it seems like this might be something really important to bookmark today.
We'll, as always, we'll see. Um, but so much of the financial economy right now is dependent upon this buildout. Uh, it had better not be the important thing that you think it might be because it's not just the data center guys. Not just the model guys, it's Caterpillar, it's GE Vernova, it's all kinds of labor and construction on the ground, private companies, it's HVAC companies. I mean, this is a massive stimulus program. And um, it's going to be a massive stimulus program unless we begin to see the capital markets shut it down. Um, and we'll see, we'll see if the debt markets... the debt markets have weakened. They're still not prohibitive. I should add that to our discussion. I mean, CDS's of 225 means they're raising money at 8%, you know, whatever it is. Um, that it's not prohibitive yet. Um, if it gets into double digits, debt only, then it's prohibitive.
Yeah. But Oracle to fulfill these obligations are going to continue to keep raising money. They said they're not going to... Larry Ellison canceled a seven-and-a-half billion dollar stock sale last week or two weeks ago. And you know, the irony is, is like he was probably raising that company capital to put into the Warner Brothers deal for his son. Right. So they're...
Well, he also had layoffs that weekend, which made it a little bit...
Right, but there's... and they're laying off tens of thousands of people. Right. So, so Gary, when you... like you asked the question, it's not like this is stuff that's not on the tip of your tongue. I mean, when you think about, okay, what these companies, the model, you know, the Frontier Labs, what they're spending, how much they've raised, how badly they need to go public, right, to continue or just to kind of raise capital, what does that make you think about? Like you're focused on the pace of innovation and what these models actually do. Is there going to be a moment where the rubber hits the road where ChatGPT 5 comes out not that good, you know what I mean? Whatever the next model is, I keep hearing about the next Gemini. Do those announcements matter? Because if you're spending tens or hundreds of billions of dollars to create an infrastructure to train these models and the next, your competitor comes out, it's just a little better than the last one. Who gives a... Because, you know, every week there's going to be a new announcement. I mean, this is part of the insanity of it all is that nobody has a lasting lead here, nor will they. And now China is in the ring. So, like you just alluded to, Google has a new model coming out. Maybe it'll be a little bit better, but nobody keeps a lead for more than two weeks. That's why we have this commodity structure. It's great for consumers because the models get better and it's great for consumers because they have to compete on price because nobody can really say my model is, you know, substantively better.
And there's no frictional costs, right? I mean, to change, there's a little bit, but less and less, and things like OpenRouter are designed to reduce that friction. Um, and so yeah, there's basically no friction at this point, or converging on friction. We're basically converging where it's not clear there's going to be any profit for the model providers at all.
All right, let's talk... because you're focused on that. So, who has the monopolies or... maybe Nvidia?
Nobody does, and that's part of... Well, Nvidia does. They're right. Nvidia's making shovels in a gold rush. They're in a wonderful position. Although the degree to which Jensen keeps kind of trying to hype the stuff suggests to me that maybe he's concerned, right? I mean, we were talking before we went on about how much money he's spending, you know, from the money that he's making.
It's now almost a dollar for dollar. So, it's not quite. Last quarter, I think we calculated it was 70 cents or something close to that on the... of every dollar he makes, every day, he has been going back out the door in investments.
And another, you know, big number, 80% of his time... I mean, I just made up that number, but a large fraction of his time is spent trying to pump the market for...
Is that the way you see it? Because bulls, and if you're long semis and you're long Nvidia and you're long this trade, they just see it investing. They think it's a brilliant way to use this cash flow that they've been generating.
There's an argument for that and there's an argument it's going to run out because let's talk about the rest of it. So I wanted to bracket, you know, Nvidia out. They're in a different position. They're selling shovels in the gold rush. They're doing a lot to try to get people to buy more stuff in the gold rush. You could say if the organic demand was there, would they need to do this much hyping of it and so forth? So you can wonder about Nvidia separately. But then you have the model providers and then you have the end users. The end users, every time somebody does a study, it comes out that there's not that much return on investment. Right? Most people are doing this because they feel like they must do it because they'll get behind on their competitors. But when they actually do it, it's rare that they find, you know, any real return on productivity. Probably coders, it's real, but many domains it's not.
I mentioned to Dan before we started that one of the more empirical places we can check on this sooner rather than later is my world, the investment world, in that some of the biggest end users of these products are trading firms and financial funds. And it'll be interesting to see if anyone does the work to see if the funds that are spending the most per dollar of AUM on AI are outperforming those that aren't.
I think that's an interesting test. Here's another one. Um, is, you know, coding is a place where it helps the most. And if you look at the App Store, we have like, whatever, I don't know the exact numbers, like five times more apps coming out per month or something like that than we did before. But the actual sales are not really up that much. Like most people are making new apps that nobody actually wants. And so it's productivity in the sense that there are more apps. Like you can count that as a measure of productivity, but it's not productivity in terms of anything is all that great. You know, all of the great things that we have aside from, you know, whatever you may think of chatbots, were invented a long time ago and built a long time ago. Web browsers, word processor. It's not like we have, you know, harnessed this and transformed anything really. The only thing that's been transformed, I would say, is some aspects of formal math. But the rest of the world is not, you know, we're not having better books. We're not having better movies, better software.
I think we discussed this on the last podcast, but Gary might not have heard this and be interested in it. We took a look at the world a decade before Netscape and a decade post Netscape. So, basically '86 to '96 and '96 to '06. And we looked at US GDP growth. We looked at total factor productivity and we looked at S&P 500 earnings growth. And there was slightly greater US GDP growth, but it was a rounding error. It was 3% after Netscape in the decade. It was 3% in the decade prior to Netscape per year. Real GDP growth in the US. Total factor productivity was slightly better. S&P 500 earnings were actually slightly lower post-internet. They were 5% per year versus 6% per year, which is, by the way, the long-term curve is 6% per year in S&P earnings growth. Um, so the internet came and went and was highly disruptive, brought forth lots of new businesses, destroyed a lot of businesses, but in the grand scheme of the overall economy, you would not have been able to say, 'Oh, the advent of the browser and the public's acceptance of the internet changed the world.' It'll be very interesting to see... GDP growth's been very mediocre the last three, four years. Now again, we're early on, but um, you would not be able to say from late '22 the ChatGPT moment changed economic growth.
And in early 2023, that's what everybody, not me, but everybody else seemed to think was, 'Oh my god, this...' You know, people talked about like 10x programmers or 100x this, you know, single-employee billion-dollar companies, the only one of which I know verges on fraud. Um, but I said that carefully, I guess. Um, you know, people had all these fantasies that like the whole world was going to change. And then, yeah, you look at these productivity numbers, it's like, yeah, did something happen?
Here's the counter to me. What is interesting is that OpenAI has more than a billion users. You know, Gemini is going to have about a billion users. The adoption in such a short period of time. But again, I'm just taking the other side of it, is like there's not a person that I know in my life, either personally or professionally, who's not using an LLM. You know, and many in our cohort are paying $20 a month for it. But we've seen with Anthropic, and this speaks to, Gary, what you were just saying, is where are the kind of most evident sort of products or productivity? It is coding, right? And so, you know, I think about it, I've been able to create a lot of cool little things. I'm using this Instinct. I didn't create it, but I don't even need to vibe code. This is something I get for free. It's a personal assistant, you know. Um, we'll talk about Muse in a second. I just think there are a lot of... I think there's a lot of really useful things that are created that I don't actually have to pay for right now, and as a consumer, that's great if I have...
Um, it depends. I mean, we'll talk about tokens and we've been talking about the cost of compute a lot. I mean, if the prices keep coming down and ultimately, let's say, the models become more useful, that's a great scenario if you're a consumer of these products, you know. And so I don't know if what you just laid out is maybe a great apples-to-apples, is all I'm saying, because the pace in which things are being created in and around... I see my kids, they're 21 and 23, and how they're using this stuff.
Shouldn't that argue for a greater impact now? Yeah, I mean, there's a lot of like... I mean, we're not seeing...
Well, I mean, like maybe we're just in the early stage. Maybe it is 1999 and we haven't had our Amazon moment or we haven't had our eBay moment or, you know, I mean, the list goes on and on. I mean, Google was born out of that period. It was a pretty good business model. I'm just kind of taking the other side a little bit here.
So, my view is that AI is eventually going to be absolutely transformational, but it's not clear that generative AI is.
Would you explain the two?
So, generative AI is stuff like chatbots, right? And that's one way to build AI. You could argue it's the best one that we have right now. You could argue that the best way of doing that is actually this hybrid that I was talking about before with the generative AI plus these harnesses and whatever. Um, but even with the harnesses, there's still reliability problems. That has been the core problem all along. So take vibe coding, right? People love vibe coding, at least when they start. It gives them this moment of feeling absolute power, which those of us who did traditional coding got in the old-fashioned way by you directly coding, but not everybody gets that or, you know, had the nerdy tendency to do it or whatever it is. You know, so, you know, when I first learned to code, I thought it was the most amazing thing in the world, and I still do in some ways, and vibe coders get that. But vibe coders don't typically have the training to know how to fix the things that they build, and so a lot of the things that they build don't endure. They don't really know how to make them last. They make a demo of, I took a picture inside my refrigerator and it told me what ingredients there were and gave me a recipe, and it feels great, but eventually, maybe if they're smarter, they realize like, oh, some of the ingredients are behind others and this is more complicated than I thought. And um, you know, sometimes the lighting... well, maybe it's not a perfect example, but you know, things occlude one another and the things don't work as well and they wind up with this mess. I have a friend who's writing a piece. I don't know if it's out yet, but he visited some people who were doing these vibe coding projects and like while he was visiting them, like they would work and then they would stop working, then they would work again, and they had like no idea, you know, whether... and there's this, you know, sort of magic alternating with terror. And some people will, you know, it's like a rodeo horse and they'll try to stay on, and some people will get off. And so like a question even for the coding is how much of this will endure. For the professionals, it is a genuinely useful tool. For the people who already have the training and know how to debug things, well, they can do that. They already know, you know, the system will make errors and they'll be able to deal with it. Although there's some questions about them losing their skills and not paying enough attention and so forth. But I think a lot of the vibe coding is not going to work out as well as some people think. I think it's going to be a fad. I mean, remember, you know, before maybe the time of our listeners, we had things like pet rocks that were really popular. Some things endure and some things don't.
We had CryptoKitties five years ago. We had CryptoKitties because you were all over that too, by the way.
And so like it's hard sitting where we're standing to know which of these things are going to be like that in the long term. Absolutely. And I'll talk briefly about agents and we might want to come back to them because of how they're playing out and a whole bunch of other things. But agents are phenomenally useful in principle, but if they make mistakes, that can be pretty serious. So, if you tell an agent to, I don't know, place transactions for you and it misplaces the decimal place, that's pretty bad. Or if it...
Does that also... my wife...
I'm just telling you. I'm just, you know, we end up getting literally, we get flights going to a totally different city every once in a while.
Well, the famous one is Auckland versus Oakland.
He knows my wife and you're just looking at me like this. She doesn't listen, Jim. Okay. So, like she's never going to hear it.
Sorry. So, the thing with agents is sometimes they work and sometimes they don't. And so people, you know, maybe they decide that this is, you know, like right now Muse is super popular and there's something I want to throw in about Muse, but you know, the reliability may not be there and it might be popular for three weeks and go away. We don't really know yet. You know, in the long run, we will all have these agents and they will be great, but the technology doesn't really support it right now in a way that's good enough to trust. And, you know, trust is important.
Well, all right. So, Jim, this is interesting. Everything you just said, Gary, you know, Instinct, it caught the tech world by fire. This is a personal assistant. I've been using it for a month. I haven't actually seen... you know, you can coach it a little bit like any sort of thing. It's been very useful. I'm just telling you that right now. I really...
You sent me to the wrong place this morning, first.
Oh, stop it. That was Amanda, by the way. Um, sorry. All women, too. I'm going to get in a lot of trouble.
Well, here I'm going to make my point about Muse right there. Actually, it's her. Well, I'm going to precede it by saying there was actually a thing before that, that Facebook, before they were called Meta, released called Facebook M. Got all kinds of headlines. Everybody was excited. Earlier this week, I pulled together the clippings of people saying how great it was. Um, it turned out there were humans behind the scene. The reason that they never scaled it out, they only had like 10,000 users, was because their fantasy was, 'We'll just collect data from 10,000 humans. We'll scale it out, we'll start to make money.' And after four or five years, they realized that was not working. Now the technology is better now than it was. And maybe it'll work this time. But I discovered from what was really Reuters' work that they have humans there now. There are still humans behind the scene.
So it's still not existent. I, you know, I was just going to say one thing and then Jim, I'd love to get your take on Muse because, you know, Meta stock has gained literally $400 billion in market cap off of its lows. The stock got destroyed after their Q2 earnings. They came out with their capex number. The stock was trading at multi-year lows. They put out a personal agent and the stock goes straight up, gaining again hundreds of billions of dollars. What's interesting to me about this moment...
Speaks well for the investors, like they have no, you know, long...
And that's your... and you're... he's going to get all up in there. But two things: Instinct needs to go out and raise money because they are compute-constrained. That doesn't surprise any of us, right? Quick adoption of this thing, it's pretty good. I'll bet you someone buys them, okay? I'll bet you like Microsoft or somebody buys them as they're out there trying to raise money, okay? Muse, people, investors were really disappointed about how much capital they're raising. They already squeezed out a lot of productivity in their ad business, you know what I mean? And but now they're like, well, they don't have a cloud business, GCP, AWS, or Azure, right? And then they come out with a personal agent and it's the fastest-growing app that's ever existed, right? And then the market or investors reward them. So help me think about that because all...
So the two largest returns on incremental invested capital in the hyperscaler world are Microsoft and Meta. Um, and both of them have end markets, right? Microsoft has the enterprise, Meta has the consumer. Both of them have the best economics right now. Looking backwards, right now in the hyperscaler space, and you can imagine which ones have the worst. There is something to be said for building upon a customer base and having it turn profitable or at least not be a huge cost sink, um, as the others have. So it does help. I mean, it does help to have end markets, and those two of the hyperscalers are probably to us the most attractive of the hyperscalers, um, for that reason.
What about Google? Because they're going to come out with their Muse and it's connected to Chrome, Gmail, Calendar. I mean, the list goes on and on. That will likely be one of the most useful agents if you think about across their productivity suite. I won't use... I'm not... I don't have Reels or Instagram or Threads or any of that stuff. I won't use Zuck's products. I don't trust them. I don't trust Altman. I don't use their crap. I kind of trust Sundar. I, you know, I have... a lot of billions of people have Gmail and all that sort of stuff. So to me, and then if you're an Android user, they have the ultimate distribution. Um, so I don't know, that one looks interesting to me. I'm just saying ultimately, and your point about Microsoft is Copilot adoption right now is minuscule. It's like, you know, less than 10% of their, you know, Office 365 users, and it is growing. But let me ask you this then. Okay, so Microsoft, $13 billion going back three, three and a half years, right, into OpenAI. They're in the catbird seat with that technology. They spent on the infrastructure to build out Azure, but they didn't spend tens or hundreds of billions of dollars building models. Like, is that something that looks interesting to you ultimately? Is that kind of what you're saying also?
