The Most Dangerous Investment Cycle In History | Paul Kedrosky

Hey, thanks for having me. It's the first appearance on this podcast. Unfortunately, it's not the first appearance on any podcast, but yeah.

So, I come at it from both the finance and technology side, I suppose. I mean, my day job, one of my day jobs is I run a series of venture funds with a partner of mine that we've been doing for almost 15 years. So, you know, so I come at it from a sort of financial technology side. But way back, way back, I was actually an equity research analyst at one of the major brokerage firms. And so, where I was coming at this largely from a financial standpoint, by which I mean, you know, financial systems and how they function and how companies do what they do. And so, I come at it from sort of a synthesis of those two perspectives, I guess. And way, way, way back, um, in grad school, my PhD work, my grad school work was in sort of complexity and path dependency about how systems get into fragile states and what the consequences are. So, you know, so in a sense, in a very loose sense, it's all kind of linked together. At least it is for me in my brain. And so, you know, that's how it all fits.

In terms of independent research and so on, you know, I was a university professor way back when at the University of British Columbia, University of California, San Diego. I still sit in as a fellow at MIT at the Institute for the Digital Economy. I'm also a fellow at the Mansfield Oxford Institute in Oxford. And so, I still kind of maintain affiliations with the universities, but I just found it was more interesting for me to do stuff semi-independently and just maintain affiliations with places that have interesting people doing interesting things. So, I suppose I've been doing that for the better part of 15 years plus in various forms.

Yeah, I use that as a vehicle for a lot of my stuff just because, you know, what's the old line from The Dark Knight? If you're good at something, never do it for free. I think is the line the Joker uses. So, don't do it for free. So, I do a lot of stuff like this where I like to get out and talk to people, but I also have a number of institutional clients that I spend a lot of time with, but a small number just 'cause it eats your calendar if you're not careful. And mostly hedge funds and large institutional investors.

I was indeed. Yeah, yeah, yeah. So, that was my scar tissue. And I think, you know, it's one of the ways I approach all of this stuff, not just in terms of systems, but thinking almost in terms of demographics. And we often repeat cycles in capital markets as people age out, and so a new cohort comes along who doesn't have any scar tissue from, you know, pick your favorite financial meltdown. It could have been the telecom bubble, it could have been the global financial crisis, it could have been the dot-com period, whatever. And so, you know, having had that exposure, I think was for me anyways was incredibly useful. I was... a friend of mine who's one of the largest short sellers in the world was recently at a conference and he put his hand up and asked... There was like 800 people in the room and he asked what percentage of you... These were all fellow traders and hedge fund managers were active during the dot-com bubble, and it was less than 30%, which I think is really important because that kind of... you know, we're all human, right? So, that kind of cognitive and emotional scarring in terms of creating a, you know, receptivity to noticing things that look more fragile and dangerous, it helps if you've been through it. And the reverse is also true that if you're completely immune to it and you don't care because you've never experienced it before, you often have an incredible run of luck in the markets because you have no sense like, you know what? I don't know what's going to go wrong. Nothing's ever gone wrong for me in my career. And well, how long's your career? 8 years? Oh, okay. Well, I get it then. So, it can work both ways.

So, for me, it's a general-purpose technology, a GPT, so which is to say, its best analogies are things like the canals back in the early 19th century and 18th century. It's like the railroads in the late 19th, electrification in the early 20th, um, the interstate highway system in the mid part of the 20th century in the United States. These are all general-purpose technologies, which is a fancy way of saying that it's a general framework on top of which other things happen. And so, for me, this is the way I think about it, and so that leads you to a bunch of metaphors in terms of thinking about it like a utility, about its ubiquity, and about it as an underlying substrate on top of which a lot of things happen. So, much like, as I said, much like railroads, electricity, the interstate highway system, even fiber in the late 20th century, these are all examples. So, that's the broad metaphor. It's unfortunate that, you know, ChatGPT stole the GPT part and kind of repurposed it, but nevertheless, GPT has a much longer history than ChatGPT. So, it's been something of a term of art within the industrial economics and economic research community of technology, which is this idea of a substrate on top of which things happen. So, AI is a GPT.

Yeah, so broadly, and I think that's a really important distinction to make, is that as a GPT, it's very unusual because while it's a substrate, it's much more of a substitute substrate than necessarily a complementary substrate. So, think about it like... let's take the example of the railroads. The railroads were really lethal for, say, horse-drawn carriages, as were highways, in the sense that they created a really effective competitor, but it didn't really change much in terms of humans. There were, at the margin, there were people who changed occupations, there wasn't as much horse grooming going on, and so on. But they were much more complementary in terms of introducing new ways of doing things, potentially increasing productivity. The difference with AI is that it's an incredibly broadly applicable technology that for the most part, at least in its current incarnation, is really focused on symbolic work, what we sometimes call white-collar work. So, and in those domains, it's at least as substitutive as it is additive in the sense that it's very straightforward to see how it can replace tasks and occupations. And I, you know, I consciously make that split as well as making people potentially more productive. So, in that sense, it's much more consequential in terms of its likely impact on humans as opposed to saying, you know, I don't know, the emergence of railroads makes it easier to move boxes around the country, but it didn't necessarily change that much in terms of individuals at scale. So, the two things that are different: it's much more substitutive, and it's happening much faster. So, the two together make it much more consequential in terms of the likely disruptive force that it has in the economy, and I think while that's better understood than it once was, I still think that's part of the problem in terms of people... People have a bunch of rote mental models in terms of how they think about GPTs. You'll hear lines of pattern like 'technology always creates more jobs than it destroys.' And this is based on a very small sample set of prior GPTs, and most of which had the characteristic of being very much complementary, not substituting. And so, yeah, that was sort of true, but it was true in a very small number of cases 'cause we only have say N of 5 as an example, maybe, of prior, you know, wide-scale GPTs in a Western economy. So, you have to be really careful about applying these blanket rules, like it always does this or never does that, because we just don't... We don't know. That's just not a... You can't say that from the data we have. And in particular, you can't say it given that most of the analysis was ex post. We know what happened afterwards, but we don't know what would have happened... If every time we had a prior bubble, people said don't worry about it, invest as much as you want because they always create more jobs than they destroy. That turns the thing around, right? Suddenly, you take something descriptive and make it prescriptive. You're telling people, here's what to do as opposed to here's what happened in the past. So, that's for me one of the key distinctions is that it's much more of a substituting technology than a complementary one, and we see that in the early data. There's new data on call centers, for example, that is really compelling here. Lots of other things out there that you can point to. Even Anthropic had some new data recently, a report last week talking about it changing the nature of really sort of entry-level hiring in terms of them doing less entry-level hiring. And you're seeing exactly what you would expect for a symbolic substituting technology being introduced at scale.

Oh, absolutely. I think that's baked in here. And it's probably even more so because of the speed at which it happens. So you can separate the two aspects of it. One, that it's kind of a heat-seeking missile aimed directly at white-collar work, and also because of the scale and speed at which it's happening, that's unprecedented. You know, you can think back about the railroads which took like 25, 30 years to really reach full deployment, or think back about the industrial revolution which was 30 or 40 years in terms of, you know, the true scale of what happened, or even the fiber moment was like five or six. All of these were much longer. This is happening with a much, much larger amount of spending and at a much higher speed. So we're not well-equipped to deal with that, in particular whenever we keep applying this old metaphor of like, you know, sort of the... I don't know how familiar you are with... there's a number of observers of this stuff. Like one of the people that people like to point to is Carlota Perez who famously wrote this book Technology Revolutions and Financial Capital which sort of describes how some of these moments happen and how we go through these wrenching changes and then there's sort of this element of moment of disruption and then there's losses and write-downs and then on the other side there's a new sort of golden age of technology. And the problem with these kinds of ways of thinking about what's happening and applying it to AI is again, that's descriptive and ex post. It's not that we went into the telecom bubble saying it doesn't matter how much you spend because it always works out, or that we went into the railroads build out and saying it doesn't matter how many route lines of railroad we build because it always works out. This is just something, a conclusion we've come to after the fact, and so that's one of the reasons why this moment is more dangerous is because we keep applying the wrong metaphor, these sorts of descriptive metaphors as opposed to a sort of a prescriptive one and saying, 'Okay, how is this going to play out?' And the other reason why this one's really unusual, and this is more of an economic history perspective, is that most of the largest financial crises in Western economic history have had one of four factors driving them. So, it's either been about loose credit, the easy availability of money, that tends to drive things. The GFC, the global financial crisis was in part about that. Real estate has been at the center of many crises in the US and elsewhere, including the GFC. So, speculation about real estate. Technology. There's four of these, by the way, just in case you think there's 100. Technology is often at the center of this. A really good technology story can cause these moments, these crises. And then the other one is government policy, obviously. And government policy where, you know, the government's encouraging people to do things and creating an implicit backstop. So, generally speaking, one or two of these are involved in almost every crisis in Western economic history. This is the first one that has all four. We have a great technology story. We have a real estate angle. Data centers are implicitly commercial real estate and a lot of speculation around that. We have very loose credit, which is changing a little bit in the last 2 weeks, 3 weeks, as we kind of have this, you know, debt market freakout over CoreWeave and others. And we have a government angle, right? We've been told repeatedly, at least in the United States, that we're in an existential battle with China that we must win. And it doesn't matter the costs or the consequences. So, the notion that for the first time one of these episodes, and I can sort of define what I mean by it, but one of these episodes sits at the intersection of all four forces is why this one in particular is both dangerous and misunderstood.

Yeah, that, but there's also a race dynamic, right? There's an implicit idea that if I don't continue to spend, my competitors will. This is likely to be a market that's dominated by a small number of players, at least in terms of inference and the provision of tokens, these the underlying sort of atomic units of AI. And so if you subscribe to that, then you immediately enter into this race dynamic where I must continue spending because I'm at risk of becoming an also-ran or even irrelevant. And so that's what sort of locks in place at a deep level why this competitive race dynamic exists. And you saw that in the recent calls from Dario Amodei at Anthropic asking for, you know, and this is a bizarre term, but pacing the frontier, which is just, you know, a fancy way of saying let's try to slow things down, but you can't because you get into this, you know, sort of prisoner's dilemma problem that, yeah, okay, fine, what if we all agree but China doesn't? Well then we're into this problem where China wins again and we're back to this existential crisis. And so the race dynamic is baked into the very nature of what we're doing, and to your point, it's exacerbated, it's made worse by a kind of pseudo-religiosity on the West Coast of the United States with respect to technology and this idea that there's no point in trying to slow anything down because it's inexorable. It has its own force. It has its own morality and consequence, and slowing it down is anathema to this group of people. And so you've got these two phenomena operating and reinforcing one another. And then a capital markets that sees this happening, a race dynamic and this religiosity, who's more than happy to provide endless amounts of capital to try and keep the dynamic going because they're making money from fees. They're not holding the debt. They're making money from an agency business of helping to create these special purpose vehicles, off-balance sheet financing, syndication, all these things, but they're not holding on to the instrument. So, they make money from issuance, not from holding. And so, you've got all of these things intersecting that creates the dynamic that we're seeing, which now means that, you know, we're now on a... as a percentage of GDP as a capex moment, we've now exceeded all prior episodes in the last 150 years. It's larger than the telecom build-out, inflation-adjusted dot-com, PCs, rural electrification, railroads, canals, take your pick. We're only behind interstate highways and World War II rearmament. So, good for us.

Sure. There's lots of ways to get at that. So, I'll just give you one or two, maybe. So, if you approach it from the standpoint, just deterministically, let's just assume for the sake of argument that the current capex spending, the amount that's being spent on building out data centers is rational, which it isn't, but let's pretend it is. Okay, so if we pretend it is, and then we say over the next four or five years we'll probably get to 10, pushing on 10 trillion dollars in data center spending. And if we apply a reasonable hurdle rate in terms of the investor expected rate of return and then a margin on top of that in terms of... you can do the math fairly straightforwardly. It's a classic sort of return on capital argument, which a group in Columbia did recently, Bain has done as well. You get to numbers like four to six trillion dollars in revenues needed by say early 2030s. These are all very squishy, but they're ballpark. And so then you say, okay, fine. This is... I'm going to get to your question sort of the back door way. You say, well, where's that going to come from? Right? So there's only two places it can come from. And either it replaces existing spending or it's all new spending, okay? So when it's happening that fast, generally speaking, it's replacing existing spending at least in large part. So then you can say, okay, so where is this coming from? Well, we've already said that this is sort of symbolic automation. Well, that's white-collar work. So four to five trillion dollars is about 40% of the US white-collar workforce in terms of... if you look at white-collar work as a TAM, a total available market, it's about 25% of the global market for white-collar work, a little less, maybe 23. So you can start to say, okay, so in one way or another a significant fraction of the US workforce is at the very least going to see their work repriced. It doesn't necessarily mean layoffs, right? So you have to be careful about making the distinction, but what AI is doing is repricing the value of cognitive work. And as it does that, as it reprices all of that, it changes the nature of people's relationship with companies. So it would be unsurprising to see, you know, numbers at least in the low millions in terms of actual layoffs, but I think it's better to think of it in terms of the assumptions driving all of this require at least a quarter of white-collar work to be repriced. And so then you can sort of draw your own conclusions. Does that mean fractional work then for these people? Does that mean massive layoffs? And then to add a sting to the tail, once you reprice that much white-collar work, the rest doesn't stay the same price. Right? It doesn't work that way. You don't get to say, 'Well, that corner is now cheaper, but the rest of white-collar work we still pay the same.' No, everything gets repriced just to a greater or lesser degree. So, it's consequential both in terms of layoffs and in terms of, you know, what we pay for this good that we call human work. And that's baked into the nature of what we're creating if you simply accept that these capex numbers require return on investment.

Well, it's the same in the sense that if you believe that this is naturally a scale business, so there will be consolidation which is implicit in having only a couple of frontier model companies, for example, then inevitably you want to be one of the largest scale providers. And so then you're competing on the basis of the marginal cost of production in the limit. And the marginal cost of production of inference, for example, is just, you know, heat, power, and light, right? I mean, at the margin it's just what it costs me to keep the business's doors open, if we're truly competing away all the margins and competing in a truly scale sense, right? So, the response of the frontier model companies is the price will never decline to the marginal cost of production because we're so smart that we're creating all these new models for which we can charge premium prices, right? So, you'll see this constant flow of new models coming out of Anthropic or coming out of OpenAI. And the way to think about that is it's a defensive tactic to try and make sure that they can charge what the marketers call, you know, have a halo price above the cost of actually producing tokens. And that's the game that's being played as you try to produce new models and justify how powerful they are by all kinds of talk about their exploits and other things. And it's a way of making sure that the cost of selling tokens doesn't collapse to the marginal cost of production.

Oh, it does absolutely. So, there's two or three different elements of this. I'm not going to try to... I don't buy their arguments whatsoever. So, one is that you can think about what's happening elsewhere, like China's a good example, is China's trying to essentially become almost an industrial manufacturer of tokens. So, producing tokens at very close to the marginal cost of production that are say 8 months behind US models in terms of their performance, which seems to be borne out by statistics and benchmarks from people like Epoch and others. And they're producing these tokens at a vastly lower price, and that's important because let me turn the story around. There's a really important misunderstanding of the nature of what happened in the last 3 years that kind of drives many misunderstandings of where we're going from here. And that is, and this is an old tenet of being a venture capitalist, is that never trust your earliest adopters because they're really not representative of the mass market. Because by definition, they're early adopters. They're risk-takers. They have extreme needs. They're willing to pay a premium price. They're willing to use risky technologies. That's not like normal people. Normal people aren't anything like that. The first four years of frontier model usage were dominated by whom? Software engineers, right? We know that from the data. Something like 70 to 80% of frontier model usage is by software engineers. Software engineers are very extreme cases for a whole bunch of reasons. Many of them have relatively complex problems. But also just the nature of software. Think about how it works. I prompt a model with a relatively compressed prompt. Say, you know, help me build a fitness app. That's like what? Like eight words? Something like that? What comes out the other side? Might be a million lines of code. So, it's highly expansive. Right? Which is really unusual. Why is it unusual? It's unusual because think of the nature of white-collar work. What's an awful lot of what you might use a model for is... I'm from the finance world. So, I'll say, you know, take this Goldman Sachs research report that's 30 pages long and tell me if there's anything there I care about. So, it's compressive. It's the opposite of software engineering. An awful lot of the use cases, what normal people are using AI for both inside and outside of large organizations, are for highly compressive uses. So, why does that matter? It matters because it changes the nature of what you're extrapolating. If you're extrapolating from how these tools were used in the first four years on the basis of these early adopters called software engineers, you're making two huge errors. One, you believe that everyone cares about being at the absolute frontier of performance, right? Because software engineers cared. Normal people don't care. They might not even notice. I've run the equivalent of the old Pepsi Coke taste test where I hide the model behind a harness and see if people can tell the difference. And normal people have no idea. They can't tell. It doesn't matter to them. And the second obviously big error is this one, that white-collar usage of large language models is very different from how software engineers use them. So, to take it back to your original point, that makes the margin pressure even more dramatic because if you can't charge a premium price because normal people don't care whether they're at the frontier, so you launching a really advanced model is irrelevant to them, the pricing pressure becomes even more extreme.

It's both. That's the honest answer and I hear this all the time from inside. So, there's a story that's required and this is why you continue to see them launching new models and talking about how this improves on, you know, pick one of the following 17 million benchmarks, and that's to maintain the momentum of the story and say, 'We're still making material progress. Your money is contributing to us producing new models at the frontier.' And the untold part of it is if you look at composite benchmarks as opposed to just one-offing, because an awful lot of the benchmarks have been ingested by the models themselves. It's like they saw the test in advance. So if you look at composite benchmarks, what you'll see over the last 3 years is while models are improving, it's becoming somewhat asymptotic, meaning that on an absolute level, they're not improving as much as they once were. We went from like say around 18, 19% performance year-over-year on composite benchmarks 3 years ago to more like 5 or 6%. And this is hugely important. So it's being missed largely in the conversation and yet, to your point, it's really crucial to these companies because it's what allows them to continue to raise money and it forces the race dynamic to continue because you have to continue to show progress. Otherwise, the question is, and I say this all the time, that if it weren't true, the first company to stop training new models would actually be the winner because their costs would collapse, right? In a perverse way, the rational move of somewhere in here is for an AI frontier model company to say, 'We're cutting all training costs because the return on training has really declined sharply, at least in a dollars per unit improvement basis. And so now we're just going to provide inference.' The data suggests someone should have done that already. It's not happening. And the answer, of course, is you cannot do that in this race dynamic without it having the catastrophic consequences for your ability to raise money and keep the race going.

It's wildly unusual. And I'll take you back to my equity analyst days. And one of the first things you look at when you're trying to evaluate a company isn't just its growth rate, it's its customer profile. And it's not just, you know, do I know their customers? It's how many do they have, because one of the biggest risks you can have when you're following a public company as an investor or an analyst is an air pocket in terms of their revenues or their earnings. And one of the causes of that can be if they have a very concentrated list of customers, a small number of customers who account for most of their revenues, it doesn't take very much for suddenly, you know, an earnings per share forecast of a dollar 42 to get missed and it's a dollar instead just because one customer held off on some purchases. So, taking this over to the large language models, we have not seen yet. Obviously, there's no S-1 filing from Anthropic yet, but there's, you know, endless leaks about the nature of what the revenues look like. And we're told that something in excess of 70% of the revenues are eight companies. One alone, Meta, might be 17 to 18% of Anthropic's revenues. This is for a company allegedly on a 50 to 70 billion dollar run rate, this is unprecedented. I've never seen that in a company of this size to be so dependent on a small number of customers. The only place where you sometimes see it is perhaps in the defense industry where obviously your only customer is like four governments worldwide, so you might see it there, but that's very different because you're in long-term contracts and you're in relatively secure positions over a cycle. So, to see a commercial company operating in a private space with such customer concentration creates unprecedented risk as an organization because you in a sense are over the barrel in terms of your dependence on them and their ability to say to you, you know what, change this, do that, whatever else because I can see how important I am to you. So, with respect to its comparison to prior eras, there is no comparison. We never saw this level of concentration in any prior episode like this, whether it was the internet, fiber, GFC, railroads. Then one of the most unusual elements of this moment is the level of concentration and dependence on a small number of customers, and within that small number of customers, the dependence on a small group of people who happen to be doing a very anomalous activity called writing code.

Sure, we can answer both actually. Yeah, yeah, yeah, yeah. So, you can answer both pretty relatively cleanly. So, I've been tracking this idea. One of the things that first got me interested in this episode was sort of this anomalous continuing stubborn growth in the US GDP that...

didn't seem to be tied to orthodox economic statistics. And so, when you dug under the hood, you found out fairly quickly that for at least the last eight, seven, eight quarters, something like—and it's varied from quarter to quarter—25% to 60% of GDP growth on a quarterly basis has been tied to AI capex, which is again wildly unusual, incredibly distorting, but obviously hugely consequential because when you're that such a high percentage of GDP growth, it masks the weakness in everything else. So, something on the order of between a quarter and a little over half, depending on the quarter. And that's up there with prior episodes. It's probably on the order of the railroads, something on that order in terms of its materiality with respect to GDP growth. And you can ask yourself, you know, get into all these numbers and so on, and it's like, well, why do I care? Why does that matter? Well, it matters because it masks that nothing else is growing or even some things are shrinking. And the analogy I often make is that when you have a causal model of how GDP growth is working and it's wrong, you make bad decisions. And I often joke that it's sort of like, you know, when the mailman comes to my house and my dogs bark, they bark and bark and bark, and the mailman leaves, and they think, you know, I did it. I made the mailman leave. And it's like, no, you have a bad causal model. The mailman always leaves. He's got other houses to go to. So, this is the same thing with GDP growth. Let's think of it in the context of the US experiment with respect to tariffs. If you believe that tariffs work, you look at GDP growth and say, 'Oh, look, it had no impact. As a matter of fact, GDP was growing in a really sprightly way.' And that's just a misunderstanding of the nature of what was driving GDP growth. And so, this is important not because we're trying to get at some kind of, you know, Guinness Book of World Records statistic on a thing that contributes the most. It's because GDP growth leads to wealth, and we're misunderstanding the cause. And the cause is almost monolithic. It's this great, you know, sort of spending hairball called AI capex. And so, that's why it matters. It's because it's masking all kinds of other phenomena. And the same thing is true in inflation data. AI capex is badly masking what's happening in inflation, which is causing us to raise rates when most of the economy is actually fairly weak, right? But nevertheless, this is obviously hugely important, this misunderstanding. And so then, take it over to capital markets. You can sort of do the same kind of untangling and say, 'Okay, how much of the S&P 500's growth over the last few years has been driven by AI-related stocks?' You know, euphemistically, people talk about the Mag Seven, the Magnificent Seven stocks, which are, you know, largely AI-related. They were almost 45% of S&P 500 growth for the last five years. So, seven stocks, seven stocks driving almost half of the S&P 500 performance. And it gets worse. Something like 55%, properly understood, of the S&P 500 is now tech, which means that if you think you're buying a diversified index fund, you're actually buying not just a tech bet, but a concentrated AI bet. So, people are far less diversified than they might hope to be. Let's say I'm nervous about this moment and I say, 'Okay, honey, I'm going to buy an index fund.' You're just buying AI again, right? And so, we know how much it's driving the S&P 500, and we know what fraction of the S&P 500 it is. And both of them are at levels last seen, well, again, during the railroads episode in the 19th century and during the real electrification. And as a side note, the Dow Jones—which the S&P didn't exist at that time, the Dow Jones did, right? As a major financial index. And the Dow Jones, in response in the late 19th century to how the overwhelming share of railroads of the Dow Jones was on the order of half, created the Dow Jones Transports to, in a sense, to kind of like shuffle that stuff off to the side so it wasn't causing so much trouble. So, you can see prior historical financial bubbles in the indices themselves because we have vestigial remnants like fossilized tracks. The Dow Jones Transports is a remnant of one of the last major bubbles. And so, we are at the levels where the Dow pushed out the railroads, if you think about it in terms of what's happening with tech. And yet, there they are dominating the index. And the only other sort of data point I'll add to that, why I care about that, is because more Americans own stocks than ever. So, more Americans are exposed to the stock market. It's something like 52%, something on that order of things, are exposed to the stock market. And a slightly higher fraction of them have, in terms of their total financial wealth, is exposed to the stock market. So, it's one thing to say I own a couple of stocks, but it's another thing to say a significant fraction of my wealth is tied up in the stock market. Both of those are at near record levels. So, what that means is that most Americans are vastly overexposed to AI.