Paul Kedrosky - AI Risk and Institutional Portfolios

Yeah, thanks for having me. I'm glad we could make this work.

So for me, this is just one of a series of very similar episodes over the last 250 years of Western capitalism. So I think in broad sort of historical sweep terms, and then I like to narrow down and think about the underlying systems that give effect to what we're describing. And so this moment in terms of building out this particular technology that we euphemistically call AI or SI, I suppose, depending on who you're talking to, is one of those moments. And so you can go back through history, back to, you know, railroads and canals and rail electrification in the '20s, obviously railroads and canals in the 19th century, or telecom in the late 20th century. And all of these represented the emergence of what economic historians in the histories of technology call a GPT. Unfortunately, that particular acronym has been repurposed by OpenAI and used in the context of ChatGPT. So to an economic historian, a GPT is a general-purpose technology. It's a kind of platform technology on top of which other things are created. You can think of it as almost like an economic substrate if you're thinking of it in chemical terms, or a casting if you want to think about it in factory terms. It's something on top of which we build things. And so the railroads were a GPT, the canals were a GPT, electricity, all GPTs, right? They're all substrates on top of which we build things. And to that way of thinking, AI as currently instantiated is just another GPT. And so that at a very high level is sort of the way I come at this in terms of thinking about it. And then each of these GPTs have had similar themes and consequences in terms of how, I'll say, investors, but let's just think about it in terms of, you know, entrepreneurs and others and capital providers have thought about it, and then in terms of what the aftermath was of the buildup of the GPTs. And there's all kinds of dogma around this where people have sort of received wisdom and ideas about what, like, you know, all such waves create more jobs than they destroy, or even if there's mass business failure, on the other side there'll be useful infrastructure created because it is a substrate. There's all these sorts of mantras, dogmas, and so on.

Sure, sure, sure. Absolutely. All of those things. And so they're all worth keeping in mind. But at the very highest level, that's the way I think about it, is this is just another in a line of GPTs. And having said that, however, this is a very unusual one because the capital expenditure required is now higher than all prior in inflation-adjusted terms, and also it's happening faster and is repricing something that's in a sense never been repriced. So historical GPTs were mostly adjuncts to human behavior. So, you know, railroads and rail electrification you can think of as being augmenting in terms of making it easier for humans to do certain things like move large amounts of goods across space and time. Whereas AI can be augmenting, but it's actually in many ways a replacement technology. And Daron Acemoglu at MIT has written really compellingly on this, who recently won his Nobel for some related work. But the idea being that it's in a sense repricing a lot of white-collar labor, and this is why there's so much unease about this particular technology, because while it's like prior GPTs in many ways, it's very unusual in terms of the capital required, the pace, and what's being repriced.

That's the other lens I like to bring to bear on this. So now that we've established it's just one in a history or a series of GPTs, all of which have some patterns and things we can identify that have consequences for investors. And then the other aspect of this that makes it unusual is that in general, sort of capital market paroxysms, waves of enthusiasm, we shall call it in lieu of saying bubbles, these enthusiasms tend to be organized around either sort of one of four factors. Very commonly is real estate. Some of the largest episodes in U.S. and Western financial history have had to do with real estate. Obviously, most recently the GFC, the global financial crisis, but also technology. So now we have two: real estate or technology. And technology obviously underlay the dot-com moment, the fiber moment. You can even think about it in the context of electricity; it was a kind of enthusiasm about technology at that period in the 1920s. And so those are two factors that often underlie these things. So, you know, real estate, technology. The other one that you just alluded to that's really often seen is loose credit. So the provision of credit in either new amounts or in new methods can be really consequential in terms of creating these moments. And obviously we saw that in the GFC; we're seeing aspects of it again today. But you can trace it all the way back to like the South Sea Bubble and Tulip Mania. So that's the third factor. So now we have technology, real estate, loose credit, and then the fourth factor that often underlies at least Western economic bubbles is government policy. So to the extent that government is goading people into doing something, or even worse, creating a kind of put underneath it and making people feel like it's okay, we've got your back on this, that can be an immense factor. We saw that in the global financial crisis where there was this notion that Fannie and Freddie were sitting underneath it, but we've seen it in many, many historical bubbles. So to wrap on that, what's unusual about this moment is it sits at the intersection of all four of those. There's a real estate component, there's a technology component, there's a loose credit component, and there's a government component. And that's why I sometimes refer to this as a kind of rational bubble, in the sense that you have four different actors, all of whom think they're acting within their expertise and acting rationally, but it's the intersection of all their activities that multiplies the chances for issues.

15 years of scar tissue, I like to say, maybe more, 20 years.

So anything to do with large language models, training models, refining models, all of those areas have been hit with a point mass of capital, right? It's kind of a, you know, deuterium load of capital that's just hit it, smashed it to bits, and left just sort of wreckage in the landscape. And so, while I suppose it's possible maybe four and five years ago, you might have been able to do something there, I guess, but I still think you probably would have lost money. There's really nothing to do in the sort of architectural underpinnings of most of what's happening. So we're not particularly, we're not very active in AI at all, to be honest. I mean, so one of the sort of thesis we have as we look at this in terms of saying, I'll give you an example that there's many products out there that are very difficult to use. And so I'll give you like, say, Photoshop. God help you if you ever have to use that thing for anything useful. And so current-generation AI can often make it easier to use these kinds of products, and that either exposes a moat or exposes that the product wasn't necessary in the first place. So in the case of Photoshop, most people have discovered, you know what, I didn't actually need that because I can now do it with large language models. And so one of the lenses we bring in terms of thinking about it and trying to duck this sort of, you know, high weapon load of everything that's happening out there is to say, what are some products and services where AI makes it easier to do a thing that should be done in larger sort of quantities? And I'll use a public markets example. You can think of a company like Cloudflare. So Cloudflare is this caching service that was sort of, you know, came around almost 20 years ago now, but since doing that, they've added a vast number of services that are very useful to many people but absolutely impossible for normal people to understand. Who might be doing what you're doing? You might be hosting a podcast, you might be creating a web presence or whatever else. And so what's happened in the last 18 months is suddenly there's been an explosion in their business because people didn't have to actually know how Cloudflare worked and could just set up tools that would say, I need to deploy whatever it is that Cloudflare has that will help me. So we look for people, we look for areas in the private markets where there's an opportunity to rediscover rent by using these tools to help deal with complexity, and that's about the only operating thesis we have to try and avoid most of the damages that's out there, because that tends to be a more straightforward way of, you know, in both investing and avoiding the high capital cost of some of the other things out there. And it turns out there's a huge, a huge amount of...

Right, right, right. And so it's just discovering whether once you reduce complexity, is there rent or is there nothing? And that's what's interesting to us in terms of our... And, you know, we're also doing some work in security specifically, where we have a couple of investments related to introspection, model introspection, understanding the dynamics of what's happening under the hood inside of models, because right now most of security is about trying to prevent certain behaviors in terms of as they're exhibited, as opposed to understanding the process via which a model comes to a decision to do something. And so that I think as we go forward will become much more important, and it's a very, very hard problem because these are, you know, millions and billions of neurons interacting in a complex and statistically non-deterministic way. And so, you know, there are things you can do. I just think you have to be very at the margin at this stuff because the point mass of capital directed at most of the orthodox stuff will crush you.

So the transmission mechanism is very direct because the exposure in the context of, say, just pick for example the two favorite suspects, the two largest frontier AI companies, OpenAI and Anthropic, you know, they're sitting at in excess of a trillion dollars in private market valuation, which is just mind-boggling, but nevertheless, and this, you know, posited that they might see, you know, $2 trillion at an exit and IPO. Everything in venture right now and much of private credit is priced on the basis of those numbers. So that price plays backwards into what my anticipated multiples and valuations look like in the event of an exit, but it also plays back into what private credit's willing to do in terms of extending whatever kind of external financing for continuing to purchase things in terms of data centers or lease GPUs or whatever else. And so everything now, it's in this perverse mode where everything is being priced on the back of two highly illiquid companies. So if you think about it in terms of historical public markets, I would always say, you know, well, what's a public market comparable, right? Because the reason why obviously we say that isn't because I'm somehow, you know, fetishistic about public markets. It's because this is a liquid market and it's being priced on the best available evidence and at the margin by active participants. Private markets are completely the opposite, right? In a sense, the prices are capricious. They're almost ad hoc, and they're priced in anticipation of sort of almost what the last person thinks we could get whenever we eventually exit this in public markets. And so these are not great comparables in terms of setting prices or trade sales, right? GP.

And so to turn it around, the notion of pricing much of public and much of the rest of private markets on the basis of these highly illiquid investments is a very dangerous place. So the repricing of those particular instruments, OpenAI and Anthropic, rips straight backwards in terms of all portfolios in venture, most private equity portfolios, and most of private credit, because that's now the pricing comparable used for justifying almost all transactions. And of course, as we both know...

Like everything, the FT had a great article about this the other day, how everything is over-indexed on AI right now, right? In the sense that it's the primary factor moving everything. We see that in the S&P 500. We've seen that in crazy places like the impact on the 10-year. It's becoming, it's a one-factor world, and that one factor is AI. So the repricing of the two highly illiquid instruments that are driving the pricing cannot be anything but consequential.

Yeah, absolutely, vastly more. And of course, you know, every one of those buckets you just named obviously is highly correlated with the AI trade. And this is the problem, is that, you know, for example, you think about energy, right? Energy, well, the XLE ETF, which is obviously one of the more liquid energy ETFs, is now outperforming the tech ETF here today. So you say, 'Oh, great, I've got some diversification.' No, you don't. You have AI in a different basket, because the two things driving energy obviously are, to an extent, the Iranian adventure and what's happened with gas prices, but also this incredible demand for building out more energy supply, because the essence of AI is that it's tokens, and tokens are driven by, you know, megawatts and gigawatts of power. And so it's the same trade, as I say all the time, it's the same trade with a funny hat and a mustache.

Well, there's lots of ways you can come at it. So you can start off by looking at private, sort of like the public market comparables, where something like, you know, 55% of the S&P 500, which is obviously a market-cap-weighted index, is now tech if properly defined. The S&P changed back in 2018 some of the definitions. So if you put those companies back in, you're at like 55% is tech, 45% of the recent performance is AI. So as a starting point, and you can do the same thing with the Wilshire, you find out that it is this 30 to 45% of most broad indices now are in some way shaped and pushed by what's happening in AI. And so my general view is, and I talk to a lot of people in the private markets, is that it's if anything even more dramatic on the private side, because if I'm a concentrated portfolio holder in energy, for example, and I'm active in energy, you're probably even a higher level of exposure, you're probably even more over-indexed to AI than you realize, in part because, you know, everything's being driven by what's happening with, you know, gas turbines, where we've got a three-year queue, right? It's being driven by natural gas prices. Natural gas prices are being driven in large part by the provision of behind-the-meter power.

Which is kind of an interesting observation. There's no positive correlation, even though there is a positive correlation.

Right, right, right, right, exactly. So now I'm at the margin a bidder in natural gas markets for a mini gas turbine, which will have an effect on winter heating. Right. So we're giveth and taking away because of the size of what's happening. The externalities become, you know, hugely important.

We obviously don't know until we get a stress test, and we haven't had that stress test yet. So it's impossible to say, but it looks like for the most part these things are wildly under-collateralized, in part for the reasons you're describing, the depreciation base or the depreciation, I suppose, schedule of the GPUs themselves, which are about 60% of the cost of data centers. So insofar as you're collateralized against that instrument, then 60% of that data center is, leaving aside the lease obligations and what you plan to do in terms of bucketing, that 60% is the GPUs themselves. And there are things that NVIDIA, for example, is trying to do to put a floor price underneath the price of GPUs, which I think are in the limit indefensible. They can't do what they're promising to do, which is to say that there will, you know, be a permanent floor under GPU pricing because I say so. That's just putting your balance sheet at risk in a new and different way.

Because I say so.

So that's part of the problem. So insofar as that becomes the collateral here, one of the reasons why people like to point to... So let's dig deeper into this collateralization issue with respect to how you ask the private credit firm exactly how they're thinking about it. One of the things people will point to is the stubbornly high price of A100 and H100 GPUs in the rental marketplace. And they'll say, 'Well, look, they're holding their value. These are relatively, in chip terms, ancient products. You're telling me that the value is falling off.' If I look at, you know, SemiAnalysis or wherever else, and I look at the rental prices, even Bloomberg now has a GPU rental price index. So you look at those, you say, 'Well, they're very, very stable, right? These prices are stable.' But this is an unusual moment in the sense that that's driven by a little bit of artificial scarcity as well as by this rapid transition in terms of what's happening inside the data centers, that we really quickly went from predominantly training to predominantly inference, which is to say...

And so GPUs are somewhat more useful than people expected in terms of doing inference, but that's changing very, very fast because they're very cumbersome, they're power-hungry, they require a lot of cooling. And so you should note this window is wildly anomalous in terms of the support it's given to the price of GPUs in rental marketplaces. So I hear this all the time from people on the private credit side and people on the commercial real estate side that these are actually much more valuable assets, Paul, than you realize, as evidenced by what's happening in the rental market. And if you actually walk...

So that's a classic trap for people, that they don't realize what the dynamics are that's supporting the price and holding them at these levels, and they'll probably, you know, be a half or a quarter of current levels within three years, four years at the outset, at the most. So yeah, so this is the problem.

And not surprising that that augmentation of existing remit is the way people are employing this right now. But, you know, if you look at McKinsey stuff, they're saying, 'Whoa, whoa, wait a minute, you got to rip apart the entire business model and start right down.'

So I think you have to look to exactly what you're describing. And so it's not clear in terms of durable productivity gains from augmentation that we have a lot of good evidence, even in the most exposed areas like specifically coding, which is obviously where the most intensive application of AI has taken place. You're already seeing some, you know, dramatic cutbacks in terms of the amount of tokens people are allowed to use because the productivity benefits have been less dramatic than people hoped for. Which isn't to say it doesn't work. It's incredibly useful in coding. Perhaps more useful in coding than almost anything else.

And Rell, which does some really interesting and unorthodox labor market research, has some data out right now showing what's happening in two areas. One is call centers. So in call centers, there's been a fairly sharp reduction, not just in the United States but increasingly worldwide, in terms of overall employment. And this is a sector that had seen 25 years of very consistent employment growth, and you can trace the changes almost as an inflection to the recent emergence of AI models. Now again, correlation and causation, but call centers could hardly be a more ideal application for using large language models because I have a script, there's a grammar, there's a very rigid structure, and in turn I can generate, you know, now using tools like ElevenLabs and others, very compelling human voices or even chatbots on a website. And so this would be an area where you'd expect dramatic change. So is it being used to augment existing people? No, not at all, because it's much more effective in terms of driving productivity and then the bottom line by replacing people. And so you're seeing dramatic changes already worldwide in the call center. And so why is that important? And it's important because, you know, call centers are large regional employers, you know, in the southeastern United States, in places like something like 6%...

Philippines, Salt Lake City is like 6% or something of employment in Salt Lake City is call centers, which is a wild statistic. But similarly, yes, India, Pakistan, right, right, Malaysia, these are areas where it's not just, you know, a large employer, it's something like 7 or 8% of GDP, call centers. So this is not to bring in the whole macro trade and go full Stan Druckenmiller or something, but this is potentially destabilizing in terms of this sort of sovereign health if you start thinking about it in terms of mass substitution of sectors that have, you know, a large share of GDP. So to go back to your question, I think that's right now are the most obvious things in terms of people making the pitch that there's some kind of lift in terms of likely that's going to flow down to margins. Unfortunately, it's mostly in replacement. So if you have a credible case for replacing employment and headcount, which tends to be in areas that have tight grammars and are very bounded...

...that's where your uplift is. If someone's saying that across the board, I'm going to, you know, make my marketing department 10% more improvement, I would immediately call on that because it's simply not the case.