Paul Kedrosky: Partner at SK Ventures & MIT Fellow on Debt, Data Centers & What Breaks First

Generally speaking, the biggest bubbles in US history were associated with one or two of those. What makes the current moment wildly unusual is, in a Venn diagram sense, we're at the intersection of all four. So you have four people who independently think it's rational, but the intersection of their rational activities is a shitshow. We have far more ways for this system to fail because of this overdetermined nature and all these different intersecting actors. So this is a long way of saying that it will fail. It will fail catastrophically.

So, it's a classic prisoner's dilemma problem, right, in game theory. So, if the US frontier companies all, for whatever reason, agree to slow the pace of the frontier and the Chinese companies don't, which they've already said they will not do, there's your prisoner's dilemma. It breaks on the spot. No, it's an incredibly important and useful technology, but the underlying finances in terms of the economics of the business are an absolute mess which will shatter.

Oh, they've already started appearing. But see, here's the thing, the way I look at it is...

Hey Franco, thanks for having me.

Sure. Yeah. So, one of the ways that I try to get people to kind of get a deeper understanding of this strange moment is by putting it in historical context. And so, one of the ways to do that is to get beyond talking about, you know, the telecom bubble or whatever else and just think more broadly about the last 200 years of Western economic growth and some of these episodes where we've had huge amounts of capital spending, whether it was on railroads or the interstate highway systems or on the savings and loans, you know, fill in the blank, right? And so or even go back to the South Sea bubble and other moments in Western European history. And so if you look at all of those and you kind of bracket them up and think about what they have in common, one of the things that is I think really striking is that usually one of four different forces are kind of involved. And so usually there's some aspect of loose credit. It's relatively straightforward to get credit for things that you might otherwise have struggled for. You can think about that in the context of the global financial crisis and the housing bubble and you know all of the things that went on there was in part having to do with loose credit. Sometimes it has to do with real estate that tends to be one of the driving forces in the US in particular by the way but it's also in other countries. Japan's had episodes, Australia, Canada recently but nevertheless so real estate is often a factor like it was in the global financial crisis. So there's two forces now, right? We've got loose credit. We've got real estate. And then the other one that and this is the one that the tech bros obsess about is technology obviously, right? So there's technology is often a central factor driving this stuff. So it was true in the fiber optic moment. It was true in the dotcom. It was true in white collar office automation in the 1970s. It's been true at various points. You can even think about rural electrification in the 1920s as a kind of technology-driven.

And so there's our third force, right? So now we have technology, we have real estate and we have loose credit. These are all in interlocking sort of interlocking and overlapping ways often associated with these moments. And then the last one which people often forget because it's often invisible but it's really important is government policy. So often government policy is a major factor. It was in the global financial crisis where there was a big push through Fannie and Freddie to originate more mortgages to low-income people and in turn those were syndicated because they were seen as risky and then in turn those went out in things like credit default swaps that were insuring against them and then they were syndicated worldwide and nearly took down the global financial system. So now we have the four these four and I'm not pretending that there's not other factors here but these are four that are really important and are often associated with. But we have government policy, a great technology story, we have real estate, and we have loose credit. Generally speaking, the biggest bubbles in US history was associated with one or two of those. What makes the current moment wildly unusual is in a Venn diagram sense, we're at the intersection of all four, right? We have loose credit, incredibly loose credit with respect to the sorts of financing being extended in the AI capex cycle. We have the probably one of the greatest technology stories of the last 50 years in terms of the consequentiality of this thing we call AI or apparently is now SI. I just learned.

So, so this moment has loose credit. It has government policy which is really surprising and it is because it's seen as almost an existential battle for the future of life, the universe and the western economic way and you hear this all the time like we have to do this. Why? Because otherwise China will and we can't have China winning at this. So there's a government policy component that's incredibly important and provides a kind of backstop to a lot of what's going on here and you'll hear that from David Sacks and also about this but anyways government policy. So it also has so it has this loose credit component. It has this real estate component that's often overlooked, right? AI is in a sense data centers are really just apartment buildings. Instead of like carbon based life forms like humans, it's carbon based life forms called GPUs, right? And so GPUs live in data centers the way we live in apartment buildings. And from the standpoint of people making loans to these data centers and extending them credit and putting together these giant syndicated packages, they really think about it in the same way. There's tenants and the tenants are paying rent and the rent in turn it gets bundled up and comes to me as one of the owners of the property and I like the yield, right? So, it's no different. I like the yield on a high quality apartment building. Well, I also like the yield on a data center. And I've had people who are real estate investors that they tell me all the time is they don't honestly care what's going on inside the data center as long as they're getting a yield that's competitive with a reasonable spread over treasuries. Like I often joke like there could be hide-and-go-seek competitions going on in there and they don't give a rat's ass, right? They don't really care. And so, so now we have the third component. There's real estate. And there's all kinds of ways that real estate play in even in speculative ways through things like powered land companies that are buying up rights to put in shells. They don't even have any specific tenants. So it's almost like Chinatown, this incredibly speculative component to this. So there's this real estate component to it. So we've got a technology component. We have a loose credit. We have a government policy. So that's what makes this moment unusual is it sits at the intersection of all of those things and the last point on it is each of the people who are quote experts in credit or real estate or technology or even in government policy I'll look at it and say yeah makes total sense to me. So you have four people who independently think it's rational but the intersection of their rational activities is a shitshow which is where we are.

Oh, they've already started appearing. But see, here's the thing, the way I look at it is because it's at the intersection of these four things, failure is kind of overdetermined, right? In a statistical sense. So the way to think about it is like if and I often kind of use this in a figure when I'm talking about this stuff to groups is that you can think about it is like if there's six different ways for something to fail and each has a 5% chance of happening. What's the likelihood that something fails? Well, people often get this wrong, right? Because they think, well, each of these things has a low probability of happening. But if I have six different events, each with a 5% chance of happening, you have to flip the odds and say, well, what's the likelihood that none of them happen? Well, that's 0.95 to the 6th power, right? 0.95 blah blah blah blah. And the reality is it's more than 30% chance that something fails. So, we have far more ways for this system to fail because of this overdetermined nature and all these different intersecting actors. So, this is a long way of saying that it will fail. It will fail catastrophically. Predicting which one of the many ways it can fail is kind of a mug's game. I like to think it's probably going to be one of two things. The most unexpected ones would be something in sovereign debt which we're already seeing a little bit of and we can talk more about that if you want to because of the intersection of these large capital raising and what's happening in sovereigns or something completely bizarre in government policy because of this really unusual aspect of this thing where it's seen as existential. So you could imagine government policies changing like I have to own 25% of all frontier model companies because they're essentially defense strategic now right because you're seeing them they've been they they're appearing now inside of you know Pete Hegseth's been very straightforward about this that Anthropic played a major role in terms of their there the DoD's use of it in the context of the Venezuelan adventure they were trying to do the same thing in the context of Iran. So, in a weird way, AI has already become so geopolitically strategic that it's essentially allowed people to lower the barriers to war, which is a crazy thought. But turning it around, if I'm the government, what are my thoughts then? Well, I don't want this to go to anyone else. In the same way, I don't want nuclear weapons to go to anyone else. So you can imagine a world in which a large equity stake gets taken in these firms which changes the valuation changes the likelihood of them being able to export that changes the likelihood of you being able to service the debt so there's that I think is a really material way it could fail and the other way is that we're increasingly seeing cracks because at the medium term in the treasury market one of the conflicting pressures going on is that the 10-year has obviously gone out materially over the last little while and you know sitting in excess of at 5% but one of the factors there is it's now competing against debt being issued by the large hyperscalers which is really unusual.

And one of the reasons why it's happening and Peter Berezin at BCA who you've had on previously is really good on this topic is that the structure of the sovereign debt marketplace is not your mom and dad's market anymore. It used to be that the dominant buyers in that marketplace were yield insensitive to use the central banking term of art. They had to buy treasuries because they were trying to sanitize dollars that they were receiving in the context of trade flows. So it was just it doesn't really matter what the yield is. I got to buy this stuff anyways because I have all these dollars and I have to do something with them. That's completely flipped in recent years. Now it's much the buyers are increasingly strategic and yield sensitive buyers, institutions and individuals and others and they care what the yield is. So when the yield goes up, all of a sudden the flows change dramatically. And so they'll move into and out of this market in ways that are, you know, much less predictable, which is causing some of the cracking in the marketplace because you now have data center yields competing with sovereign yields, which changes the nature of whether or not people have a higher or lower likelihood of becoming buyers. So, so those are the two, I think, most likely ways that you'll see the cracks. You're already seeing that, you know, the Project Hyperion, one of the larger data center deals, spreads around to like 10%, it's trading at like 89 cents on the dollar. This is only like an 18-month old deal. This is, you know, really important signals of stress in the debt marketplace already very early. You saw Oracle inserting force majeure clauses into its contracts so that it can potentially rewrite some of the terms in the event the data centers, they'll get built. So you're seeing all of the early cracking and sorts of things you would expect to see as the stresses grow. So yeah, I think that's one place it could fail. And the other, as I said, unexpected place would be because most capital markets investors aren't used to the idea that government could do something haywire in the middle of this. And that's a huge factor here given the sort of existential nature of it. So those are the two that I'm watching most closely. Not so much technology, not so much the real estate component, which is kind of perverse because it is a technology story.

Yeah, I mean the debt investors don't care whether or not token prices are falling 80% year-over-year. They're buying yield. Right. Back to the original conversation from their standpoint. This is just a commercial real estate a sort of a CMBS project. And so I just want to make sure I get the yield that I promise to my investors. And the way that I get it is by holding your feet to the fire with respect to the yield from this stuff. But unfortunately what's happening under the hood is it's not hide-and-go-seek competitions and not laser tag. It's data centers and the data centers are predicated on producing that yield by selling in a sense at an industrial scale this really unusual and new commodity that we call tokens and not the crypto kind but the AI kind and one of the weird hallmarks of tokens is that they're profoundly deflationary because of well for a host of reasons which we can get into or not but the upshot of which is they've been falling anywhere from 70 to 80% year-over-year for the last four to five years and there's no reason to expect that to change. It could even accelerate. So, as a result of that, you could the maths are pretty simple. Just to stand still from a revenue standpoint, given a commodity whose price is deflating 70 to 80% year-over-year, you have to grow in excess of 400% year-over-year just to stand still, right? And that creates this sort of, you know, the extra level of consequentiality where if you actually want to grow, I'm promising Wall Street, I'm going public. I'm Anthropic. I'm OpenAI. I'm whomever. I need to I can't just stand still. Standing still gets me no multiple, right? As an IPO. I need to deliver growth. And so, if I'm going to deliver growth, I can't just grow 4 or 500% a year. It's probably going to have to be double that. I'm going to have to grow in excess of 1,000% a year in terms of token usage. But the pro, you know, just to deliver Wall Street kind of growth in the most optimistic case. So, you can see the problem, this geometric problem that it faces. No commodity in technology history has grown that fast on a sustained basis. Doesn't mean this one can't, but it creates this really unusual pressure because of the deflationary nature of this, you know, this wacky commodity that we call tokens. And that's ignoring that increasingly tokens are really turning into a commodity. We have now for the first time like an industrial producer of tokens in the context of China. China as you can think of is kind of producing at an industrial scale token production. Think of like the giant Chinese companies as kind of like inference factories as opposed to training factories. And so we've seen in the past what happens when China decides it wants to penetrate a market as a low-cost producer. We saw it in photovoltaics. We're seeing it increasingly in EVs and cars. I expect the same phenomenon here where you have this low-cost producer competing against the frontier AI companies who are trying to maintain a pricing halo and that halo is coming under huge pressure but the halo is that pricing halo is the only thing that allows them to service their debt so now the debt comes under pressure so this is a huge huge problem in terms of the dynamics for these companies as viable functioning financial entities nothing to do with the technology this is not saying that you know AI is like I don't know some weird fad or something no incredibly important and useful technology, but the underlying finances in terms of the economics of the business are an absolute mess which will shatter.

Not only that, I mean, here's the thing, and you know, it's great to go back and read the history of railroads because that it played out over 40 years, the deployment of railroads. We're doing it in we're doing in four what we did in 40, right? And so the opportunities for excess are just so extreme. Over the course of 40 years, the railroad buildout triggered at least three major financial crises. So each time with serial failures of banks, railroad companies, investors, and others, you should expect some level of crisis on par with those, but it'll probably be all in one fell swoop because it's all happening so fast. There's no opportunity to have a, you know, have a crisis come and kind of clean out the madness in the system. It's not working that way because we're doing it so quickly. So the railroads are a good example of how the underlying infrastructure was useful and it rolled but it rolled out over a much longer period of time and precipitated a whole series of crises unlike the current moment which is happening so much faster. So the usual response from the tech bro crowd is that well yeah but the railroads were useful even if all the companies behind them went bankrupt right and so which is absolutely true I mean that I didn't use a railroad for 10 years doesn't mean other than a bit of weeding or something that doesn't mean it became useless but that has no bearing on this unusual capex that we're building out now a GPU in 10 years is functionally worthless it may be worthless in five but is certainly worthless in 10 even the data centers themselves because you'll get the argument that these data centers the shells that contain the power and the cooling and everything else and you know are essentially the apartment buildings for GPUs will be worthless worth something. But part of the problem is because of the speed of the buildout a lot of the power is what's called behind the meter. So it's people are bringing power with the data center because you they can't just connect to the grid because you and I as consumers complain and then we tell our local congressmen and congressmen jump up and down and then there's a midterm election and crazy things happen. And so, by bringing behind the meter power, which seems like a solution, what they're really doing is creating a new legacy problem, which was we have all of these kind of white elephant gas turbines that are being thrown behind data centers, which solves the problem of grid connection because I bring my own power, but now we have a new somebody new bidding in the natural gas marketplace for the same commodity, which will drive up winter heating prices. So, that has a consequence. And then over time, we're going to have to replace all those gas turbines. So, we're able, you know, paper the country with gas turbines. Kind of like what's that island? Easter Island. It's going to be like Easter Island statues all over the place, but there'll be gas turbines, not just some guy's head. And no, everyone will wonder why did we have so many of those things? It's because of the speed of the buildout, which caused this real misallocation of dollars behind an unusual and easily orphaned source of power. So these are all things that are kind of set in motion already that are really easy to predict that will those will all turn into problems but because of the speed of the roll out it's happening at a pace that you know it's sort of baked into the system.

Let me by the hang on just one sec just before we get there. I went off on a tangent about Easter Island before I finished my point and I forgot where I was going. So, what I wanted to say was the bro response is all of these assets will one day be will still be useful in 10 years. And of course, as I just the answer is no, they won't. The GPUs won't be useful in 10 years. Our Easter Island of natural gas turbines won't be useful in 10 years, right? They might even be outlawed in 10 years because of emissions issues. I have no idea. But unlike prior episodes, railroads, canals, rural electrification, fiber is a great example. Unlit fiber is useful long after you lay it in the ground. This doesn't have that characteristic in any meaningful way. And people who tell you that, it's pure wish thinking. It's or it's cargo cult stuff where they're like, well, it kind of looks like past episodes and in those cases the stuff was still useful. But yeah, look at the nature of what's being constructed and what's likely to be orphaned and what's likely to be useful. It's vastly different. So that means that all of this capex has to pay for itself now. You can't count on it being paid for later because there is no later because of the nature of what we're building out. And the comparisons to fiber or to railroads or whatever else are very flawed and kind of like you know really just wishing. So anyway, sorry that's I meant to wrap that point up.