Plain English with Derek Thompson - Plain English BEST OF: This Is How the AI Bubble Could Burst

Episode Date: January 27, 2026

Throughout December and January, we’re going to be re-airing some of our favorite episodes of the past year and beyond. This list includes interviews that really stuck with me and some others that you guys had tons of feedback and thoughts on … including this one! “This Is How the AI Bubble Could Burst” originally aired September 23, 2025. If you have questions, observations, or ideas for future episodes, email us at [e-mail]. Host: Derek ThompsonGuest: Paul KedroskyProducer: Devon Baroldi Learn more about your ad choices. Visit podcastchoices.com/adchoices

If you're a fan of the inner workings of Hollywood, then check out my podcast, The Town, on the Ringer Podcast Network. My name's Matt Bellany. I'm founding partner at Puck and the writer of the What I'm Hearing newsletter. And with my show, The Town, I bring you the inside conversation about money and power in Hollywood. Every week, we've got three short episodes featuring real Hollywood insiders to tell you what people in town are actually talking about. We'll cover everything from why your favorite show was canceled overnight. Which streamer is on the brink of collapse? And which executive is on the hot seat? Disney, Netflix, who's up, down, and who'll never eat lunch in this town again?

Follow the town on Spotify or wherever you get your podcast. Hi, everybody, Derek here. In December, my wife and I welcomed our second baby girl into the world. I'm going to be taking some time off, but we wanted to keep the pod going through the holidays. So we're going to be re-airing some of our favorite episodes from the last 12 months, a kind of best of compendium. And this list includes interviews that really stuck with me and others that really stuck with you. and you had lots of feedback and thoughts on, including this one.

I'll be back in the new year with fresh content, but until then, happy holidays and happy new year. Today, the AI bubble. This year, American tech companies will spend about $300 to $400 billion on artificial intelligence. That's more in nominal dollars than any group of companies has ever spent to do just about anything. And notably, these companies are not anywhere close to earning back that $400 billion that they're about to spend. This is why you're starting to hear some people wonder if the AI buildout is turning into the mother of all economic bubbles. Sometimes you'll hear this case from critics of the technology.

Critics will sometimes point out that we're on track to spend trillions of dollars this decade building something that might be all smoke and mirrors. I'm more interested, though, in the boosters of artificial intelligence. They'll sometimes argue that we are living through a transformative tech akin to the creation of the internet or the railroad or the telegraphes. I think they might be right. I also think they don't realize what being right would imply. The infrastructure build out of the internet created an enormous bubble in the late 1990s and early 2000s. The infrastructure build out of the telegraph created another bubble in the 19th century. The construction of the Transcontinental Railroad System, as we explained in a previous episode, created several bubbles, ending in the panic of 1857, the panic of 18507, the panic of 18th

1973 and the panic of 1893, a half-century of panics. In the 20th century, radio was a bubble. The dawn of automobiles and aviation companies also quite bubble-licious. In short, if AI's boosters are right with their comparison of AI to the greatest technology of the last 150 years, their own analogy anticipates that their product, too, will pass through a calamitous crash on the way to changing the world. This should absolutely scare you if you care about the U.S. economy. Half of GDP growth comes from infrastructure spending on AI, on data centers, chips, and energy. More than half of stock market appreciation in the last few years comes from companies associated with AI. If you open up

the hood of these biggest companies, meta, Microsoft, Alphabet, Amazon, AI infrastructure spending or CAPEX accounts for, you guessed it, nearly half of their revenue. If the AI spending project blows up in the next few years, as our next guest says it might, the implications for technology, the economy, and politics would be immense. Paul Kedroski is an investor and writer. Today we talk about the AI boom, how it works, who's paying for it, and how they're financing it. We put the AI build out in historical context, and then we spend a great deal of time walking through what could go wrong and when it might go wrong.

I'm Derek Thompson. This is plain English. Paul Kedroski, welcome with the show. Hey, Derek, good to be here. Before we start, who are you? What do you do? Yeah, that's a good question. So I have a couple of day jobs. One day job is I'm a partner with the venture capital firm called SK Ventures,

where we're mostly doing early stage investing, which is to say high failure rate, low capital, most things break. And then I also sit in as a, as a fellow at the MIT Center for the digital economy. So this is sort of closer to the spirit of some of the things we're working on. And then I also have a newsletter that goes out to a bunch of hedge funds and generally to hedge funds and buy side firms and things like that. Just because my background way back when was I was on the sell side, I worked for a brokerage firm.

And I've just never been able to shake that. So I can't help myself. Sometimes I just provide, I want to give them advice and whether they like it or not. And so I still do a lot of work with a bunch of hedge funds and buyside firms, which takes us back to data centers and AI and blah, blah, blah. Well, you should know your newsletter doesn't just go out to hedge funds. It also goes out to podcast hosts, which is one reason why you're on the show. I heard, yes.

One thing I find so interesting about your analysis is that artificial intelligence is sometimes talked about as being the technology of the future. And I'm trying to ring the bell very loudly that AI is the most important economic phenomenon of the present. It is here. It's happening right now. And you've been sounding the alarm maybe more than just about anybody or more effectively than just about anybody. And just how massive U.S. investment in artificial intelligence is by historical standards.

So why don't you just start with your thesis statement? How big is this? So, yeah, let's maybe go, can I go back and tell a quick backstory here first? Just because what got me interested was I sort of what you're describing, which is there's a huge amount of money being deployed. it's going to a very narrow set of recipients, some of these chip firms and others, and it's going to some really small geographies like Northern Virginia. So it's an incredibly concentrated pool of capital, and yet it's so large that when you do the aggregating of the math and do the math, it seems to be large enough to affect GDP.

So I was saying, okay, fine, this is crazy. I should do the math. So I did the math, and it's found out that in the first half of this year, the data center related spending, so spending on these giant buildings full of GPUs and racks and servers and what have you that are then used in turn by the large AI firms to generate responses and train models, that that probably accounted for something like half of GDP growth in the first half of the year, which was absolutely bananas. And I was like, I did the math four or five different ways trying to prove myself wrong. And then I said,

okay, fine. This feels like something I should mention. And so I said it. And it's, I think it's a startling figure for a whole bunch of reasons, one of which you alluded to, which is that even compared to historical spending, whether you pick the telecom bubble or railroads or whatever else and we can dive into those, it's unprecedented. It's also unprecedented because of the nature of the spending, which I think is incredibly important because of the railroads are very different from GPUs, not just in the trivial sense, but in some very deep and important ways. And all of this gets missed. But the upshot is spending is huge. It's driving the economy. It's driving the economy. People are very confused about this, and as a result, you end up making bad policy decisions,

because you think policy decision A is driving the economy when it's this wacky stuff over here on the left. So we're talking about infrastructure boom that is on par with the broadband build out of the 1990s, early 2000s, still behind, it seems like the railroad boom of the 19th century. But we're talking essentially about an amount of spending on one emerging technology that is without precedent in at least 60 to 100 years. How does AI CapEx break down? We're talking about capital expenditures. So money that's being spent on essentially machines rather than people. How much of this is chips versus energy versus building the actual data centers themselves?

Is there a good way to think about where all this money is going? So a little more than half the cost of a data center are the chips that are going in. So say 60 It varies depending on the model of the data center. Because there's a whole bunch of different styles of data centers, if you will. There are some that are built almost on spec. Think of companies like Corweave where they're buying it. And it's almost like they're hoping to tend into building. Think about it as commercial real estate.

And I'm hoping to get people to move in. I'm building a shell and people are moving in. And I'm hoping to get tenants. And then the tenants pay rent. So think of it in those terms. And then there's the metas and the Googles and the Amazon's where they're using a huge amount of what they're building, which again, roughly 50 to 60% of it is the GPU cost. The rest is a combination of cooling and energy.

And then a relatively small component is the actual construction. So think about the frame of the building, the concrete pad, and purchasing the real estate. So you can break it out that way. So it depends a little bit on what you're planning to use it for. If I'm trying to build something that's for training, well, I'm going to buy more expensive GPUs. I need the latest products from NVIDIA. If I'm building something that's more for inference, meaning that I'm just going to have using it largely for people that are trying to generate responses, I'm hoping, well, then I don't need the latest GPUs and I can cut costs and cut some corners in there. So you can think about it as that continued.

You compare the infrastructure boom several times to the railroads, the fiber build out. You also indicated that there's ways in which that analogy does not hold up. So I want to get into that, right, what the analogy misses. Like the rail that we laid in, say, the 1860. was still around in the 1890s, maybe some of it's still around the 1950s, right? The fiber optic cable that was laid still works for years. But I keep seeing news headlines about GPUs getting better all the time.

So I wonder, like, are companies buying something they're going to have to replace in like three years? Talk a little bit about how AI is just fundamentally different than steel rail or fiber optic cable in a way that's really important in understanding what it is that we're building here. So I'll start at a high level. This is my complaint about economic statistics in general. This is plain English, right?

So capital expenditures is such a misnomer in such a, I'll say misleading, but it's not consciously misleading. It's just that it's so aggregated up. It assumes that everything I'm spending money on, that's capital expenditure, and AI data centers are lumped into that loosely, as we're railroads, as are the dot-com construction. but they're all very different for exactly the reason you allude to, which is the lifespan of the thing you're creating is wildly different. So when you think about comparing railroads, which let's say I built something during the railroad bubble in the 19th century, say 1855, I only got around five years later to running traffic down the line.

What were the biggest issues I potentially had? Weeds? I don't know. Weeds, maybe some cows had settled in. It's not clear to me what it was what the forces I was pushing against were. So in all likelihood, I could, quote, light that line very, very quickly. I could put it into use relatively easily. Similarly, with the fiber boom back in the late 1990s, if I was putting in some gear from, you know, Sienna and JDS Unifase and all the great names of that era and building out fiber and it was like, wait a minute, nobody wants me to light the fiber.

Nobody wants to send data down it. Netflix isn't doing that yet. Okay, so I wait five years and I light the fiber. What are the things that then are pushing against me? Nothing. Maybe someone accidentally put a backhoe through a fiber line, but I can very quickly put that back together. So it's not as if the fiber optic cables themselves are depreciating asset, that they're becoming less useful over time because I didn't light it to send some streaming stuff for Netflix down it. It has no bearing. Now we turn to the current wave of

quote capital expenditures, and that's why this is very different. The lifespan of a GPU is on the order of two and a half to three and a half years. This is nothing like the spending that's being done on railroads were five years later, completely the same value, or in the case of fiber, a couple years later, light it for Netflix, no problem. If I build a data center today and populated with Blackwell GPUs from Nvidia and hoping three years, I can get the same rental prices that I could have gotten today, I'm dreaming. There's likelihood I'll get a fraction of that, if anything, at all. So these assets decline exponentially, which is completely different from all these historical bubbles that we were spending on the same levels. So you have this problem that you need to recoup your

investment if you're doing this very, very quickly because you're sitting on a depreciating asset, which creates this perverse problem. So when people look at the fact that these hyperscalers are spending $200, $300, almost $400 billion a year on, we can call it CAPEX, we can call it, you know, just AI infrastructure, there's a way in which if you're rooting against a bubble, you could say, Well, it's like building a railroad. You can use it forever. So the $300 billion that's being spent right now can be made useful 10 years from now, 20 years from now. That's the railroad analogy.

On the opposite side of railroads, there's like bananas, right? If you spend $300 billion on bananas today, your CAPEX isn't worth shit in like two weeks because all the bananas are brown. And, like, I don't think GPUs are like entirely like bananas, but they're also. not entirely like... They're closer to bananas than anything else. It's worrying that they're closer to bananas than steel. And so, like, what does this tell us at a high level about the value of this kind of spending and the threat that these companies are just not going to be able to return

capital from all this upfront investment? So it's... And I hate to say this, but it's reminiscent in some ways of Bitcoin Treasury for companies. So this idea that it made no sense that companies were like strategy, like micro strategy, Michael Saylor's firm, were being rewarded for putting Bitcoin into Treasury. Their market value was increasing by $2 roughly for every dollar of Bitcoin they put into, quote, treasury.

So we have the same perverse phenomenon happening here, which is that the market's rewarding you for doing this even though it makes no economic sense to spend at this level because there's no way I can recoup the value of the capital spending I'm making over the next three years. So then you're forced to do these kind of wacky shell games where you say, well, it's okay because the shell itself, not to use shell for two different purposes here, but the shell, the building itself will actually be valuable in five years. It'll still have energy. It'll still have water.

I'll still be able to cool things. The walls will still be standing. I'll just swamp out the GPUs. But as we talked about at the outset, the problem is the GPUs are the majority of the cost. So the shell to a first approximation is the thing I'd like to write off. I don't want to have to write off GPUs every three years because they're the most of the cost of the thing that we call it a data center. So you can't play the game of saying, well, there actually is capital here that money that I'm spending that will be valuable in a couple of years. Again, unlike telecom, unlike the fiber boom, unlike in railroads.

There is actually two assets here, one that's long live, a building, which is essentially a small. fraction of the cost of the center. And then there's one that's very short-lived, which is the GPUs, which is the thing we'd like to have last and doesn't, but yet represents as much as 60% of the cost of the data center. So there's the perversity, and that's the problem. And before we talk really deeply about how this could lead to a bubble or a crash, I do want to talk about how this is affecting the economy right now. It's eating tech jobs. There was a University of Maryland study that found that if you subtract out AI jobs from all IT jobs, basically all IT is declining when you subtract out AI.

Like tech is just in large part becoming an enormous employment bet in the future of artificial intelligence. I'm also looking at the fact that like, you know, construction jobs are declining, mining jobs are declining. Manufacturing jobs are declining in America, despite the fact that the tariffs are nominally about re-industrialization. It almost feels to me, Paul, like AI is like this, like this star that is pulling in all of these resources gravitationally from throughout the economy, in your own words, and hopefully I got

you started along the right track, how do you see AI spending warping the 2025 economy? Yeah, so the analogy I draw is looking back. You can see how a similar effect happened. That is, massive capital spending in one narrow slice of the economy during the 1990s, caused a diversion of capital away from manufacturing and away from small manufacturing in the United States. There's been some good studies on this showing exactly the effect. And it's not surprising because people were rewarded for accepting that money to build out telecom and they were rewarded for spending that money because, look, I'm spending in an area with high returns. But what that did was it starved small manufacturers of capital, which made it very difficult for them to raise money

cheaply, which raised their cost of capital, meaning their margins had to be higher. Now, let's follow that along. During that same time, China had entered the World Trade Organization, tariffs were dropping. We've made it very difficult for domestic manufacturers to compete against China in large part, not entirely, but because of the rising cost of capital, because it all got sucked. To use your death star term, it all got pulled into this death star of telecom. So in a weird way, we can trace some of the loss of manufacturing jobs in the 90s to what happened in telecom because it's It was the great sucking sound that sucked all the capital out of everywhere else in the economy. The exact same thing is happening now.

There is no reward for spending money. If I'm a large private equity firm, if I'm any kind of large capital allocator, anywhere else but in data centers, which is why if you watch the announcements from places like Black Rock or from Blue Owl or from any of the large private equity firms or private debt providers, the thing that they're making the most noise about and they're most excited about of these giant multi-billion dollar checks, they're writing towards data centers. And so, again, the same phenomenon. If I'm a small manufacturer, and I'm hoping to benefit from the onshoreing of manufacturing

as a result of tariffs, so it's leaving aside whether they're good or bad economic policy, but I want to benefit from it. So I go out trying to raise money with that as my thesis. The hurdle rate just got a lot higher, meaning that I have to generate much higher returns because they're comparing me to this other part of the economy. that will accept giant amounts of money, huge checks I can write for this to data centers, and it looks like the returns are going to be tremendous because look at what's happening in AI and the massive uptake of open AI. So I end up inadvertently starving a huge slice of the

economy yet again, much like what we did in the 1990s. It's such an interesting interpretation because the story that we tell is that trade with China took our jobs. The China shock, as economists like David Autour call it, moved manufacturing to China, and that is what's hollowed out the Rust Belt. You're saying, yes, trade with China might have been a factor at the margins, but also the telecom buildout took capital once allocated as manufacturing and moved it to tech. And what's so interesting about that is if you fast forward to the 2020s, Trump is trying to reverse the China shock with the high tariffs, but we're recreating the capital shock with AI serving as the new telecom. So rather than reverse

the conditions that led to the decline of manufacturing, the Trump administration is ironically recreating those conditions in a way that's hurting manufacturing even more, with all of this money moving toward AI and away from traditional manufacturing. It's such an interesting idea. Yeah, and it's even more insidious than that, and this requires some inside baseball, and it's insidious because, let's say you're Derek's giant private equity firm, and you control, I don't know. Let's say you've got $500 billion burning a hole in your pocket. What do you not want to do? I do not want to allocate that money, one $5 million check at a time to a bunch of manufacturers. Because all I see is a nightmare of having to keep track of all of these little companies doing who knows what and everything else.

What would I like to do? I'd like to write $30-50 billion checks or $30, you know, I'd like to write a small number of huge checks. And this is a dynamic in private equity that people don't understand, that capital can be allocated in lots of different ways. But the partners of these firms do not want to write a bunch of small checks to a bunch of small manufacturers, even if the hurdle rate is competitive, even if they're operating at a level where they can compete against what the perceived return is on data centers, because I'm a human. I don't want to sit on 40 boards. And so you have this other perverse dynamic, this other perverse dynamic that even if everything else is even, is not equal. So we've put manufacturers who might otherwise benefit from the onshoreing phenomenon at an even worse position, in part because of the internal dynamics of capital. What about the energy piece of this? So electricity prices are already rising. This revolution

in a way is just getting started, and these data centers are incredibly energy thirsty. How much do you think is this going to result in energy inflation that becomes an economic consumer and even political problem, that these data centers are essentially seen as a lever on electricity inflation such that you've got your average Joe saying, why is my economy essentially a temple to AI? And all it does is make it harder for me to keep my child's room 69 degrees while she's sleeping. Like, how is that going to play out? So that's already beginning to play out in the strange as possible ways. The most obvious way is that we're already seeing energy inflation, in part driven by, again, consumers being outbid, if you will, by data centers. Because

it's almost like the private equity phenomenon. If I'm a utility, I would love to have a bunch of large people I can put on, large buyers, I can put on the grid because I can manage them. They're good for payment. They're not going to go away. But a bunch of people in some exurb in rural Virginia, yeah, I don't know. Maybe they'll pay their bills. Maybe they won't. So again, it's the same phenomenon is happening at the larger scale in terms of the capital allocation. I'm perfectly happy to put these large buyers onto my network, meaning the power grid, because I feel like there's some security of payment. But there is this perverse thing happening at the same time. And we saw this in the PJM recent regulatory filing that got kind of ditched, but they were proposing to add people

to their network. They're an interconnect provider in the Northeast, one of the largest. And they were proposing to add data centers to the grid with the proviso that any time the utility is the grid is under stress and I have to cut back because of a heat wave or because of heating or whatever else, I can disconnect the data centers because, you know, it's fine for a few hours if they disconnect the data center. So they're trying to have their energy cake and eat it too by connecting these large buyers but saying it's okay, don't worry consumers because we'll, we have a provision in our agreement with these data centers to disconnect. them if things become really difficult. Well, of course, if you're a data center, you're like, yeah, no, that doesn't work for me. And that's not going to work. And so you can see if you look underneath the hood, how the tensions are beginning to play out in ways that cannot be resolved straightforwardly. We have energy inflation on the one side, and we have the somewhat dodgy interconnection agreements being proposed where we propose that someone will actually be cut off from the grid, meaning a data center will be cut off from the grid.

And that's just not, it's not going to fly. Even if I sign on to that now, rest assured, I will sue you in two years if you do it to me. 100%, the lawsuits will be just massive. Even if I agreed to it today, I will sue you in two years. So what's a reasonable prediction here? If, you know, consumers are going to, I think, start to be much more explicitly nimbly about the construction of local data centers. If they see these data centers as essentially, not just, you can either say it,

stealing energy or basically a very clear lever on raising energy prices. The data centers have enormous political power of their own, enormous economic power of their own. These are large, rich companies that can spend a lot, pay a lot maybe for that energy. I mean, how is this going to play out, do you think, in two years? So I think you're going to rapidly see an offshoring of data centers. So that will be the response. It'll be increasingly be that it's happening in India, It's happening in the Middle East, for example, with massive allocations are being made to new data centers. And it's happening all over the world in China, where to the point that there was recently a warning from the Chinese government that every city does not need its own data center. Because what they're trying to do, obviously, is create a massive oversupply at the local, regional level.

There's an incentive to create these things. And so you'll see lags aside because it's nice to have a data center located locally to you in terms of actually providing services. nevertheless, the focus will increasingly move offshore for exactly this sort of NIMBY-esque reason. And there's been some great – Bloomberg had a great story the other day about an excerpt in Northern Virginia. It's essentially surrounded now by data centers. This was a – was previously a rural area and everything around them. All the farms sold out. And people in this area were like, wait a minute, who do I sue? I never signed up for this. I never signed up because at night they would go outside their houses and they hear hum. And it's like I didn't sign up. This is the beginnings of the NIMBY.

because it's become visceral and emotional for people. It's not just about prices. It's also about having this six-acre building beside me that's making this noise all the time. This is not what I signed up for. So I think the pushback has already begun, and it'll become much larger within two years. And increasingly, the largest construction will move elsewhere. I want to talk about how some of this might go badly in the next few years. And I want to preface that discussion by saying that when I talk about AI as a bubble,

I think some people see me as being pessimistic about the technology. The railroads were a bubble. There was a panic of 1857, of 1873, 1891, I think. There were constant railroad depressions, and also the railroads changed the world. Broadband was a bubble. It also changed the world. Big infrastructure buildouts that change the world often passed through a bubble phase. So it's not particularly pessimistic to say that AI is currently in a bubble.

You could say it's actually incredibly historic. historically in tune to say that we could, we are very likely in the middle of the bubble, because every industrial revolution like this passes through bubble phases. So let's start here. How close are the hyperscalers, meta, Google, Microsoft, the big boys, how close are they to aligning spending and revenue in the AI space, right? Or how far, I guess you could say, On the other hand, how far are they from seeing what could be plausibly called AI revenue catching up with AI spending? Nowhere near. I'll say first, I agree with you about the bubbles.

I mean, my general argument is you never know if you've spent enough on capital until you spent way too much. So it's like Michael Kinzley used to say this, sort of had a similar wording. But the notion being that you're never going to have a rational expenditure of capital on new, on new equipment and do it in a way that makes economic sense all the way up. You will eventually spend too much and then pull back. So let's take that as a given. It's, as you said, it's just part of the process of building out. But the deeper issue is, are you going to get to a point where it's obvious that the companies are stretched in terms of, let's take, for example, most of the publics that we're familiar with, the hyperscalers, are spending as much as 50% of income on

CAPEX, which is unprecedented. This doesn't happen. Normally, if I did that as a Microsoft or an Amazon, I would absolutely be taken to the woodshed and beaten by investors because that's such an incredible investment, not just in terms of capital expenditures, but on one narrow slice of CAPEX, that you're going to be punished for that. So they're not being punished for that. So what are they doing instead? And this goes to your point about what we should be watching for in a sense. There's a way of thinking about it. It goes back to economists like Hyman Minsky and others that what you start looking for are whenever the mechanisms that they use to raise money to do this become increasingly opaque. So what I'm watching is how they're moving the

financing off balance sheet. Because that's a way, for me is a reflection of I don't want the credit rating agencies to look at what I'm spending. I don't want investors to roll it up into my income statement. So what we're saying increasingly are these SPVs, these special purpose vehicles being created where I have a stake in it as meta. Some giant private debt provider, credit provider has a stake in it. And yeah, okay, fine, the data center at the end is under my control. But hey, hey, hey, I don't own it, right? And so you don't get to roll it back into my balance sheet in terms of assessing my creditworthiness. It doesn't change my credit rating. It doesn't change my income. So we're seeing for the first time over the last six, seven months, the beginnings of a

wave of these special purpose vehicles and other more exotic financing structures. We're seeing the equivalence of some of the old collateralized debt obligations emerge where there's tradable debt interests in data centers. These are all, for me, the beginning of the sign that the bubble is becoming tired. Because the market is beginning to punish, at least there's perception that the market will punish me if I continue to keep this on my income statement. So I won't. I'll move it somewhere else.

And that makes the entire process much more opaque. It's almost obfuscatory in terms of preventing people from understanding it. Like, how do I go through all these, the footnotes of all of these statements? And so that, for me, is that's the thing to watch. People get hung up on, I think, a lot of the wrong things in terms of trying to assess what's going on. Like, for example, is the rental rate of GPUs now competitive in terms of the actual costs of running the center? These are good things to look at, comparing your cost to your rental rates, but look at it from the company's standpoint. How hard are they trying to hide the expenditure? And for me, that's the factor to watch. And it's just begun accelerating. I feel like people who remember 2006, 2007 are feeling their eyes start to twitch as you talk about

this general law that I love the way you put it. It's kind of like you know your behavior is unethical if you try to keep it a secret. You know your economic activity is bubbleicious if you try to dress it up in financial opacity. Let's talk about just exactly how this works. I've seen you talk about this in other interviews. I think it's really important to understand how these data centers are being built

and specifically how it's not as simple as, oh, META just has a line item in their overall spending that says, and then we bought a bunch of land near Ashburn, Virginia, and built a data center there. What's happening is the hyperscalers like META are getting together with the private equity firms like Apollo and they're both putting money into this box, right?

These special purpose vehicles and that box is the thing that's investing in these data centers. Just take me through exactly how this works. So if I don't, so the idea, obviously, is you can look at it from a couple different ways. One is that the private credit firms, the Apollos and others, want a stake in data centers, but they also, they want their stake in data centers to be in a data center that has built-in customers. So from their standpoint, I can write a large check, and if I partner with a meta or a Google or whoever,

there's a large, a high likelihood that will immediately people will be using. It'll be populated. There'll be rental income flowing back from it because that's what they want. Think of it like interest on a debt note. There'll be interest flowing back because people are paying for hourly usage of these GPs and that flows back to me as a partial owner of this data center. And that's what I'm looking for. A stake in that rental income, no different than having a stake in a note that I've extended

to someone else in the form of debt. So now I have multiple. people participating, in part because their interest role on, but also from the standpoint of these individual public companies, because I don't have to roll it up into my income statement if I control less than 50% of it. So that's a really important provisor that I want de facto control because I'm actually using it and benefiting from it. But from a legal standpoint, I don't want legal control because then that flows back and I have to deal with it from the standpoint of my credit rating in terms of the leverage, the amount of debt I have on my balance sheet

against my equity and all of these things that are really mundane and boring, but matter immensely to CFOs. So from their standpoint, the idea of partnering with a large private credit firm to create these bespoke special purpose vehicles that are these one-off vehicles that we all sign up for and we join in and the data center gets created, they're great because now I get what I wanted, which is a new data center, and I don't get what I didn't want, which is a hit to my credit rating. And so there's a huge incentive to create these and create more, especially given that we're already at on these historical limits in terms of the amount of spending we're already making. So there's a huge incentive to put it somewhere else. Let me try to restate this.

So I understand it. So meta wants to build these gigantic AIA data centers. These projects cost tens of billions of dollars altogether. Even though meta's rich, they don't want to just borrow all the money the normal way. They don't want the spending necessarily on their balance sheets. So they solve the problem by creating like this special box, as I put it. Meta put some assets into the box. Another private investor puts some money in the box.

And now that box, that special purpose vehicle, is going out, borrowing money, paying for construction and owning the data center, right? On the surface, I guess, you could say everyone's happy, right? Meta gets money without messing up its balance sheet. The investors get high returns, I guess, without obvious risk because they're basically working with meta. But what happens if we build so many data centers, right? meta's exposed.

The private equity firms, maybe more importantly, are exposed. Maybe some of these REITs are exposed. And that means the limited partners, the LPs, whoever's putting the money into those private equity firms, they're exposed as well. So the same way that, like, if you were going back to 2006, 2007, we were thinking, if this whole house of cards comes down, who's hurt? You could say, oh, it's Bear Stearns. Oh, it's AIG.

Like, give me a sense of the kind of companies that would be most. most exposed if we saw a significant slowdown in the AI CAP-X world or some kind of significant pullback here? Yeah, so not to go full hedge fund, but if you think about it in terms of the companies that have really benefited, they're in construction, they're in air conditioning. So like carrier, for example, and think about them as being a, if I'm building out industrial class air conditioning for giant data centers

at an unprecedented scale, what business would I really like to be in? I want to be selling them air conditioning because cooling those buildings is a huge problem. And it's great if I'm an industrial provider of industrial air conditioning, for example. So these are the kinds of companies that aren't as obvious, but are huge beneficiaries of this buildout. So leaving aside you and I as beneficiaries from AI, think about the construction providers, architectural providers. Think about the carriers of this world and the air conditioning providers. And all of these people who have to, whose products end up helping turn a piece of real estate into a functioning shell into which I can insert GPUs. And all of them are delighted to be participating in this.

But then there's the perverse fact that like you as an individual investor might say, well, you know what? Let them all burn. If this goes bad, it's their problem. Carrier or these private equity firms or private credit firms or even meta and whoever else. But yet that it's not going to work out that way. reason why it's not going to work out that way is in part because it's driving economic growth. And that's, we've talked about that a little bit. But it's also because increasingly, these investments in special purpose vehicles and other related data center spending is showing

up inside of things like REITs. So real estate income trust. So if I'm creating a, if you look inside at any large REIT in the United States today, somewhere between 10 and 22 percent of it is already directly data center related. So if you're a conservative investor with a reed in your portfolio because you're saying, you know what, I don't care about any of that crazy tech stuff. I'm going to be over here safe as houses commercial real estate or whatever else getting real estate income. Go have a look inside your read. See what's actually in there today. Two years ago, there was nothing in there that was related to data centers. In some of the largest ones today were up to 21, little 21, 22 percent

is directly data center related. So you're soaking in it. You're already in there, my friend. And then putting this all together, right? Let's say you're a typical, you're an older investor, you're a conservative investor, you've got some money in reeds. You think this is just sort of your meat and potatoes investment vehicle. And now you said between 10 and 20 percent of these reeds are directly or indirectly tied to data centers. 10 or 20 percent of their assets under management. So all read, pretty much all the reads, but 10 or 20 percent of their assets under management. Assets under management.

Okay, so you're right. So 10 and 20 percent of these reads assets under management is in data centers. You told me 25 minutes ago or whatever that data center costs are like 70% GPUs, which means in effect that these reeds are basically just like significantly in Nvidia. Right. Like, I mean, these grandparents who like don't even know that they're like invidi investors are like significant investors in Nvidia, which means as Nvidia goes, so do their investment portfolios. I mean, that also seems like a significant part of this, which is that like, you know, you've got this enormous U.S. economy, $35 trillion, and it's like a significant amount of its growth on a quarter to a quarter basis, whether it's equities or GDP, like balances on the narrow read of like,

how's Nvidia doing? It really, it just seems like an enormous share of economic growth right now is like basically, how are we doing with chip sales? Yeah, absolutely. And it's, it's not funny, but it is kind of funny. Imagine you got scared. I'm a risk-averse investor, and I said, you know what? I'm only in index funds. And so a year or two ago, someone told you,

you know what? You may think you're being risk-averse, but 30% of the S&P 500 is now tied to what's euphemistically called the Mag 7 stocks. You're actually long Nvidia. You're a long-in-vite in a huge way. You're like, oh, I'm getting out of the S&P 500.

I'm going only into really safe stuff like REITs. So now it's sort of this problem. There's nowhere to run. It's increasingly, the case that you've got nowhere to run. And in a backdoor kind of way, private credit now is now allowed inside of retirement funds. You're seeing increasingly these showing up in other ways, not just as REITs, but let's say I'm an investor in private credit, thinking that as a retail investor, I'm now investing in, I don't know, take private operations for a manufacturer in Iowa.

No, you're not. You're in data centers. And by proxy, by being in data centers, you're also in invidia. So this notion of, it's a complex system, but there is a single point of failure. And this single point of failure is a couple of semiconductor stocks who are highly leveraged to everything that's going on and yet have kind of metastasized across each of these pieces from the S&P 500, to REITs, to REITs, to back during their way into new private credit ETFs. It's incredibly insidious and important, and yet most people haven't even realized how deeply it's insinuated itself. So going further along this particular train of thought, what does a bubble look like to you? What are the news headlines?

So I think the news headlines are, for starters, it would be the largest share of future building in terms of data centers is all through SPVs. So for me, it's people saying, oh, look, it's now all being done in partnerships. It's not as risky for meta. It's not as risky for Amazon. Look, they're partnering. For me, that would be the hallmark of a bubble that's hitting the point. of, okay, we need to really be paying close attention because the companies themselves are stepping away so aggressively because they see the effect this might have on them. And the other thing to watch

for is delays in terms of the provision of air conditioning and other of these ancillary equipment that's incredibly important. Interconnect gear for interconnecting racks and GPUs inside of these centers. Delays at one point we're going out to four and five months. If that continues, If that continues to come back, I'd be watching it because now it's the reverse phenomenon where it's like, oh, wait, if I can actually get stuff, that must mean things are slowing down. So these are things to watch for if you suddenly hear about, you know, I go to any industrial class air conditioner suddenly missing their numbers. Well, the only reason they're going to miss their numbers today is because they don't have

data centers to sell to. That's the only reason because otherwise they're going to blow the doors off from now into eternity. So these are some, it's these things at the edges that you need to watch, as opposed to saying, my brother-in-law, I try to open AI and doesn't like it anymore. Or even to go back to your original point, yes, AI is stealing jobs probably and people are increasingly being feel threatened, especially in white-colored jobs, especially in areas like software, maybe in law and so else. But the bigger risk remains this great sucking sound of capital being pulled out of small manufacturers who might otherwise be onshoreing and employing people and are now forced to say,

wait a minute, I can't compete against this. So this goes back to a point I make all the time about this stuff, which is that this is how you end up making bad policy decisions. So if you say to yourself, oh, wait, people say the economy's weak, it grew 3% in the second quarter, it's not weak at all. Well, yeah, but you're not factoring in how much of that was tied to data centers and how transient that spending is. And I make this stupid joke all the time, but I'll make it one more time,

which is that having messed up causality in terms of understanding, the causal nature of what's going on. It's a little bit like my dog. He barks every time the mailman comes to the house. And then he keeps barking and the mailman goes away. And he's like, dude, I totally have this. If I bark long enough, the mailman goes away. No, no, no, the mailman goes away every time.

It doesn't matter how long you bark. So the dog's notion of causality is completely wrong. We're like that barking dog in terms of understanding the drivers of economic growth right now. We think it's because of tariffs. We think it's because of all of these other factors. and it's not. So there's this perverse incentive to keep doing the wrong things because, look, they're working, and they only are working because no one's going down deep enough to understand, wait a minute, it's being driven by completely different things.

The economic commentator Noah Smith wrote a piece about what it would look like if a data center slowed down became a true financial crisis. And he put it this way, and I would just love to hear you evaluate this particular logic. He said, you know, number one, we've got this big story about how this time is different, that AI is going to be the technology to overtake all technologies. Number two, we've got a large and increasing amount of debt being used to fund one single sector, and that means that the loan's probability of default is highly correlated. If one loan defaults, it means there's probably others that are going to default as well. We've got an opaque corner, as you've said, of the financial system with private credit that's grown a lot. And finally, we have systemically important

lenders, banks and even insurance companies, I believe life insurance companies in particular, are significant LPs to some of these private credit firms you've talked about, and they're enmeshed in this new sector that might see a drawback in the near future. To an extent, do you think that this represents the ingredients for an actual financial crisis? Oh, absolutely does. It has all the pieces. So I'll pick on just one that you mentioned, and I'll just flip it slightly, which is that the connection, for example, to insurance is very poorly understood. So what's happened over the last few years, it's not so much that insurance companies are large LPs in these private credit providers, meaning that they're large investors in them.

It's the other way around. So what's happened is private equity and private credit have purchased insurance companies and they call it the term of artists. They call it a captive source of capital, which is to say the premiums get reinvested and in what the private equity firm is doing. And I can use those assets in turn to invest in. data centers or whatever else I choose to do. But what we have, and this goes back to the, to the time of Bear Stearns and the financial crisis, we have a classic temporal mismatch,

a timing mismatch in terms of when the debt comes due and when I have to make my payments and when I provide things, right? You can see how the data centers are relatively transient, but on the other side, I've got these obligations to my insurance policyholders, which are longer term. So I have mismatched assets and liabilities, wildly mismatched on a temporal, on a time horizon, no different than what happened way back during the financial crisis, which did in Bear Stearns, which was they had lent long and owed short, right? And so they ended up blowing up on that basis. So you can see how the same thing would happen through a back door that doesn't look like it has anything to do with data centers. And it's because the nature of the funding of the providers

of this debt, private credit firms, is increasingly tied to a sector whose obligations don't match the returns from the data centers. And that specifically is. insurance firms, which are increasingly owned by private equity firms. And that's not nearly well understood enough, that the nature of the capital structure in the economy that's driving this has changed. And that's created a new source of risk because of this temporal mismatch. What's the most likely way that you're wrong, or that we're wrong, that like the case for the bubble has some error in it, right? Like, I could imagine, like, if, maybe if, like, Michael Sembalist was on this call, he'd say, look,

these companies have more free cash flow than any group of companies in the history of modern capitalism. They can withstand enormous, enormous amounts of infrastructure spending for years and years and they'll still be highly profitable because they're fundamental business models, whether it's ad sales or whatever collection of businesses

you could say Microsoft is in. They can withstand an enormous hit to their balance sheet. They arewithstanding it right now. That's number one. And number two, maybe this technology is closer to a breakthrough that will yield significant income than you and I think at the moment, right? Like right now, when I think about like how Open AI makes money, you know, they've made money from subscriptions, they make money from their business relationships, but maybe they're on

the cusp of something that's like about to become like a $100 billion annual business, in which case, of course, that's going to pay for all of their investment in training and inference. What's the most likely way that you're wrong? So I've had this discussion with Michael Sembrus. So I'll tell you. Tell me how I mischaracterized his argument. So here's what here. The nut of the discussion we had was about this difference between what we're earning right now,

what a data center earns on renting out GPUs versus what its costs are. So let's say I can rent for $35 an hour and it's costing me $12 an hour, the combination of air conditioning power and the net of my debt on this. That's a $23 per hour gap. So let's say that gets halved over the next two years. That's still a huge premium over the cost, the rental cost of these data centers. So as much as you might say the capital flowing into this is going to cause a big hit, it's still very competitive as a commercial real estate play, if you will, in terms of the

amount of premium I'm earning on top of my costs. And that's a very, that's a perfectly sound argument. You could make that argument and say that, yeah, we've earned, we used to earn, I don't know, $25 or $30 an hour of straight margin on top of the cost of these data centers from a rental standpoint. And that's going to get cut in half. But, bro, that's still a lot of money. Right. So that's the argument you will get from many on the other side of this, that even the sharp decline in the margins on the rentals of these GPU assets still doesn't affect the amount that I'm going to get back. Now, the problem with that is, is it doesn't get to the question of, okay, fine, where's that money coming from?

Where's the money coming from that's the rental? The rental price is coming from somewhere. Most people I talk to are not, most people probably you know as well, typical consumers are not paying for chat TPT. Some enterprises are and others, and Chad TPT's built a nice little business on it, but they're still going to burn. What was it I lost? So, like $100 billion over the next two years, I think was the number. And so the deeper problem is there's a great subsidy going on right here. So the data center rental income is coming from people whose economic models currently don't

work and they show no sign of it working in the near future. So yeah, they're continuing to earn margin as a data center provider because of the monies that they're being spent. But that still reflects a massive subsidy to the people who are paying the data centers even at half off prices. And so for me, that's the way I'd be wrong. And the way I'd be wrong is that margin doesn't continue to decline, that even though it declines, it doesn't decline back to the point where it's no longer economically viable, given rising costs of operating a data center and declining costs or declining prices of being able to rent them, that those two don't come into line. I think they're going to come into line and it's going to become a deadweight loss business.

The argument from the other side is no, it won't. Cost will continue to improve. the providers of these services will find high margin businesses that will support those rental prices, the de facto subsidy from private credit and venture capital and others that allow these prices to stay so people continue to pay. That's not going to go away. That's the argument from the other side. Based on the math that you're describing, when is it reasonable to think that some kind of break could appear in the system? You can't look at it right now and say, oh, like things are starting to break. It seems like it's basically status quo. But when you look at the

amount of free cash flow that these biggest companies have, and they're spending levels on infrastructure, and the fact that they're going to have to buy essentially each generation of GPUs every two to three years in order to stay up to the frontier, when does the math stop making sense for you? So I'm not, I'll start with a naive projection. So a naive projection given current decline rates would put you into about two and a half years out where they're no longer earning a risk-adjusted return commensurate with the cost of doing business as a data center, meaning that it no longer becomes a productive investment because I'm not earning enough in rental prices to compensate for the cost of doing all of the things you just described, not just building the data center,

but having built the data center continuing to refurbish, continue to rebuild the GPU constellation inside of this data center footprint. So a naive projection puts you out like two and a half years. It's hard to imagine, absent something changing materially with respect to the amount that companies are earning from selling AI services, that it doesn't happen even faster than that. So my guess is two to two and a half years. We'll suddenly see these two numbers come very close into alignment. And that's when you really have to be saying to yourself, I don't see how this can continue at that point. The subsidy goes away. Data Center construction stop. Something in these moving people.

has to stop at that point because if you can't earn on that rate, a competitive rate of return is no different than having an office building whose costs are higher than the amount I'm able to rent it out to to my favorite law firm. And not to blend subject matter here, but two years from now, you're going to have the debates in the Democratic primary for president, and two and a half years from now you are in the middle of the 2028 election cycle. So I think the interaction effect between an AI bubble beginning to pop, and the 2028 election could certainly put us in line for quite a chaotic news cycle in 28. I want to end on a positive note because I think you, like I, do see this technology as akin

to the railroads and broadband in that it's almost certainly in a bubble dynamic now, but predicting that something is a bubble is not the same as predicting that it will have no effect in the world. I think that this is probably going to end up being conclusively and definitively the most important technology of, let's say, the 2030s. Where are you most bullish on or interested in the application of AI at the moment? So I wrote a piece recently related to this

where I said, I think we went, chat GPT is kind of the original sin, meaning that we as humans get sucked into things that sound like us. Dating back to Eliza, the original fake psychologist online where you could ask good questions and it would say, why do you feel that way? Chat TPT is that to a different level of very similar to. It sounds so much like us that we wanted to be, it wanted to be our friends. So the most interesting applications of these tools and of this whole

idea of large language models and predictive next tokens and all the technical gables, is at a deeper level, think about I'm a small manufacturer and I'm trying to bring on a bunch of new suppliers. All of them have different systems. I got a person who sits in the back office and tries to say, whenever they say zip code, they don't always have a dash. What do I do with that? These are all things that people do that not only is it a job, but it prevents a more competitive landscape emerging. I don't want to have 20 providers because it's too hard to bring on new providers. If you think about the ways that companies communicate with each other at a very low and boring level, it's kind of like English, French, Spanish. It's the things that large language models are good at.

They know the grammar. They can predict the next token. They say, yeah, yeah, yeah, that's a zip code. I know it doesn't look like a zip code, but that's a zip code. All these kinds of very mundane. So think about that as one of the areas that I think there's huge promise. I'm actually thinking that this is an area where we've got a company that we've been looking at in this area as an investment, but it's just broadly makes the most sense that these are all languages that these models can handle really effectively. And it's at that level. We got sucked into the idea of they sound so much like us. This must be super important. That's kind of a dead end because you can only, as you watch public companies increasingly pulling back and saying, oh, we added a bunch of chat stuff to all of our products last quarter and then analysts asked how that's doing. And yeah, not so much. People don't want chat added to everything. That's just some crazy bipedal ape thing where you want to talk to everything around us and then realize we don't actually want to talk to everything around us.

The interesting stuff is all deep under the hood, messy, boring, the language of business talking to each other in this kind of mundane business of, is that a zip code or not? And that's the stuff that this stuff is tremendous at. And it is really transformational. I can now have 20 different suppliers of that widget, not just two. And that's great for me as a provider of widgets. It's great for the economy. It's great for individuals.

It's super important and no one docs. Paul Kodroski. Thank you very much. Yeah, thanks, Derek.