Tech Brew Ride Home - (BNS) Could AI Spending Blow Up The Economy? With Paul Kedrosky

Episode Date: August 9, 2025

I spoke with the great Paul Kedrosky to discuss the significant impact of AI capital expenditure (CapEx) on the economy, exploring how it contributes to GDP growth and the implications of this spending. Qw delve into the rapid growth of AI-related investments, the short lifespan of data centers, and the potential risks associated with this economic phenomenon. 00:00 The Impact of AI Capital Expenditure on the Economy 09:34 The Dynamics of Data Center Investments 19:39 Debt Financing and Its Implications 30:09 Potential Risks and Future Outlook for AI Investments Articles mentioned on this episode: Paul Kedrosky: Honey AI Capex Ate The Economy Chris Mims in the WSJ Noah Smith Learn more about your ad choices. Visit megaphone.fm/adchoices

It's peak pollination season, and my business is scaling fast. To keep the nectar flowing, I need a phone plan with top priority data speed. That's why I chose GoogleFi Wireless. My connections stay strong even when the hive is buzzing. Plus, unlimited plans start at $35 a month. Now, that's a deal that doesn't stay. Explore Google Fi Wireless plans today. Plus taxes and government fees.

GoogleFi Wireless is not subject to data traffic deprioritization during times of high network usage. Welcome to another bonus episode, a weekend episode, but the first episode of the newly recristened Tech Brew Ride Home. So this is the first Tech Brew Ride Home bonus episode. I'm Brian McCullough. As always, we have somebody that should have been on the show many times for years. I don't know how we haven't spoken before, but we're going to talk to Paul Kedrusky. I've started reading him at Infectious Greed. blog that he ran for years. You've seen it on CNBC. He's a seed stage investor. Paul, thanks for coming on the show. Yeah, sure. Thanks for the invite, Brian. So you lit a fire. You wrote a piece that got a thousand other people to write pieces. I'm going to link to it in the show notes. But basically, the title was Honey, AI CapEx is eating the economy. And let me start with the data point that is in your piece that first caught my attention, which is that essentially AI CAPX spending by big tech platforms, but just, you know, data center spending in general,

might be contributing half a percentage point to GDP growth or more. I can't remember if that was Neil Dutta's number or your number. But basically, as as, as, you know, Chris Mims pointed out in his piece, CAP-X spending for AI might have contributed more to growth in the U.S. economy in the past two quarters than all of consumer spending? Yeah. So, all right, these are insane numbers. So what led you to poke around at these numbers? And also, why were you the first to do this?

Because this seems to be pretty important to our economy. Yeah, it was a mystery to me why no one had gone a little deeper into the numbers, because the, I mean, we can get to dive deeper into them, but just that's super, sort of a face value. The AI capital expenditure numbers, numbers are huge and growing and have been growing for a number of years and consecutive quarters. And it struck me when I was looking at Q1. First, it happened when Q1,

when I was looking at the first quarter of this year and looking at CAPX in that quarter in the context of a, well, actually a negative half percent of GDP be gross. So actually a contraction, I was curious at the time. And I said, well, how much was what was contributing to this? And then in particular, I got thinking, well, hang on a second. There's this anomalous spending going on in the quarter. What role did it play? And I realized in the first quarters, which is when I started looking at this, that it might have contributed half a percent or less in that quarter. And then I took it forward to Q2 and said, okay, well, how much? And there's all kinds of ways to get at this.

not to go completely wonky, but you can start, you can go kind of top down and bottoms up, you can start dealing with multipliers, but let's just start at a very, very basic level and say, how much did the big four players spend on AI cap expenditures, usually building data centers, usually, in the second quarter, and then look at that in the context of overall US GDP and of US GDP growth. So then you can start off, and that really is a floor as you start thinking about this. And so that's what got me into it, was thinking, well, okay, if it was material in the first quarter, well, the second quarter, it was even more spending. It has to be more material. So let's actually walk through this and try to, in a bunch of different ways, do the math. And a really conservative number, just based on the big four, you know, meta, Google, Amazon, and Microsoft.

And Microsoft, just based on their spending, what do we get to in terms of an annualized figure, were we to map it against GDP? growth and against GDP in the quarter. And as you said at the top, you get really quickly to as much, so it was 3% real GDP growth, which is obviously could be adjusted any which way from tomorrow. But let's just take it as gospel that it's 3% GDP growth in the quarter. So the annualizing the spending in the quarter gets to do about 300 billion, 320 billion or something like this in capital expenditure. And the nice thing with that number is that these are public companies. They're disclosing this in part because they have to, in part, because there's a one-upsmanship competition going on that they all want to demonstrate. They're spending

more than the next guy. So the bottom line is we have decent figures on something that's often mysterious. This is what's sort of unusual here is that we have decent numbers on something mysterious. And I can map that in an annualized basis against GDP, and in particular against GDP growth. And you get to about 0.7 out of the 3% GDP growth in the quarter might have been, very conservatively might have been AI CAPEX related, or certainly was, I should say, AI CAPEX related. But what gets interesting is if you start to become a little

more wonky about it, because this isn't how economics works. It doesn't stop when they spend the money on the data center. It's not like all the monies flow to Nvidia and then it stops. It flows to architects. It flows to engineers. It flows to site planners. It flows to the spouses of these people who then go out and buy things.

It gets recirculated and it grows via debt. We have companies like CoreWeave who aren't reflected directly in these numbers. because they're using more of a debt financing structure as our private equity companies like Apollo and others. And so this becomes a floor. And it's really quickly you can get to numbers like probably closer. My view is it was probably closer to half of growth in the quarter when you apply a 2x multiplier and include all of the ancillary spending and circulating in the economy. At least 1.5 of the 3% in Q2 was AI CAPEX related, which is,

almost unprecedented. There are very few examples in U.S. history where a single, a single industry category contributed so much to GDP growth. And so to go circle back to your original question, that made me stand up and say, okay, why is no one writing about this? This is insane. And not insane in the sense that, oh, my God, it's all going to end tomorrow, or is it a bubble, or is it not a bubble, just straight up in the sense that we don't understand how our economy is working. Our economy is leveraged to something really unusual and no one's talking about it. And the analogy I make all the time and then I'll shut up is that it's a little bit like when you get causality wrong in something important, it causes you to make bad decisions. So the analogy I always make is my dogs bark at the mailman all the time.

And every time they bark at the mailman, he goes away. So they think, oh, great, my barking made the mailman go away. No, the mailman goes away every time, whether you bark or not. That's just his job. He goes away. So it's the exact same thing in the context of the economy, where if you don't understand the drivers of GDP growth, you're likely to think. think that almost anything you're doing is causing that growth. The current administration might think it's tariffs. They might think it was more Americans working because we've had water deportations

from ICE in the context of the current immigration things going on. So whenever you don't understand the drivers of growth, you can on one hand make up whatever reasons you want, but on the other hand, it's almost certain that the policy decisions you make and the levers you pull are the wrong ones. And so that's the reason why this actually matters. It's not just a freak show thing, because there's a tendency to just say, oh, big number, isn't that fun? And my view is that if you don't understand the levers of growth, you make terrible decisions and you need to understand how your economy works. And that's why one of the reasons when I started was that I saw this headline in the Wall Street Journal that said,

the weirdest quarter ever. And I was like, as soon as people start saying that something at big and complex, like an economy that they at least superficially think they understand the drivers of, isn't working the way they expected, you should always pay attention. Right. Because, I mean, number one, I think the reason the struck a court is the correct way to think of this, that we've essentially got this huge non-governmental stimulus that may or may not be propping up the economy right now. Like maybe the economy would look worse if this wasn't happening. There's no question. So if you take it, even at the conservative figures, at 0.6 of the 3% of GDP growth, we would be sitting at the analysis. The analysis. or the context to put it in is that the U.S. economy is fundamentally predicated on having at least 2% GDP growth. Absent 2% GDP growth, it gets much harder for us to service debt. It gets much harder for us to provide social services because the obligations start to grow faster than our ability to pay them.

And that's leaving aside whether or not deficits are growing. But just in a static model, it's really important that you understand why we need to have levels of growth 2% and higher. because of these fixed obligations. And so as soon as you discover that, wait a minute, this is a transient phenomenon. This is not a reflection of a healthy economy. This is a reflection of an anomalous amount of spending going on in one particular sector. Well, also one thing to note, even though we're using all these big numbers, is kind of how recent this is. Like, according to Brian Saussi, like 10 years ago, CAPEX spending by the tech giants was like a tenth of one percent of GDP, and now it's approaching 1%. So that's that's 10. 10x over a decade. But it's also basically doubled in just the last four years from where it was

four years ago. And we are talking about numbers in the hundreds of billions of dollars a quarter at this point. And that's just the influence from the big tech platforms themselves. So the point is that this is a firehose of money that kind of just got turned on. They just got turned on and that there's all kinds of people playing in different corners of this who are directing more firehouses of capital at this. So it's not just Microsoft, Amazon, Google Meta. We also have private equity companies coming into this and saying, I will finance a shell, what's called a powered shell, and then we'll try to tenant it, almost like an MTAU, like a multi-tenant apartment building, and we'll try to tenant it with other users who will then want to rent access.

So they feel very clever because they've created this structure that can generate, they hope, perpetual income that's largely debt financed. And they think that the rental income will exceed, and this is more sort of finance wonkiness, but will exceed their weighted average cost of capital. So as long as the rental income exceeds the weighted average cost of capital, you build data centers. You build data centers forever because the math tells you to do it, because you can earn a higher return than your cost of capital adjusted for risk.

And that's the calculus that people like, well, I just saw another story today, but Apollo moving into this area and starting to build data centers. And it's remarkable. and it's all driven by the perception that the income from building these things in their models, especially these powered shells where they don't actually even get involved with buying GPUs, are going to exceed the weighted average cost of capital. And that's a really striking moment. And we can get into why.

But the notion that this fire hose of capital is coming from so many directions at once is also unusual. Yeah, I want to come back to especially the debt angle of that, but, you know, where all the investment is coming from right now. But there's one more data point that I think is useful to frame this for folks. And people that know me know that I love data points from history. So you're pegging AI data center spending as being around 1.2% of GDP right now. And then you estimate that at the peak of the railroad boom in the 1800s, railroad CAPX was like 6% of G&GV. So we're not there yet.

And you can answer if you want if you think we're heading in that direction. But at the height of the dot-com era, telecom boom, although I think your number was the year 2020, the telecom boom after 5G. Anyway, the point is, people forget that it wasn't just the dot-com boom. There was also a coincident boom of fiber billout. And that maybe was about 1% of GDP. So we're at that level now. Right.

Right. And that context is incredibly important because there's a big different. When you think about those levels of buildouts, let's take railways and fiber as an example, the two historical examples, they are unusual and in a really important way compared to what we're doing now. If you think about it, the hallmark of the fiber buildout was we didn't light a lot of that fiber for years. The hallmark of the railway buildout was we didn't use, at least in a heavy way, a lot of those rail lines for years. They didn't necessarily sit empty, but they never carried the amount of carriage that people thought they would initially. But it didn't matter. And the reason why it didn't matter was these are long-lasting assets. So the difference this time is the mean life of a data center's GPU install is about three years.

So if I don't, quote, light the GPUs within three years, I have to rip out everything in my powered shell and replace it again. So this is very different in terms of, if you think about it, it is almost a perishable good. It's very different from what we spent previously with respect to these waves of KAPX. Okay, let me pause here and underline that for a second. one of the hallmarks of the rebirth of tech in the Web 2.0 era was the fact that there was all this dark fiber that you could get for pennies on the dollar. So if you're a Mark Zuckerberg and all of a sudden, you have to do data centers all over the place and your data intensive startup, you can get it for cheap. And there's a whole generation of startups that could get that sort of stuff for cheap. But okay, go into a little greater detail for me about why data centers are. So fibers, sort of future-proof. Rails are sort of future-proof. Once they're laid, it's not, I mean, maybe 20, 30 years down the line, there's a generational change to detect, but is it, is the problem with these current data centers, the fact that three years from now, Morsla or whatever, like,

the actual compute moves so fast that it's out of date almost as soon as you built it? That's right. And so the models that most people were working from who are doing actual data center construction suggests three-year lifespans. In terms of the like, period over which you will have to earn a sufficient income on those GPUs via rentals to justify replacing the beginning of that cycle. So there's a double bet going on. So you're saying, I'm sorry to interrupt, you're saying that the key here is to make back your investment on the chips. Because someone said to me recently, they're like, yeah, but once you make the put the roof over the thing and once you have all the racks in place or whatever, then you're just swapping out

chips. But what you're saying is the key economic thing here is return on investment for putting those chips in. Right. Because of the if you structure the costs of building, let's say the average new data center is 100,000 square feet, around two acres, something along those size. About 70% of the cost of construction is the GPUs. So fine, you give me a powered shell, but I still have to put in the GPUs that I'm going to actually use to justify my rental ROI transaction or my calculation. And so I don't get to then just in perpetuity earn anomalous income from this powered data center because I put a roof attached water and power supplies.

70% of my cost is still reflected in the replacement cost of the GPUs. And the next generation of GPUs, while more powerful, is not going to be cheaper, which I go back to the same problem then, that I still have to earn a reasonable rate of return on top of my weighted average cost of capital, which is a blend of debt and equity in most cases. And so to your point, there still is this very short and anomalous window in which I have to earn my return on a perishable asset. Think of it like, I don't know, bananas with a three-year lifespan. Yeah. And unlike railroads and unlike fiber, both of those had to be maintained,

but the notion of a useful railroad didn't go away because we hadn't used a line in five years. We might have had to pull weeds. Similarly with fiber, it didn't go away and become less useful because I haven't lit it in two years. We could light it and have Netflix, send all kinds of things across it. Just real quick, do the economics work right now for that rental and ROI model? They do. But it's in decline really sharply. So a year ago, you were seeing roughly 50% return on about a 14% weighted average cost

of capital. So that was huge. It was basically the market telling you build as many data centers as you can. That is in sharp decline. It's now closer to 22 and 14. So if you start thinking about it in terms of risk and illiquidity, that's a very, very, very, very, that's actually a relatively tight margin. And now in the commercial real estate marketplace, if you're buying it purely on the basis of income as opposed to construction, you're actually

paying less than the cost of capitals, meaning that I'm actually being, people are willing bidding for assets at five and six percent return, knowing that the cost of capital is closer to 12. So we're already, at least in the resale marketplace, not the construction marketplace, underneath the cost of capital. And that's even though, as far as we're aware, the demand for compute is not leveling off. That's right. But the demand for compute isn't leveling off, but neither is the demand for new construction in data centers. And so those have stayed far enough ahead, and there's such a land grab going on that people are pricing. So if you think about what the floor price is on a

transaction, it's a combination of things, right? So meaning a transaction like a rental transaction in terms of GPUs, hours of GPU usage. It's probably about what I paid, but no one cares that much about what you paid. It's probably about my cost of capital, but it's also about my baseline power and maintenance costs. So if you think about that as the absolute floor, let's say I can do nothing else, so keep the lights on and the water flowing. We're not that far away. We're about 80%, 70% above it at the lowest, at least with open source models being provisioned over those kinds of rental transactions. That's not enough to justify the kinds of capital flowing in here. So we're rapidly hitting the point where we're selling it.

if you want to think about it in widget terms, only marginally about the cost of good sold. So in your piece, you mentioned that all of this cash that's being thrown at this land grab is coming from six sources, which I won't name all six, but primarily thus far the internal cash flows from the big guys, Microsoft, Google, Amazon, and all that, which is ironic because for years everyone was like, how come they're sitting on their hands and not spending this money? And then there's VC, private equity, equity and follow-on offerings. Like, you know, Corweave, you know, raised a lot of capital recently. But also, as you mentioned, Corweave and others have been going to debt issuance recently.

And so this is apparently a rising role. I also found another piece from the economists that said, the hot center of the AI boom is moving from stock markets to debt markets during the first half of this year investment-grade borrowing by tech firms was 70% higher than the first six months of last year. So it's just year over year, 70% increase in debt issuance. That's right. And some of it's obfuscated because what they're doing and Mata just did this in a recent transaction where they're creating SPV special purpose vehicles where they do it in partnership

with a provider of private credit, like a Blue Ridge or someone else. And so in partnership with a private credit provider, they'll create a vehicle that into which they both contribute capital. but then the vehicle nominally controls the data center asset. And the reason why you do that is it lets you get away from having to worry about a bond rating agency saying, oh, my God, the amount of debt on your balance sheet is soaring. So it allows you to keep the debt off your balance sheet and avoid re-rating issues. X did that recently with a transaction where they moved some, where they had provisions without getting too deep.

They had provisions on existing notes that prevented them from going above a certain percentage of debt on the balance sheets. They moved the new financing. into SPVs that are off balance sheet. And so we know this story. Once this stuff becomes obfuscated and opaque, it's much more difficult to manage the risk and see what's actually going on in terms of the debt servicing obligations I have as a company

because I've moved some of this away from my balance sheet and into these SPVs that can be very murky in terms of understanding exactly the obligation. So yes, you're right. The debt component is really the story right now, but in particular how it's moving off balance sheet. and then the secondary thing, how it's moving into some unusual places like real estate income trust. Real estate income trust have become a huge player here because, once again, they think of these things as multi-tended apartment buildings.

They just see them as yield engines. And so there's a huge incentive on Blackstone to both help create the financing structures for new REIT vehicle or for new data centers and then incorporate them in REIT. One of the largest REITs in the world would I think is Blackstone, which is around $55 billion. Last I checked, from zero to three years ago, 18% of the assets under management are now tied to data centers. That's astonishing. So if you're the average conservative, let's say conservative recent retiree dependent on income generating assets and you're holding a REIT because you think it's a relatively conservative investment, it's kind of like you're actually holding Nvidia. Right. Because it's highly levered to the GPU marketplace and in particular to the building of these data centers, which I,

I think should be eye-opening for people. At the very least, you should know and be aware that this is what's going on. Right. And you've been very careful to say that, that you didn't write this as a, oh my God, this bubble is going to blow things up. But, okay, we got to go this way. How could this be bad? Like, if, I mean, just from first principles of if this is having such an outsized impact on the economy, how could this hurt the economy with the obvious thing is maybe first by going away at some point? Yeah. So there's at least four ways. So one is that it goes away and we realize, oh my God, we're sitting on a giant air pocket, right? So it's just clear air turbulence and boom. We're flying at 30,000 feet and now we're flying at 15, right? Because we have GDP growth when you

replace, or sorry, when you take away this component of GDP growth. So that's one way. And that's calamitous. Goes back to my point with respect to debt service in the United States. And the the necessity of having 2% plus real GDP growth and so on. So that goes away what's hugely problematic in terms of the U.S. fiscal obligations. The other way you can get into an interesting problem is for exactly the reasons of the creation of all of these SPVs and related structures. So we're already seeing stacking of these structures analogous to the CDOs and the CDO squared, like we saw back in the financial crisis, which further obfuscates the risks underlying them because, you know, if I roll enough data, centers, incomes into an instrument, how could it possibly go wrong?

Right. Well, we learned how those things can go wrong. So, again, we're seeing some of the same things play out. Another thing to keep an eye on that I think is actually a weird outlier risk, but maybe not so outlier. Think of what's going on is kind of like OPEC in the sense that a bunch of companies, countries in the case of OPEC, all think they have a giant, there's a giant reserve. They have a reserve of oil in the case of OPEC.

These companies are trying to control a giant. reserve of tokens that they can cheaply provision and provide for whatever sorts of services they want to provide internally or rent out. So imagine as things the demand tapers off or put a different way, demand stays at the same level, but the data center construction stays just as high. What do you do as an owner of a data center? You begin, just like OPEC, you begin dumping tokens onto the market. And I don't mean tokens literally. I mean rental access to tokens. So the cost collapses. And when the cost collapses of tokens, suddenly a bunch of things, that didn't seem economically feasible, become more feasible.

Like there's an acceleration in call center employee replacement. There's an acceleration in some of the things that were probably coming along over time. But imagine all of a sudden we're just dumping GPU access onto the market essentially for free. What sorts of things become accelerated? It's much like what happens if suddenly oil becomes free. Studies and play. Come together on a Windows 11 PC. And for a limited time, college students get the best.

of both worlds. Get the Unreal College deal, everything you need, to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 premium and a year of Xbox GamePass Ultimate

with a custom color Xbox wireless controller. Learn more at Windows.com slash student offer. While supplies last, ends June 30th, terms at AKA.m.S. College PC. Ready to soundtrack your summer?

With Red Bull Summer All Day Play, you choose a playlist that fits your summer vibe the best. Are you a festival fanatic, a deep end DJ, a road dog, or a trail mixer? Just add a song to your chosen playlist and put your summer on track. Red Bull Summer All Day Play. Red Bull gives you wings. Visit redbull.com slash bright summer ahead to learn more. See you this summer.

Right. People, you know, when at least in the West and in the U.S., when oil prices are low, we're like, yay, that's great for the economy. But, you know, oil producing countries can go through the entire debt. decades of, you know, bad, bad economies because their underlying core commodity is depressed. The, you mentioned obliquely there, like the idea, I mean, the idea to me of, you know, Sam Altman says that currently, you know, their compute costs are eye watering or eye bleeding or whatever he said. But the concept that you could have a crash, like an oil market crash, where then AI becomes so cheap that it would accelerate this sort of adoption of AI.

And so that would be potentially a job killer to entire sectors. Yeah. Yeah. And I mean, and so as a case in point, Goldman had a report out today or yesterday, Goldman Sachs, and they were talking about the likelihood of replacement in very, you know, the usual story with the likelihood of job replacement or augmentation in various occupations. But so much of their model was predicated. on something that I think is a somewhat slippery notion, which was what are token or what are GPU rental prices going to look like going forward

and ignoring that we're building in a sense, we're building like a massive supply, a Permian basin of GPU rentals. And there's a huge incentive once you've built it. I've got costs associated with this. I'm paying interest. I'm paying all these different costs that I have. And so at the very least I want to earn back

enough to cover power and water, right? Back to this basic idea of what's the minimum amount I can charge to just keep this thing from shutting down. So you're going to see, I think, inevitably, in a sense, this kind of price crash and then dumping of, if you will, GPU rentals or just tokens onto the market with consequences that kind of blow up the model

that Goldman used in terms of saying today, I think they said like over a decade, we might see 7% replacement in a small number of job occupations. occupations. I thought it was remarkably naive of them. Just a quick aside on that, these data center buildouts are not extremely jobs intensive. Like compare it to an Amazon fulfillment center where once you, again, build the walls in the roof or whatever, then for decades, hundreds, thousands of people can work there. But that's not necessarily true with the data centers. No, data centers, the NADESAIMA is if you remember the final scene in Raiders of the Lost Ark, where once you put the box in the warehouse, it's just like,

We leave the place, shut off the lights, and everyone goes. They're like that. They're very much like that. So they are lights out facilities that largely operate on their own and relatively autonomously, at least in the broader sense. There's not thousands of people wandering around sweeping up after God knows what inside of the place. So they're not big employment creators. They're much more like the warehouse at the end of Raiders of the Lost Ark than they are like an Amazon

fulfillment center. And Amazon fulfillment centers in the great scheme of things are increasingly less employment dense places too. Right, right. I mean, that's part of the automation, everything, yes. So Noah Smith's piece on this tried to get into the idea of, you know, are we setting the stage for another like systemic financial crisis? But he kind of was making the point that like in the 2008 crash, it was because there was a ton of debt held by banks in one form or another. And that's not yet the case. Like if everything shut down tomorrow, you wouldn't have like that cascading wave of defaults and things like that.

But are we concerned that things might be going in the direction of, as you said, like things chopped up and put into different debt instruments? No, we 100% should be. That's the real story here. The story is not to be found in the banking system, or at least in the orthodox banking system. It's to be found in the private credit system. So the private credit system is largely supplanted banks in much of this, in much of, well, I was just looking at the staff the other day that we've gone from 80% of mortgages

in the United States being issued by banks 15, 20 years ago, and now it's completely flipped. It's everyone but banks that's issuing mortgages. The exact same phenomenon exists in the context of data centers. The private credit institutions out there, the Apollos, the Blue Ridge, and the others are the fastest movers in this area because they don't have the same obligations with respect to meeting some of the financial security things that we set up after the financial crisis,

and they can go out there and take on huge amounts of debt, issue huge amounts of debt, and be highly levered to a relatively small number of places. So that's the place to look, not in the orthodox banking system. So the way that I would say that you have to watch for this to trickle backwards is through things like REITs, that REITs become, let's say, as an example, become increasingly dominated by data centers. I mean, who wants to own like commercial office space anymore?

It's really difficult to earn a rate of return there. But if you load up my REIT with data centers, think of the incentive as an ETF manager to load up on data centers because I can offset a declining asset with what looks like a gaining asset. And so you start to think about now, then what happens if they're now 25 and 30 percent of READ ETFs are now made up of data centers, and then we see a sharp decline in their ability to generate income. That's a huge problem because now we have a massive income generating asset class that's

no longer delivering. And you could even see a complete collapse in a large ETF as a result. The last few questions, and this first one is drawing directly off what you said, but it's purely speculative and so unfair to you possibly. But in the next, let's call it six to 18 months, is there anything on your radar that you think could be a trigger that would get people to stop spending on this AI bailout and then maybe start the dominoes falling for bad things to happen? It would be something macro. It would be a withdrawal because of something larger. in the world, not something specific to the AI buildout itself, because the momentum behind the buildout and the notion that there needs to be,

there's a land grab that'll be followed by consolidation is too appealing to the incumbents. Like, that's the story people tell is I don't even care that there's 100 people building out data centers or a thousand. Or in the case of China, we had Premier Leiping saying the other day, I wish every city in China didn't think they had to build a data center. And I was like, oh, my goodness. I mean, this is just a staunt. Because that's what they did, if you remember, back in 2000,

it was building apartment buildings. And then you had ghost towns and so on. So we're going through that same phenomenon. So everyone thinks they can build out and then consolidate. And so there's no incentive to hold back right now, because if you think you're going to be a consolidator, you're perfectly happy with people building out frantically. And so in answer to your question,

I think it has to be something macro. And the U.S. is like an infinite monkey engine of macro nonsense right now. So there's lots of ways, the U.S. could blow this up accidentally, which takes us right back to the very beginning, that the U.S. could blow it up completely out of ignorance. If you didn't understand that the reason why GDP growth wasn't as low as it likely was intrinsically, you're likely to make other bad decisions about tariffs, about immigration and about other things because you feel emboldened, because you misunderstood what happened in the last two quarters.

But another way to say this is you're betting that Mark Zuckerberg is not going to drop this AI stuff like a hot potato like he did with the Metaverse. You know what, though? That's a great example. And I was just talking to a friend of mine at a hedge fund about this the other day. For years, he told me, there's no way next year they continue to spend like this on the VR stuff and the Metaverse stuff. And yet he persisted, continued to do it with a less defensible economic rationale than what he has in AI CAPEX. So there's no effing way that they slow spending on something where the rationale seems. even more defensible than it was in VR. So I actually think the reverse that there'll be,

if anything, an acceleration in spending. And Zuck might be a unique case, but also, you know, Google's got to defend the innovator's dilemma of all time, defend for its life. Everybody's motivated to kind of keep spending at this point. Amazon's motivated because the fear is, and we didn't get into this, but it's worth mentioning that historically when one of these big KAPX waves happen, you have to think about it, the money comes from somewhere. So the question is, well, who's not getting money. Well, you can think about it in terms of some manufacturers no longer looks defensible spending on that cap-ex. But it's not just that. It's like AWS Cloud XAI isn't looking very healthy anymore. And that's part of what's going on. So Amazon now has a huge

incentive to spend even faster on AI to defend what looks like a relatively flatlining asset in non-AI-AWS. Right. And their AWS numbers weren't as great as maybe some people had hoped recently. But also, I'm glad you brought that. up because that was in my notes too. You said one of the problems that this could have would be at the same time all this money is being invested in this particular corner of the economy. Things are not being spent in other parts of the economy. Yeah. And that's something that people missed because this was true in the era of the railroads. This was true in the era of the dot-coms. I can make a strong argument that one of the reasons why manufacturing left the U.S. as quickly as it did

It wasn't just that China got most favored nation status and manufacturing moved overseas in the late 1990s and into the 2000s. It's in part because capital dried up for manufacturing because it was diverted into being massive CAPEX on fiber and other things. So there were multiple things going on and the latter one gets overlooked. That had huge consequences because then manufacturing moved overseas. People lost their jobs. It created an entire political movement around it. We've led to this current situation in the U.S. U.S. But in some ways, you can tie that back to capital drying up for manufacturing as a result of a

massive diversion that happened during the first tech boom. Second unfair question to you possibly is just what is what is your brother thinking about AI as an investor, the AI moment. Like how bullish versus wary are you as an investor? So we're very, I'll say, careful. It was obvious. at the beginning to us that there was no point in investing, for example, in base models. It was simply too expensive and it didn't make any sense. There was no point in investing in frontier models for similar reasons. So there was no point in doing any of the baseline work.

And then as we went on, we did a couple of selective things in medicine, for example, where it was obvious that I'll turn it around. One of the things we've been looking for is places where it's obvious that not using AI will be a liability. So it's obvious that in medicine, where now it's a novelty, within five years, you will be in legal peril if you didn't get a second opinion from an AI. There's no question in my mind. So that will represent a reversal where it goes from being a kind of acute extra to being the default mode. And so those are the kinds of, that's the way we think about it, which is, isn't, it seems kind of obvious.

But, you know, where are the places in the economy where it has to become a default for no other reason than say legal peril or, comprehensiveness of a solution and that kind of thing. The other stuff, honestly, bores me to tears like, what's the AI of accounting? I don't get a rat test. By the way, when you say we, you're still at SK Ventures, right? That's right. Yeah, everyone and I. Yeah. Right. Yeah. Okay. Last one, and this isn't so hard. But what do you personally use AI for these days? Ha. That's an interesting question. What do I use? Mostly whenever I screw something up technically and I say, oh my God, here's the error I'm getting in some strange log file online, how do I fix this? It's on that level for the most part. That's how

I use it. It's because my rationale is if you understand the nature of how large language models are trained, then if you give it a corpus of definitive answers, which you can often get in technology, like I see this error in this kind of log file online, that tends to mean that the cause is this. Like, it's really stupid, wonky stuff. But in areas, domains where there are definitive answers, they can be very effective. And so I use it in those contexts using tools like warp. This is sort of an AI terminal product and then directly in some of the others, where understanding that the propensity to generate the most common solution actually is the correct solution, as opposed to just being a homogenizing effect. So in those cases, that's where I use it.

Outside of that, honestly, I don't use it much for anything like not in riding, not in a bunch of the usual places like that. Just because of this, this It's a race to the median, to the mean, right? And so I just, given the nature of the training, you're just generating the statistically most likely outcome. And honestly, that's just boring. So I don't do that. Anything you want to plug. Again, I'm going to link to this piece in the show notes.

His blog, Paul Kedruski.com is where I found it. But anything else is you want to tell us about where to find you? No, that's it these days. Mostly I do things through there just because it seems like the right place to do things. completely withdrawn from social media. I don't do very much on social media at all anymore. I used to do a lot. And I joke, I feel like I'm just, I don't know if you're familiar with the old Kafka story, the hunger artist, but I feel like I'm, I've always felt like on social media, you're having to do more and more extreme things to the point of, quote, death to keep people entertained. And I was like, why am I doing this? I'm out before they replace me with a tiger in the cage. Well, and then look what happened. You did a dumb old blog post like we did 20. years ago and look at look at what happened so yeah who knew yeah hey paul thanks for coming on and and

talking about all that with us sure ambition comes in all shapes and sizes at first citizens bank we roll with your goals because we're built for what you're building fit for your ambition for citizens bank