The Hidden Recession Beneath The AI Bubble w/ Paul Kedrosky | Better Offline

Hey Ed, good to be back.

Yeah. Like it's funny because almost a year ago, one of the things that first got me interested in this bananas topic was that AI capex specifically was, you know, blah blah, more than half of GDP growth, blah blah, and so on, which was just, you know, unprecedented and goes back to the railroads and everything else. And so that was interesting. And I started talking about it and then I was surprised that no one else was sort of noticing this, how consequential it had become.

And now, of course, that's become commonplace. And so then I went on and I've been paying attention to that and that hasn't varied very much from quarter to quarter over the last sort of, I don't know, 12 to 18 months. It's been consistently between 30 and 70% of GDP growth. And then it started, it was interesting how it's almost in cancer terms sort of metastasized across the global economy and become even more consequential. So the next example was about a month ago where all of a sudden we had this treasury market freakout in the 10-year, and one of the reasons driving what happened was this sudden spike in the 10-year up over 5% now.

That's right.

Correct. People required more.

No, no, it's fine. Keep going.

Yeah. So normally the Treasury market, which is a multi-trillion-dollar market, is so large and liquid that unless you are a sovereign, meaning another country issuing debt on the scale and with the same sort of security as the United States, it's really hard to influence it just through issuance. Meaning that just because I'm selling Treasuries, or not Treasuries, I'm selling comparable duration debt, it's really hard to have an impact on Treasuries because it's such a big, large, and unusually liquid market. So one of the things that happened a month ago was that...

this thing we call AI capex and this staggering amount of issuance, much of which has a duration, meaning the length of time over which the debt becomes due, it varies from 5 to 15 years, but let's call it a median duration of 10 years. It began to seem like something unusual was going on in Treasuries. And as we sort of work through the math, and this has now become fairly widely accepted, that AI capex issuance had actually become so large that it was beginning to bleed into the Treasury market. And so one of the reasons why we had this Treasury market freakout is because of the unprecedented amount of AI-related debt being issued. And that in turn was causing a sell-off in Treasuries as people said, you know what, given the choice, and this is a remarkable thing, given the choice between owning 10-year Treasuries backed by the full faith and credit of the United States or owning 10-year debt backed by a hyperscaler's pristine, and I have that in air quotes, credit rating, I'd rather have the latter at the margin. That just doesn't happen. I mean, there's no reason why someone historically would do that because obviously sovereign debt is backed by the full faith and credit of the country, and in particular in the case of the United States, you can print your own currency. So it's like there's no prospect of default other than deflating it away. So having that happen was an artifact of two things. One was the crazy scale of issuance, which is now approaching in excess of a trillion dollars.

And it's more than 60% of AI financing is now debt-financed, up from something like 15 to 20% a year ago, and it's now the largest piece of the investment-grade marketplace. It's now the largest piece of the high-yield marketplace. It is literally taking over global debt markets. And we had this unprecedented phenomenon literally a month ago where it in part caused this global bond freakout in the Treasury market because the scale was so large and the counterparties, the hyperscalers, for crazy reasons were seen as such secure credits that it began to literally compete with the highest-scale sovereign in the world's debt issuance, which is the United States. And that was, I think...

The currency of the world was that AI capex as a currency was beginning to compete with the dollar in a really sort of loose sense. So that was the next one that really caught my attention because that has consequences, because in turn as the cost of financing the US deficit rises, and its existing debt, the cost of servicing it rises, then obviously that has consequences in terms of the United States's fiscal position. And so you're having this crazy bleed over from what we see going on in AI capex into the fiscal health of an entire country and of the entire world. So that is just a remarkable moment. And then it gets crazier, right? And I'll stop on this next one and then we can sort of loop into it in whatever way you want. But just last week, the WTO put out a report saying, 'Whoops, our bad. We completely screwed up with respect to trade data on the goods side, not the services side,' because the WTO, the World Trade Organization, tracks goods and services global trade data, because they had been forecasting a fairly sharp decline in part because of the US tariff policy and the consequences of it, both in terms of uncertainty and changing trade flows, and actually the opposite happened, right? And so it was up something like 4.5% or something like this, which was mean in real terms.

So now they broke down the data because they, you know, like any good economist, they said, 'So, you know, we were wrong.' Why? And it turned out that something like 19% of global trade in goods in the last three quarters was AI-related. That's...

It gets worse because that's Nvidia GPUs.

So that's Nvidia GPUs. And then if you take it the next step, well then how much of growth? So if 19% was AI-related, how much of the growth that was anomalous? Almost 55% of global trade growth was AI-related. So in a sense we've created this global now illusory phenomenon that's being driven by this incredibly anomalous spending, much of which is concentrated on one specific thing, these things we call GPUs, but it's bleeding over into sovereign debt. It's bleeding over into the cost of financing the US deficit. It's bleeding over into world trade. It's masking the effect of the Trump tariffs because it's so consequential. So it's having all of these incredible effects that are rippling across the entire economy. And I'll give you one last one, which is one of the craziest ones. So you probably saw last week the Fed raised rates because US inflation remained stubbornly high in excess of 3%. So let's break that down because you know it seems like, okay fine, consumers must be spending a lot, and of course the answer is that consumers are actually more nervous than they've been in a long time and sit with relatively high levels of debt. And so it's like, well wait a minute, where is this inflation coming from? So if you break it down and sort of tear apart the pieces of this, you find out that the two largest pieces that are driving inflation in the United States over the last year are, well, you'll sort of sing along with me on the chorus, is AI capex, which is driving energy prices higher, which is driving all sorts of externalities across the economy. It's driving all sorts of services higher, construction costs, materials, all of those are being driven higher. And then this is the sneaky one. The other piece is obviously energy in the context of the war in Iran, which is the other big piece, and you can argue, and I did last week at a thing I was doing at Oxford, that in a sense that's also AI-related. And the reason why that's AI-related is the Trump administration was emboldened by its success in Venezuela, which I call the first AI war. Pete Hegseth and others have been very straightforward in saying that that entire campaign was organized and administered using Anthropic, which is one of the reasons why they had a seizure when Anthropic said you can't use this for military purposes. And emboldened by their success using AI to conduct a campaign in Venezuela, they said, what's next? So Iran was next, and they've obviously used AI extensively in conducting that campaign. So it's created a kind of arrogance about what's possible whenever you use AI to backstop all sorts of things. So in a perverse way, now let's go back to this inflation data. Both pieces of inflation have underneath them as predicates this incredible and unprecedented growth in AI. And whether it's the energy side directly and the war in Iran, which is a sort of AI-enabled, it's almost we've almost sort of turned war into an app, and on the other side what's happening in AI capex. So these examples for me are just startling in terms of understanding the breadth and scope of this phenomenon and the consequences across the entire global economy.

That's correct. So that's for on-the-run bonds that have already been issued. So it obviously has a similar consequence for new...

If you're issuing new debt, obviously it has to be at the higher rate. So either way...

Correct.

Yeah. So that's exactly right, and that's in part because the structure of the Treasury market and broadly the US debt market has changed a lot over the last 20 years. It used to be dominated by large sovereign purchasers like Japan and China and others, and they were, in this great economics term, they were relatively yield-insensitive, meaning that they didn't give a rat's ass what the yield was. They just had huge dollar flows coming into their country and they had to do something with them. So they bought Treasuries and it was like, whether the rate's one or five, I don't really care. I just have to do something with all these dollars. And now the Treasury market and broadly the US debt market is much more taken up by yield-sensitive investors, individual consumers, by people buying Treasuries or having them embedded in ETFs, by corporates, by institutions and others, and they care what the yield is. So now you have this phenomenon where people are yield-sensitive and they actually care, and they say, you know what, at the margin, given the choice between owning hyperscaler debt at the same duration with a pristine balance sheet, they argue, or owning Treasuries at the margin, I'd rather own the hyperscaler debt, and up goes the yield on the Treasuries.

Right. So that's the mistake people make, exactly right there.

the causal link here, and they think, oh, this is because people have suddenly lost faith in the US's ability to pay its debt. The US prints dollars. They are not going to default. They may deflate it away, but don't confuse the issue by introducing the idea that somehow because of some madness the US is doing, that its debt is more suspect. This is literally an artifact of competition at the margin from hyperscaler issuance in a newly yield-sensitive market causing people to say, I'd rather own X than Y, and that has sovereign consequences for the US as an increasing fraction of its budget gets eaten up every year by payments on the debt, and in part because of AI issuance.

Social services broadly and defense, you know, Medicare in sort of all of those are the three big buckets, but behind all of those very closely, depending on the math, comes in something like in the mid-teens, interest payments.

Yeah.

No. And it doesn't even, it's a double-locking doom loop. I'll give you the other piece of it. So if you're the government and you say, okay, you know, this has consequences for us in terms of increasing current and future projected debt payments as a result of competing at the margin for these yield-sensitive investors, what do you say to yourself as a solution? Well, we need to grow the economy faster because that'll solve the problem, obviously. And how do I grow the economy? Well, I increase productivity. Well, how do I increase productivity? Well, people tell me this AI stuff is really good. So there's your double loop. So now all of a sudden AI is not just the problem, it's the solution.

So it's literally just output over input. So if we look at how much is being produced per unit labor and capital, and how that's growing. So we have a measure of how much labor and capital are going into the economy and we look at what the output is that we're getting for that in GDP terms. There's your productivity. So it's really just simple...

So the labor measurement goes back to sort of what the size of the labor force, hours worked, all of the BLS aggregate data plays into that. So it's a good question though because there's multiple ways obviously to increase productivity. People can become more productive, but I can also introduce a lot more capital into the economy and replace people and also become much more productive. Right? There's multiple ways to increase the productivity numbers. And so you have to be very careful because people keep citing data showing that the US is becoming more productive, but it doesn't have very much to do with individuals becoming more productive. It's because we're increasingly replacing individuals with capital, and that too shows up as higher productivity in the economy producing more.

No, no, no, it's not. But every once in a while you'll see a blip and of course instantly people seize on it and say, 'Ah, there it is. It's showing up in the economy.' And generally speaking, right now that's almost entirely an artifact of capital spending, not labor. And so it's not showing up on the labor side, it's showing up on the capital side. And you see this all the time. It's classic in behavioral finance terms. It's classic motivated reasoning. I see some data, I have a bias, this proves my point, it must be right. And so that's what's happening. And you know, we see this in healthcare data. There was an Andreessen Horowitz graph making the rounds yesterday trying to show that mortality was sharply decreasing in US hospitals in the ones that were, you know, adopting AI the fastest, which is a ridiculous claim, right? And you...

Well, there's so many issues with this kind of thing, and we can get into it, but it in part has to do with what is AI spending, what are they spending on? Well, they were spending on AI scribes. Well, which hospitals spend the most on AI scribes? Well, the wealthiest hospitals spend the most on AI scribes. Well, what's happening also inside of the wealthiest hospitals? Well, they have the best outcomes and the healthiest patients. There you go. Voila. You just found another way to filter hospitals and say it's really just a proxy for hospitals that have the healthiest patients and the wealthiest balance sheets because they have the largest discretionary budgets. And yeah, right, it's exactly what you would expect to have happen if you found a new variable that filtered hospitals for healthy patients and wealthy hospitals. And so it's this kind of stuff, going back to the other point though, these are all examples of motivated reasoning, and you have to always be very careful when you find a data point that proves something you believe fervently or even religiously because you're so desperate to believe it. And you see this in the productivity data, you saw this in the A16Z health graph I was talking about, I mean it's just sort of everywhere.

Or it's down. AI's...

Seizing on these data points to prove something that you really want to believe. I'm always very suspicious of anything when I feel almost an emotional triumphalism when I see something, I say, 'Aha, this is exactly what I thought.' And I'm like, 'Okay, you know what? Let's cool down a bit here because this probably is an emotional, motivated reasoning response.' Anyway, so across the board, you see this now. So going back to the original point, it's now in the Treasury data, it was originally in GDP data. And so just on the inflation debt, I'll take it one step further which is even crazier. If you back out all the AI capex-related inflation, and if you grant me that energy increasingly is being driven by this campaign that was so AI-enabled and thus AI has a role in it, back those two pieces out of the US economy.

Okay, so there's multiple components to this. So the inflationary component from the standpoint of US consumers is utility prices and gas prices. So let's not get into the embedded price of gasoline or energy in terms of how it plays into the products you purchase, but that obviously is another issue because it shows up in shipping and other places. But the direct one that people feel most consequentially are prices at the pump, whether you're an individual with gas or a trucker with diesel. And then obviously utility prices, both of which have gone sharply higher over the last three years, in part driven by either competition for energy from the standpoint of data centers or the Iran war. So these two things, and I'm arguing that the Iran war is in part a function of the sort of emboldened US government treating everything as kind of an AI-enabled app in the wake of Venezuela. So if you grant me that, then what you find out really quickly is actually the US economy, far from being very strong and actually growing so strongly that it's justified for the Fed to raise rates to try and choke off inflation, back those pieces out and by my math the US is actually in a deflationary mode about a quarter of a percent. So what we have happening is something very analogous to what happened in the run-up to the crash of '29, which is you had the Fed raising rates because it misunderstood what was actually happening in the economy, and it thinks it's choking off inflation, but the inflation is actually something exogenous, something different than what they expect. It's being driven by this capex and by this anomalous war. So as a result, as this unwinds, and it will unwind, the war will end and capex will taper off, you actually have high rates in a weak consumer environment, which is a classic precursor to what... there's an economist named Richard Koo who's written a lot about the Japanese 15-year recession that sort of went on and on and on. People called it the happy recession.

So those are classic precursors to those kinds of recessions because you now have companies that have loaded their balance sheet up and it will take a decade for them to unwind all this debt. So rates will have no impact on them. They'll simply be trying to cut spending to reduce debt, and that will play out over the next 10 years. So we've got all the pieces in place right now for a very long balance sheet recession analogous to what happened in Japan in its last decade.

Yeah.

And the hyperscalers will, yeah, 6 or 7%. They are going to keep raising no matter what.

So what happens is, and leaving aside for a second some of the more, how shall I say, less investment-grade prospects out there, the CoreWeaves and others, because they face a different problem, which is as they try to roll over their debt it may come at terms over the next 5 years that force them into some species of insolvency, and that's fairly typical. But let's say you're a hyperscaler, some of the largest companies out there that don't face solvency risk, they're not going to likely go broke. What they're going to spend the next decade doing is cleaning up their balance sheet because they just went from some of the least indebted companies in the market. Tech historically had very low levels of long-term and short-term debt to some of the most indebted companies in the economy. And we see this as they sort of soak up all their cash flow and in the case of some of these guys increasingly teeter on the edge of having their credit rating downgraded. So what a balance sheet recession means is essentially I'm having to put myself into a kind of debt workout and say, I'm going to spend the next 10 years getting all this debt off my balance sheet again because I don't want to have it there anymore because it's no longer required for ongoing purposes because I'm not spending that heavily in future and I'm being punished for it because I have this huge debt obligation. So think about the consequences. You're not spending on growth, you're not spending on employees, and you're trying to unwind all this debt and get it off your balance sheet. So a balance sheet recession is a workout where heavily indebted companies, regardless of interest rates, continue to try and get themselves less leveraged, get the debt off their balance sheets. And so the problem is of course this is going to come at a time when there's already other pressures on hiring with the AI effects at the margin. So the effects overall on hiring will be even more dramatic because companies will be focused entirely on deleveraging. So at the same time as they're looking for opportunities to use AI to have fewer people around. So this will be really consequential over the next 5 years. And it's really not well understood that we kind of sit on the edge of a Japan moment in terms of the likelihood of a balance sheet recession given that there's this really profound misunderstanding of what the causes are of inflation in the economy and what the effect that rates are going to have, and that we're actually sitting in a kind of deflationary moment masked by what's happening in AI.