AI 驱动的繁荣与泡沫之争

Host: Right now, AI is the driving force behind the gains in the US stock market, behind US economic growth. It is sucking up so many construction resources it now appears to be crowding out housing, manufacturing facilities, like basically everything else you might need to build. And the question, even just from a financial perspective, is is this all worth it? Is all of this investment going to pay off, or is AI a bubble? And one of these companies is going to miss an earnings target, or get downgraded, and the whole thing is going to pop, taking all of our retirement portfolios with it. So, I wanted to have two different perspectives on this.

Host: Dr. Paul Kedrosky is a partner at the venture capital firm SK Ventures. He's a fellow at MIT's initiative on the digital economy, his popular newsletter at paulkedrosky.com, and he's been very vocal that he thinks we are in an AI bubble. Azeem Azhar is a British technology entrepreneur, an author, and an analyst. He is a creator of the Exponential View, which is both a great Substack that I read often, and a research platform where, among other things, they've been creating these dashboards tracking whether or not there's financial stress in the AI industry. And his view is somewhat more optimistic, I think it is fair to say, than Paul's. So, I wanted to put the two of them into conversation together to try to see why you might think, and why you might not think, we are in an AI bubble. Paul Kedrosky, Azeem Azhar, welcome to the show.

Paul Kedrosky: Hey, Azhar, good to be here.

Azeem Azhar: It's good to be here. Thanks.

Host: Paul, make the case to me this is a bubble. I mean, that's a big it's a big claim, right? You're saying that all these investors, all this money pouring in, all of these sophisticated analysts at BlackRock and Goldman and wherever, and you know, the hyperscalers like Google and Meta, that they're all getting this wrong on some level, and you're not. So, why do you think it's a bubble and they're all idiots?

Paul Kedrosky: Well, I wouldn't go quite as far as they're all idiots. That it be true, though. I'll just I'll step back on that one. So, what's unusual about this? What let's just talk about really quickly what a bubble is. A bubble is whenever asset prices detach from any semblance of economic reality. We have one of the largest CapEx competitive CapEx spending waves in history. So, in Western economic history, there've been like eight or nine of these moments where we build out some giant general purpose technology, a GPT, but not not the the ChatGPT kind, but a general purpose technology.

Paul Kedrosky: And so, when we're building it out like we were the railroads, like we were electrification, like we were the canals, like we were the interstate highways, and every one of those moments has led to overbuild. There isn't a major economic CapEx episode in Western economic history that didn't end in a rolling series of crises, vastly inflated asset prices, and vast oversupply of capacity of whatever the thing was, right? Railroads, fiber, take your pick. And so, that's the bubble. The bubble is, once again, we're vastly oversupplying something because of the availability of capital and the eagerness of the suppliers of the capital to continue building out in anticipation of future demand.

Paul Kedrosky: Now, none of this is to say that like AI is frivolous or it doesn't work or that there aren't incredibly useful cases for it. That's not the point anymore than that was the point for railroads, right? Railroads were incredibly useful didn't still ended up where something like 50% of the track miles ever laid turned out to be useless, right? So, the bubble is this vast oversupply of something largely fueled now by credit, which in turn will lead to massive business failures and write down of assets. And all of that credit has spread through the economy. It's inside of insurers, inside of private credit, inside of sovereign wealth funds.

Paul Kedrosky: And we're on the order of, you know, trillions of dollars now, and it's probably going to hit 10 trillion before it's done. So, this all of this investment is at risk and it's across the entire economy and sitting inside of all kinds of but it's even sitting inside of you know your individual you know your 401(k)s, right? I mean we've all seen the data that something like the top 10 stocks in the S&P 500 are like 40% of the market capitalization of the index. So if you're holding an S&P 500 index my friend, you are very very long AI, right? You might

Host: Get that out of my portfolio Paul.

Paul Kedrosky: Yeah, get that crap out of my portfolio. So you're very very long AI and this is the problem is the expression you hear all the time it's a one factor world. Everything is leveraged to AI across every asset class across all the retail investing and then the last stat I'll give you and the reason why that matters is we're at record percentages in terms of the percentage of people's net worth that are tied to equity. 50% of people's net worth is now in equity.

Paul Kedrosky: So follow that math through. If 50% of net worth is in equities and that's driven by the AI trade, people are even more over leveraged to this single trade than they ever were to dot-coms and that creates this incredible fragility. So all of this this this sort of nexus of of factors is what creates this bubblish moment which has nothing to do with whether or not AI is useful but has everything to do with why the other side of this will see you know a rolling series of crises that could potentially become really catastrophic.

Host: I do I want to focus for a minute on the rationality of the core players. Your Metas, your Alphabets, your OpenAIs and Anthropics all of whom believe that there's only going to be a couple AI winners.

Paul Kedrosky: Right. I mean Zuckerberg gave this amazing quote where he said if we end up misspending a couple hundred billion dollars, that would be unfortunate. But the risk of not acting aggressively enough is considerably higher. And so there is this dynamic where I think they believe this is going to be a highly concentrated or oligopolistic market where maybe it's one model that rules them all or three models that rule them all. And so there is this race to be one of the three, but that creates a kind of pressure to invest way beyond what the market can collectively sustain.

Paul Kedrosky: Right. That's the classic competitive dynamic that's existed in every one of these capex build-outs. It existed with the railroads. You could not afford to be left behind. You had to invest in And that's why we ended up with duplicate routes between cities. I mean, that's what That's the essence of the competitive dynamic crossed with a kind of, you know, prisoner's dilemma that I I would like to slow down, but I can't because no one else will.

Paul Kedrosky: And so, we have to we all continue on with this frantic competitive dynamic. And so, that's the what we're seeing incarnate there. see that even whenever, I don't know, a couple of weeks ago, Dario Amodei at Anthropic was suggesting that we need to And this is a ridiculous expression, but we need to pace the frontier, which sounds vaguely like we're training for the Boston Marathon. But it's this idea that we need to slow down. Well, of course, everyone in a in a classic prisoner's dilemma thing said, "Go for [snorts] it. You slow down. I'm not slowing down."

Paul Kedrosky: And even worse, they said, "Even if we all agree here in the US frontier companies to slow down, China won't." And then you're getting into this terrible discussion with with the whataboutery with respect to what China will do. And so, this is the problem is there's a a race dynamic that's been created that if you stop, there's a fear that either my competitors will gain an advantage or my shareholders will punish me. And so, as a result, the race dynamic continues. And that's what guarantees the overbuild. As long as you have a supply of capital and a race dynamic, you're guaranteed overcapacity, much like whether it's canals, railroads, electrification, take your pick.

Host: Azeem, you you and your team over the past year, year and a half, have been building out this dashboard, trying to track how much stress there is in the AI industry financially, trying to think through what it looks like for this to be on a healthy financial path, and what

Host: ...it looks like to be a bubble. So, why don't you walk me through sort of your perspective, the indicators you watch most closely, and why you're a little bit more optimistic right now than Paul is.

Guest: Well, when we're looking at a technology boom, a technology cycle, we've got to keep into consideration two different things. The first is, are customers finding it useful? Are they paying for it? Are they increasingly paying for it? In other words, is demand real, and is demand there?

And when you look back at bubbles that burst, it is almost always connected to a problem in the financing mechanism, the refinancing of shorter-term debt. It runs into a problem, and the capital goes away, and you see all of the things that we associate with bubbles.

The question is, where are we today? And where we are today is that demand is very strong from American and global enterprises. There is not an overbuild yet. There is vast undersupply of the compute needed for AI. And the race seems to be: can we get enough chips and power enough chips over the next two or three years to keep up with the demand?

And on the financing side, we are certainly moving off the cash that the hyperscalers have earned over the last 20 years or so, and we're bringing debt on board, but it's not yet at the levels at which you would say this is a historically high and complex position. So for me today, it's in that productive boom, with the hints of maybe getting—forgive my French—a little bit stinky.

Host: So I want to focus on the supply versus demand right at this very simple level. What Paul's basically saying is that there is a strong incentive to build, build, build. But if you actually do the math on the cost to produce tokens versus what's being consumed, there's very little actual monetization that comes back. Is that an asset that people are investing in or not? So, I'd like you to walk me through how that demand picture looks to you.

Guest: Well, we have to start a layer below the tokens. The asset we're investing in is data centers, 65% of the cost of which is the chip, the GPU, or an ASIC. And those chips can run not just this generation of models, but the next generation of models, and the generation after that. Those GPUs have an economic life of three to five years.

As long as the utilization on those GPUs stays in a reasonable range—say 60%, 70%, 80%—whether you're running training workloads or inference workloads, they generate cash. And over time, you can cascade older chips down to lower-value but high-volume tasks, like moving prior-generation chips into inference or enterprise retrieval tasks.

Crucially, what we see is that enterprise appetite for compute is not only real, but it's maturing into long-term contracts. Enterprises are not just buying on-demand spot instances; they are signing one- to three-year capacity commitments. Hyperscalers and cloud providers can underwrite debt and capital against that visibility. All of that for me is a sign of a market that is maturing week after week, quarter after quarter.

Host: Paul, what do you—how do you see that differently?

Paul: I hardly know where to start. Um so, let me just say like a lot of the discussion, and I use that word advisedly, with respect to how people talk about this... conflates fundamentally different things.

When people talk about AI demand, they often imagine this wild scenario where someone says, "Bro, build me a fitness app." And all of a sudden, seven words turns into a million lines of code. That's not anything like what you and I do for the most part. A lot of what normal people and normal businesses do is closer to kind of a one-to-one...

Paul: relationship or even compressible. Like I'll say, I got 25 emails from Goldman Sachs babbling about something or another. Are any of those things worth reading? And that's compressible, right? Because I'm taking a whole lot of stuff and producing and turning into something small. And that feels efficient to me. Or similarly, some other email or a calendar or whatever. So, the point being that normals, the people that these frontier models want to sell to now, are not really well represented by what people in the early days of the market were doing. Because decoders are really unusual. And you're already seeing some of that with uh Meta cutting in half the number of developers in-house who are allowed to use external frontier tools, Microsoft the same thing. You know, the days of token maxing where you like had, you know, charts internally that said, you know, who's the winning at using the most tokens. Those are long over. Normal companies can't tolerate that. People were blowing their entire budgets in like months. So, my point is is we're already moving away from this early days of profligate token use and everybody looking for ponies inside of giant piles of manure.

Host: Azeem?

Azeem: Yeah. In his giant pile of manure, Paul also had some sleight of hand. What Paul talked about was the frontier models and tokens. And what I talked about was the capital expenditure and the capacity we've been talking about, which is the GPUs. And I think that they are two different things that merit two different discussions. I mean, what happens to the frontier model labs, OpenAI and Anthropic, with this business model of ever more powerful AI systems that have a smaller and smaller yet sort of set of users is quite different to the question of what's going to happen with the trillion-plus dollars or so of capital that will go into building data centers upon which American and global businesses will start to run increasing amounts of AI. And I think we need to separate those two out. One is really about like the fundamental viability longer term of of Anthropic or Open AI and what they might do in the face of the pressures that that Paul has articulated. And I think the other is whether the the the GPUs and this new digital infrastructure that's being built out will find uses amongst predominantly enterprises, but of course as we rightly heard consumers through too through things like you know news and other agents. I think those are different questions.

Host: by you separating these questions. And I recognize it maybe this is an important point of confusion to to untangle because people can mean different things when they talk about whether or not there's a bubble and what it would mean for the bubble to pop or to deflate. As far as I can tell if it turns out that Open AI or Anthropic or for that matter, you know, Alphabet or whoever, that their business model here doesn't work that in fact people are perfectly happy to use Deep Seek from China for a lot less money and Anthropic like collapses in on itself that is going to pop the economy. Like that looks to me like it would become a contagion in the stock market. I mean, I don't necessarily think it throws us into a depression, but I think that would have a very very very big effect on valuations, on what people are willing to lend in the future. Do you not think that?

Azeem: Well, I think what you have to do is zoom back out and ask what is the the fundamental risk that emerges with a with a an infrastructure build out, which is you don't have enough customers who are demanding the product and paying enough for that underlying asset. And if people can move enterprises can move to open weight models that are being run directly by a hyper scalar that is just the demand moving. It may hurt Anthropic's margins, but it still monetizes the underlying capital expenditure. And then what one of the things

Host: Which which are these chips basically? The data centers?

Host: Nvidia Anthropic might collapse, but Nvidia will be fine.

Azeem: But but even so, Anthropic has got many many degrees of freedom. I mean, the last 24 hours or 48 hours since before we recorded this, it it cut prices and released even cheaper, more efficient lightweight models. These are still models around which they have a pretty decent margin. They will be doing the things that software companies have done for for decades, which is trying to construct reasons for you to use their software beyond the core capability. That's why they have Claude co-work, and they have Claude for lawyers. And that's why I think this question of if Anthropic disappeared tomorrow overnight, well, I can't quite imagine the circumstances under which that would would happen. But what I'm saying is that you're seeing the end customer demand, and that's the thing without which this will always end horribly.

Host: Can I just ask you also to because I want to actually ask you both this question of what for you counts as the bubble popping? Because there's a world here where, you know, you have, say, a giant stock market correction, but the data centers are still being used, and it sounds to me like you're saying that's not a bubble popping. That's just a correction or something.

Azeem: Well, historically, bubbles, when they've popped, have resulted in two things. One is a huge decline in the level of investment in that particular asset, and they have been accompanied with a significant crash in a relevant asset market, whether it's the stock market or or something else. And that is the test that that I would use. I think we may end up in a world where there are grindingly low and embarrassing returns for credit funds and for hyperscalers because of the investments they've made, but that degree of pain is not a bubble bursting. That is just somebody made some bad investments and they bought high when they should have bought at a lower at a lower price. I mean, I think I would use a very very classic historical definition, which is that we need to see the level of investment drop by 50% and we need to see a really really substantial and typically it's been 50% or more correction in whatever the relevant market is.

Host: Paul, same question to you.

Paul: I do okay. So, I I sort of disagree on almost every point that Azim is making. These are arbitrary definitions and I think let's just let's put it this way. Okay, so back to my original point. Something like 50% of people's current net worth in the United States is tied in equities. Most of equities is currently tech. Most of tech is currently AI. That's going to only get worse as these two monster IPOs come to market over the next, I don't know, 2 months or whatever.

Host: OpenAI and Anthropic?

Paul: OpenAI and Anthropic. And just to put that in context, by the way, OpenAI, Anthropic, and SpaceX, if you want to include it, combined, those three IPOs are larger than all post-World War II IPOs combined. Like, this is a really unusual moment.

Host: That's a crazy stat.

Paul: It's a crazy stat. And not only that, the combined value of those companies is larger than the value of the Nasdaq market capitalization at its peak. Like, the bubble is sitting right in front of people and they want to say, "Well, I expect you to come with blonde hair, not brown hair." Dude, just because it doesn't look the way you want it to look doesn't make it any less fragile. So, you could imagine, let's say, just for the sake of argument, the Anthropic IPO gets pushed out. It doesn't happen in November, it happens next year. There will be a sharp revaluation in their ability to raise funds, as well as in the companies that are supplying them, because everything is leveraged off the valuation of these companies that are currently sitting in private markets. So, I I like to say this is the failure here is over determined. There's so many ways it can all break and break really catastrophically because I'll give you another example. The the what's called the sick the Schiller cyclically adjusted price earnings multiple, which is a a sort of a longer term measure of where US equities sit. We're sitting at the levels we sat just before the crash of '29. We're sitting at absolutely historically

Paul: unprecedented levels in terms of market-wide valuation. And this is the problem is people lack imagination and simply refuse to engage with the broader data about what's actually happening in the world cuz they say, "Well, this doesn't look like the last bubble, so we must be fine."

Azeem: Well, let me But I think I'm going to I'm just going to come in directly at it. I mean, I I I think that it's very easy to use very colorful language. I'm always mindful of

Paul: Well, this is data. It's not colorful language. It's actually numbers. These are numbers like what's what's your response to like we're sitting at historically unprecedented numbers in terms of like the cyclical price adjusted earnings. Like is that good? Is that good in your book?

Azeem: Well, one of one of the one of the responses about the Cape Shiller is that Cape Shiller doesn't say anything about timing.

Paul: Well, it's because it's because it's a it's an adjusted long period horizon. It's not just a near term thing.

Azeem: it's a historical 10-year look back that can sometimes understate underlying changes in the regime of the the the economy

Paul: Okay.

Azeem: and the market. But there's there's something else that's going on, which is I think a lot of what you're describing Paul is already priced by equity investors and it's being priced by the credit markets. We've seen it just in the last few weeks. We've seen an IPO for a data center company Australia get pulled because investors weren't happy with what they were being promised.

Paul: That's not why it was pulled.

Azeem: Well, why was it pulled then?

Paul: It's pulled because of this unprecedented moment where $2 IPOs are in orbit around us waiting to come to market, and are sucking all the risk capital out of the market as people say, "Yeah, this looks really cute, this IPO we have, but there's we don't have any capacity to do that right now."

Host: Okay, but I think that in a way actually goes to Azeem's point a bit. When you talk about these massive IPOs that are circling, and we're talking about Open AI and Anthropic, and Open AI's gotten pushed back a little bit for when they seem like they're going to launch actually into the into the public markets. The thing that is interesting about that from, I think, the perspective of it's not all a bubble, is that sophisticated investors in mass numbers who really, really do have skin in the game here, are looking at these and saying they make sense. So, when you hear this from other people, Paul, and you're an investor, and you think about these things, what mistake do you think all these people are making? When you say it's just math, it's clearly not just math, cuz everybody else can do math. It is your interpretation of the math. So, where does your interpretation of the math differ from, you know, the market?

Paul: sure. So, you have to think about why why these kinds of IPOs happen, and why people are piling into them. It's partly in anticipation of what they think could be, you know, a fantastic business with a long-term franchise, da da da da da, but it's also because they think that when it comes time and when the lockup expires, they can sell the stock to someone else who thinks that, right? This is This is the essence of going back to the very beginning of why do IPOs exist? IPOs used to exist to allow you to raise money for a company who needed to grow. These are not that, right? These companies have done their growing. They're probably actually going to be, I don't know, $100 billion in run rate revenues by the time they come out, or $70 billion, or something like this. So, these are exit events. These are people inside the company saying, "I don't need my stock anymore, you can have it." So, they're selling their stock to people on the outside because they let this at this valuation they'd rather have cash than Anthropic stock, or cash than Open AI stock.

Azeem: valuation, I'd like to have cash, too.

Paul: Yeah, sure, absolutely, yeah, I'd hand it over. But, the point being that like this is not a this is not a a gesture where they're saying, you know, I really like the our prospects with the company. There's all there's going to be selling shareholders across the board because this is an exit event, not a financing event.

Azeem: is a gesture where other people are saying, I do like your prospects as a company, and I want to own this.

Paul: I the same thing was said in anticipation of the SpaceX IPO, and we've all seen what happened subsequent to that, that all of this fervor before it turned into a, you know, kind of a a sort of a light roller coaster decline. So, my expectation is you'll see the exact same thing cuz these are, in market terms, exit events. You're going to be a wall of selling shareholders showing up post lockup selling into anyone who'll show up and want to and want to take the shares from them. And that's why this needs to be understood for what it is, an exit event. People are buying in anticipation of selling to you. If you want to It's like the old line about the poker table. If you don't know who the greater fool at the table is, it's you. This is that, and I don't understand why people don't see it that way because that's what makes this so unprecedented.

Azeem: I'll just I'll just come in on that. I mean, I think what Paul describes is really what many IPOs have been like for for many, many years. And when we take a look at Anthropic and whether it can be a sustainable business, we probably need to look at what it will show in its S-1 when it's the S-1 is the document. It's the document essentially for what this company is when you're going to buy it at at IPO, and look at what's happening with its revenue trajectory and its margin trajectory, and the breadth of its customer base, and whether customers are sticking with it, whether those accounts are growing, beyond what has been, frankly, a lot of patchy reporting and leaks, some of which is accurate and some of which is is inaccurate, but not much of which gives us a cleaner picture of of the company. And I think that draws one of the really big issues we've all had to deal with over the last two or three years, which is these companies have got very big, but they are still private companies. They don't really need to disclose much if anything at all. And what comes out is a smattering of rumor, scuttlebutt, and the occasional well-sourced leak. And because they have essentially been structural to the AI trade so far, and they have been as a consequence the thing that has spun valuations of the of the wider market, whatever it is now, 38% tech. But we have not been able to inspect the way we normally can with public companies what's really, really going on. And I think what happens with the IPO is that we we won't get every answer, but we get some of the answers that we need about these these businesses.

Paul: We have a partial answer now though, just on that point though, which I think is amazing because it came from you know, what's that thing with the horror movies with the killers inside the house? This is a little bit of one of those where Anthropic accidentally leaked who their who one of their largest customers was. They said we needed to have What was it called? A mega account executive or something like this. They were trying to hire one, and then they all quickly realized that by saying who their largest customer was, you could triangulate and say that this was for Meta. So this was their account executive to handle sales to a single company. That it worked out that Meta something like 10% of their revenues. And so this goes back to

Host: Meta is 10% of Anthropic's revenue.

Paul: That's right.

Azeem: But that was in 2025, right?

Paul: So yeah.

Azeem: And let's just be clear. Let's

Paul: But just to be clear about what why that matters to them.

Azeem: No, but what matters is that Anthropic's revenue in the first quarter of this year was bigger than the entirety of last calendar year. So what we're doing is we're looking at a rearview mirror and saying, "Oh, we hit a squirrel 700 yards ago ago." What we need to look is look out of the

Host: [laughter]

Azeem: you know, the the windscreen.

Paul: This is why these companies are so unusual. We have never seen companies on this size go public with such a small number of customers. The

Paul Kedrosky: ...concentration of their customer list, it might be eight companies that make up 80% of their revenues. That doesn't happen. And so, as a I was an equity analyst years ago, this would be a colossal source of risk for me. I'm looking at it, I'm saying the dependency on such a small number of customers who are mostly using it for the same thing is historically unprecedented to have such customer concentration. So, what they accidentally leaked was the fragility of the business in terms of its exposure not just to a single customer, but to a single customer who was doing that same old thing of writing code. And this is this is the point I want to make. That's what makes these so unusual.

Azeem Azhar: The beauty of what Paul has just said, which is sort of eight customers 80% of the business, is that that's 10% per customer. That would have to be disclosed in the S-1. So, we have actually a very, very falsifiable prediction that we will be able to check when the S-1 comes out if Anthropic does successfully go public and say was it really 80% and eight companies representing 80% of the business. And that's really helpful because if Paul's right, that is a bit—that's a bit exotic, right? It's a bit spicy.

Ezra Klein: If you're not a subscriber to The New York Times, we have some news for you. You can now explore the Times for free without any paywalls at all during your first month in The New York Times app.

I think Apple is a really interesting company here. Because unlike Alphabet Google, unlike Meta, obviously OpenAI and Anthropic are AI companies. Microsoft was a major OpenAI investor and is now also trying to make its own frontier AI. Oracle is heavily, you know, invested here and is like now just like right above junk bond status. Apple just decided to not compete in making a frontier AI. While their competitors are now spending, you know, 100 to 200 to 300 billion dollars in capex every year, Apple's spending what, 15-ish billion in capex per year? Their stock this year has performed better than any of these other companies' stocks. Does it imply anything that Apple, in being the one major player here, to not try to compete in the frontier model Olympics, seems to be winning that?

Paul Kedrosky: ...the way Apple does it. Let me start with the theme here.

Azeem Azhar: Yeah, yeah, sure. Yeah. I mean, Apple has a business model that means they don't need to run a frontier model. We buy their phones, we buy their laptops, and then we might subscribe to the odd movie or something like that. So, they make money regardless of the AI. But I also think there's a really important cultural dynamic, which is Apple cares a lot about detail. I remember this story of Steve Jobs when he was rolling out the iPhone many, many years ago, phones the then senior vice president at Google who's responsible for the partnership, and blasted him out because they'd used the wrong color yellow on the then Google logo. Apple really cares about detail, and they care about control. And large language models are powerful in many ways, but they're not about control. They can get very, very hapless. And I think that one cultural reason why Apple would have held back, even though it's so well positioned to, you know, be the personal agent—it's got the device, it's got the privacy—is that it won't want the mess that a large language model will ultimately deliver if you've got a billion people using it, which is what Apple's estate looks like.

Paul Kedrosky: So, yeah, I don't think Apple was very smart about this. I think they mostly screwed it up and got lucky. Um so, they tried various times to make forays into AI. I mean, they had teams, they disbanded teams, they were going to integrate it with Siri, they weren't going to integrate it with Siri, and they didn't ever get anywhere. And what they benefited from though was by their inaction, they were a classic free rider, right? They got to free ride on everyone else's gigantic expenditures, and they come out the other side smelling like roses because they didn't actually have to make the expenditures, they can use what's being built by others, and sit on top of it.

But, I think writ large, we're going to see a lot of that because people are burning money on a bonfire, which can benefit lots of other companies, you know, so Apple sitting on top of it not having to build models. Uh, look at people like Cloudflare, for example. I mean, here's a company that largely forgotten, was kind of basic internet infrastructure, stock has tripled in the last, what, 18 months because suddenly people like, "Wait a minute. Wait a minute. If we have all these bots ranging around the internet, that's kind of like a new form of attack. That we need to make sure that we have who's out there that they can deal with all of this new traffic." And so a lot of this investment that we've seen out there will turn out to be beneficiary to a whole host of companies whose names are not OpenAI, Anthropic, Grok, or whomever else.

And I think that's a really important point. And that goes back to the, you know, the days of electrification in the 1920s. Like, who were the primary beneficiaries once the utilities all began collapsing in the crash of '29? Well, it was of course the people building things that required electricity, air conditioner manufacturers and others. And so yeah, for sure we'll see a lot of this, you know, accidental piggybacking and free riding on top of this.

Azeem Azhar: But it's not accidental piggybacking and free riding. It's—

Paul Kedrosky: Apple didn't plan this.

Azeem Azhar: It's called complementary services. That's—and complementary businesses. That's what economists, when they're, you know, being very moderate with their language, just describing the process, which is a general purpose technology shows up, it continuously improves, and then it starts to create the need for complementary services, whether it's maintenance, or whether it is, you know, repair, or whether it's something you can now do with the technology. This is the normal evolution of a general purpose technology. It finds its way in, and in some cases, it's a bit of free riding. In other cases, what's happened with Cloudflare is that we've been able to across the economy automate so many bots and AI agents that businesses on the internet are turning to Cloudflare to try to protect them if they can.

Ezra Klein: Let me ask something actually about these big companies that has come up a lot in the bubble debates I have followed. Which is the major companies we are talking about here, your Alphabets, your Metas, etc., they just have a lot of money. And so they're doing this from a very healthy balance sheet, which is not the case sometimes for these railroad companies, etc. When Zuckerberg says, you know, it would be sad if we misspent hundreds of billions of dollars, what he's also saying is we can misspend hundreds of billions of dollars, it'll be fine.

And one argument I've seen is that one reason you don't really need to worry about this is that it is being driven by companies that can absorb losses on their own balance sheet without contagion in the rest of the economy. You're not dealing with something like, you know, derivatives in the 2008 financial crisis. Paul, you thought a lot about the private credit dimensions. I mean, there has been a movement towards debt in some of the the way this is being financed now. How do you think about the question of are we protected from this, we being broader society, because these companies have ungodly amounts of money to lose if they so choose?

Paul Kedrosky: No, we're not. So, this is a funny thing. So, like a year ago, people told me to stop babbling about this stuff because these are smart companies, they have prodigious cash flows, who are you to tell them what to do? And I was like, oh shucks, you've got me there. So, I'll stop saying this. And then, of course, we're now a year later. What does the data show? Spending is moving away from cash flow to debt. So, you know, you put up your hand and you say, hey, wait a minute, wait a minute. Remember that thing you said—

Host: ...last year? It's no longer true. So, should I worry now? And it's like, no, you don't need to worry now because they have pristine balance sheets. These are great credits. And I'm like, oh, crap, I might just my timing is wrong twice on this. And then, of course, to be serious, what happens whenever you have a pristine balance sheet? What Wall Street is very, very good at is making it less pristine because they'll they'll continue to raise money for you and raise and raise and raise for you because the fees are tremendous and what and as as they do so, your credit rating declines and we're seeing that with Oracle. We're seeing that we're on the cusp of that with Microsoft. We're on the cusp of that with some of the other hyperscalers because that's what Wall Street does. They see a pristine balance sheet, an unlevered company in the terms of the of art of the industry and they'll continue to burden it with credit.

And so the way that these companies try to get around it, Meta is probably the best example, is they use somewhat more exotic structures so the debt doesn't actually sit on their balance sheet. They use things called special purpose vehicles. They use things that are off-balance sheet financing. They say, "Hey, hey, hey, hey, that's not my debt. I know I control it, but it's over there with that funny named thing that sounds like a sort of a New Orleans bagel. Um that's somebody else's debt.

And so this is all just a shell game and the reality is it rolled up, we've now at a point where tech, whose main advantage historically was fast growth, no debt, right? Tech is now the largest issuer of investment-grade bonds in the United States. This is an industry whose claim to fame other than growth was that they were relatively clean. They're now the exact opposite. These are hugely debt-encumbered companies increasingly and that are dominating the issuance marketplace to the point that it's even having an impact on sovereigns, right? So one of the reasons why we had the the bond market freak out over the last two, three weeks was because of the role of hyperscalers issuing credit at the same duration as the 10-year in the United States which at the margin was actually outcompeting the US which caused yields to have to go up in the 10-year. So that's how crazy things have become. So yeah, I know it's completely transitioned which is going to have, you know, huge consequences down the road in terms of this debt overhang that starts to look increasingly like, you know, Japan circa 1990. This kind of all these corporations with giant balance sheets loaded with debt that are trying to work it off.

Host: Well, the other side of that, Azeem, is that it's not just debt that is being crowded out where these, you know, hyper scalar companies are competing with everybody else for for money, it's also construction, it's also real materials. I mean, I know I'm sure you've seen these charts going around that imply there's real crowd out happening. That yes, we are building, we are pumping money into AI CapEx, but home building in a country that does not have enough homes is going down. Spending on manufacturing plants and factories is going down or at least like the the the sort of share. And so there there does appear to be a movement from investing in other parts of the economy at a construction and CapEx level into investing in AI. I guess if AI works out great, that's fine, but if it doesn't or even if you just think these other things are, you know, underrated as I say, maybe I think for homes, homes are more stable investment. They don't return give you returns like the like investing in an Nvidia does, but they're pretty important and I have like a cleaner way of explaining why more homes would make more people's lives better. How do you think about the the crowd out question here?

Azeem Azhar: Um when we look at uh investment rates, a year ago back in the summer of 2025, I wrote about how I assessed whether we're in a boom or a bubble and one of my observations was that at the moment the hyper scalars are spending out of their own cash flows and their own balance sheets. As soon as that starts to change and they go to external finance, you introduce lots more risk and my thresholds for more risk was a about 30% uh of the uh uh 30%, which is about the level that we we are approaching now at the end of 2026. These things are are not set in stone.

Uh the crowding out problem is a is a real problem and it's it's it's really uh a complicated and thorny issue. I mean, I think one place to to think about was um what happened in in the Netherlands in the 1970s when they discovered a lot of uh natural gas uh off the North Sea, and uh the entire economy oriented around gas extraction. You got this thing called Dutch disease, where effectively that part of the economy was sucking up all of the activities. And there are, you know, there could be a risk that that will happen in the US if it isn't already happening right now. And and I think then the question is, what do you do about it, and how do you actually uh you know, adapt to that? And I'm afraid my only answer is that it requires something of policy, something of a policy stimulus to make sure that those sectors get the activity that they need, because I'm not persuaded that market forces on their own would would do that.

Host: That seems like a tough uh prescription at a time when as they both we have very erratic leadership, but our borrowing costs are going pretty high amidst incredible levels of debt.

Azeem Azhar: Well, look, your your politics is your politics, and you know, you get to vote every 4 years and make the choices that that that you make.

Host: And then live with them.

Azeem Azhar: [laughter]

Azeem Azhar: Right, and then live with them. But but it doesn't have to involve subsidies, right? It can involve other types of you know, policy interventions. So, I'm not it's it's a it's a really difficult problem, and it will continue to happen so long as the returns on the other investments are are higher, and so long that that's that that that that is creating a supply shock on electricians and construction workers and HVAC technicians and all the other things that are needed in data centers.

But but and there is there is a but to all of this, which is that of course, there's a lot of upskilling going on. There are many, many electricians being trained. There are many HVAC engineers being trained. There are many construction workers being trained. There are construction firms who are growing, and who will start to have spare capacity. And I don't know how that ends up, but in a year or two or three years that capacity, those skills will be in the economy and able to be used.

Host: So, I want to ask this not in terms of stock markets and financial flows. I want to try to ask what does this economy look like in which companies and people and governments for that matter are using enough AI to make these investments writ large pay off. Even if, you know, some companies may not end up being in the winner's circle here and may may fail. So, Azeem, you spend a lot of time thinking about this. If you were to try to paint me a picture of what things look like at normal companies, your GEs, your Coca-Colas, um in this world where the AI bet pays off, what does it look like? What is different? What are we not What is not being spent on if so much money is being spent on AI tokens? Like, what is happening here?

Azeem Azhar: You know, when I talk to to bosses and I talk to them across Europe and the and the US, dozens every every year, they the the I I hear this the same sort of thing, which is um we're making progress with AI. It's harder than we thought. We're going to keep doing it and we're starting to get results. And the more confident ones say, we're getting more and more results.

But a large part of it is coming now, I think, will will come unless businesses grow very, very quickly from just a slow reduction in in replacement of of workers. That companies that are getting more productive with AI will be delivering more dollars per employee. And if they're not growing faster than that productivity growth, then their their head counts will will start to to to fall. But we

Speaker A: I should just add that, you know, there are very few AI native companies. When we think about electricity, the electricity native companies were people like the Ford Motor Company with its end-to-end production line. But most of America didn't get fully electrified for another 15 or 20 years after that. And I don't think I can really point to a Main Street firm outside of Silicon Valley that you could say this company has become AI native. What you can see are bits of the business that look like they're using a lot of AI.

Speaker B: Okay. Okay. But if the AI native company that you're talking about is a company where you've managed to replace people with AIs—

Speaker A: Yeah.

Speaker B: So now you seem to be really quickly tipping into the much worried about world in which AI is net negative for jobs across the economy, unless you believe there are all these new companies starting that are somehow very human labor intensive—even though, in theory, the new companies starting are going to be more AI native, so less human labor intensive. So is that actually a good world? And then if that's the world, who's spending more money in this world where people are losing jobs?

Speaker A: Well, I think there are a number of different things going on. You know, one is that there are still things that the AIs won't be able to do, and those will be the layers of humans doing verification, doing the last sign-off, doing the things that the models cannot do well. And what would typically happen is that those people will end up being paid more and better as a consequence of sitting at that choke point.

I think that the second thing that you will start to see is a whole class of roles that we can't imagine within companies. You know, whether it's things like the webmaster, which didn't exist in 1993.

And then I think the third thing is really the cultural dimension of a general purpose technology. This is where Paul was describing the GPT: they tend to construct entire new classes of work and behavior that we hadn't been able to put our fingers on. If you just think about Henry Ford, because of Henry Ford, we ended up with motor cars; because we got motor cars, we ended up with people having suburbs; suburbs led to vacations; vacations led to all the services that sit around there.

Now, I'm not blase about that. I mean, I think the challenge we may find is that it takes longer to create those new roles than anything else that happens. But the other observation I would make is that my time horizon for this runs in 5, 7, 10 years, not one or two, not two or three. Because I do think it just takes time for management teams to absorb what it is they have to do with this technology, and it's just not the case that you get the best model from OpenAI and say, "Fix my company," and it fixes it. These things take time, and that's what I see when I talk to bosses. However fast they want to move, it just takes time.

Speaker B: Paul, how do you see this?

Paul: Uh, we're basically re-pricing labor, right? I mean, this is the fastest adopting technology in history, and its fundamental function is to re-price cognitive labor. And this is the deep struggle, because the article of faith, the West Coast ideology on this stuff, is technology always creates more jobs than it destroys. And so it becomes just a thing that people repeat like a mantra that they hope will become true.

And the reality is we've never seen a general purpose technology that had such broad applicability, was adopted so quickly, and was re-pricing labor at the same time. Because when I say re-pricing labor, I mean it's making it vastly cheaper to do things like call centers, for example. And as a result, this change is going to be destabilizing long before we figure out how to turn them all into—I don't know what the equivalent of webmasters is.

And so this is the problem, right? Is that it's happening so much faster than any prior GPT that these sorts of arm-waving about creating more jobs than it destroys belies the lived experience of people who are affected by this stuff. And so it can quickly become just some kind of strange mantra like, you know, "Don't worry your pretty little head, everything will be fine." So, no, I think the disruption will vastly outpace our ability to create new positions.

Speaker B: This has also struck me as another form of economic risk for not just workers, but actually AI and for, you know, some form of this bubble popping or deflating if it is a bubble. Because you can imagine a world where actually the problem is not whether or not companies are willing to adopt AI.

Paul: [snorts]

Speaker B: That actually, particularly under more economic strain—interest rates keep going higher, maybe there's a recession of some sort—companies begin moving very aggressively to replace human labor (which needs healthcare, unionizes, and has opinions) with AI. And then there's a huge backlash. There's already a huge AI backlash: already people organizing against building data centers; they don't want those where they live.

But if you know anything about the history of occupational licensing or building homes or anything else, it's actually not that difficult for, say, teachers to organize and say, "No AI in the classroom. Actually, education is off-limits for AI." Doctors all of a sudden to organize. Maybe there's a couple of bad events where people are getting bad diagnoses off of AI, and so it's sort of locked out of hospitals.

Anthropic in their risks to their IPO says, "Existential risk to humanity is possibly one of the risks of this IPO." If there's some major control incidents where, instead of hacking Hugging Face, OpenAI hacks Wells Fargo, or it hacks something in China—

Paul: Yeah.

Speaker B: And all of a sudden regulation comes in, not a little bit, but all at once. All of those things seem like real risk—that the possibility of discontinuous regulation—

Paul: Is that right?

Speaker B: —to me than elsewhere, and also seems, I don't want to say inevitable to me, but quite likely to me. So I'm curious how you think about that.

Paul: This has been my complaint from the very beginning of the process by which these companies have launched models: that they've essentially been conducting unlicensed drug trials on the public, in the sense that these are hugely consequential technologies that are launched in a way that, were they a drug company, you would immediately have the FDA on your back. And yet, because they're technology companies, you can do these mass trials: "Oh, by the way, we're going to transform all of education. We're going to launch these things that can penetrate government websites and other things."

So this idea that you can do this with impunity in technology because we live in the timeline where so far these things haven't killed us is just ridiculous and indefensible. And I think we're inevitably headed towards a world where, after some kind of terrible event, there's going to be some kind of post hoc regulation imposed, but it'll only be after consequential damage.

And I'll give you a specific—this is the way I often try to characterize it to people, because they'll say, "Well, what's going to cause that? You can just come up with this stuff." If you think about it, there's probably 5, 10, 15 different sorts of scenarios we can come up with that might lead to regulation, none of them probably having a probability of more than 5%.

So let's say I have six events, each of which have a 5% chance of happening. What's the likelihood that any one of them happens? It's like 20%. If I have a dozen events—possible scenarios that are all bad, each of which have no more than a 5% chance of happening—what's the likelihood that one of those happens? It's more than 50%. We have way more than a dozen ways that these things could cause trouble and end up requiring regulation. So looking at it that way, it is almost inevitable. So it's like—

Ezra Klein: a 50-60% chance that something happens that drives regulation, and yet people continue to sit around and say, "Well, you know, this isn't that likely, or that isn't that likely." So, no, regulation here is inevitable. It'll take the form of something like the FDA, maybe some kind of atomic energy board, because it's just as consequential, and because there are so many different ways for them to have broader consequences in society. And it's amazing to me that people sort of glibly pass that off. Azeem, how do you see this political risk dimension?

Azeem Azhar: Well, you know, I think when you look at industries, industries grow quite well when there's a little bit of regulation. As the pharmaceutical industry has, you know, is a great great example. And I think when we think about who's going to pay for this, the AI companies from the model labs through to the hyperscalers are going to be making a lot of money. So, it's fairly clear that there is money in the system, and there are people who firms who can cover some part of the bill. And I think one of the things that's really exciting about the data center protest that you're seeing, particularly in the US, is that actually it is people expressing their sense of unease about this technology, the way it's manifesting itself. And the way to look at that is that that is a democratic expression, right? That is an expression of will, that is an expression of uncertainty. And what you would hope to come out of something like that would be a better social license, which people are more and more comfortable with.

Azeem Azhar: Now, you know, I suspect that within the US, because you have state-level governance power, that some of this may be done on a state-wide basis. What's happening in the UK is that there is an attempt to get people developing data centers to reinvest annuities into the local community. Because one of the things that we don't really talk about much is that a huge data center might employ very few people indeed. I mean it may be billions of dollars, but it's a handful of jobs and that's not really—

Ezra Klein: Like two to six people for a multi-billion dollar facility.

Azeem Azhar: if you do the math in terms of the job cost per jobs created, it's often—I was looking at one in New Mexico recently. It's $4 million in terms of subsidies per job created leaving aside the cost of the center. If you look at the tax abatements and the required improvements to you know, local utilities. The numbers are insane.

Ezra Klein: So none of that works long-term. But the useful thing is that there's money in the system because there are profits being made up and down the stack and you know, one would feel that because of that there's an obvious place for that bill to land and that bill can land through—you know, I don't think self-regulation works and I would agree that we will need to find whether it's local, national, or international real rules about how these systems work and the harms and the damages that they might cause.

Ezra Klein: Well, but the thing to me that's interesting in that—of course there's going to be regulation. That's clear to me. The range of how that could fall all the way from pharmaceutical regulation where what we want is for pharmaceutical companies to produce safe drugs. Although people may maybe argue, you know, we've made that too slow also. All the way to say nuclear power regulation where fundamentally what we've ended up wanting is no new nuclear power plants. And depending on what happens with AI and given how negative people are on it already, one of the possibilities here is that it could swing such that the answer on a bunch of things is simply no. We don't want to be replaced by robots. Which actually when you say it aloud like that seems like a fairly straightforward thing [snorts] and probably the kind of thing that people are going to express in a democracy or no, we don't want robot agents crawling all over the economy. No, we don't want to be called endlessly as a restaurant by your Muse agent trying to get reservations everywhere because now you've just made it functionally impossible for restaurant to ever answer the phone again with a human being and that the sort of like collective expansion of all this. I mean, I know people who you know are managers of different firms and they're just saying like I'm just drowning in AI produced reports all of a sudden. My email is just full of obviously AI written inquiries, which is making my responsiveness all across the board go down. I'm not just saying will there not be regulation, but I think one risk is that people decide maybe they don't want this in actually fairly large swaths of the economy or used in some of the ways people are expecting. There's a very good piece by John Herman in New York magazine about the sort of beginning agony of companies that all of a sudden are absorbing tremendous amounts of Muse agent inquiry. And like we're not prepared for everybody to have an agent calling you as a robot trying to negotiate down the price or get a reservation or sell a concert ticket or just whatever it might be and people may actually not like it.

Azeem Azhar: Yeah. They might not like it. The trouble is there appears to be some revealed preference here and that picture is not even. So only about 2 and 1/2% of Americans pay for an AI subscription, which is considering they've been available for 3 years is a very slow slow take off.

Ezra Klein: I love this stat in one of your newsletters recently that more people pay for GLP-1s—a larger number of Americans paid for a GLP-1 than for AI.

Azeem Azhar: Well, yeah, because one is really quite useful and you feel the impact within a matter of a few weeks. But hundreds of millions of people in America and around the world do use AI through the free version and when companies like Meta start to put these tools in front of you, I'm sure people will end up using them even if they vote with their social media posts saying they don't want any AI. Either it'll become inescapable or they will still have it on their phone and they'll be using it. I think we have to just contend with that possibility.

Azeem Azhar: I think we also have to realize—and this is where we think about technology as being this gigantic unpermissioned experiment, which it has been for 40 or 50 years, that what you're describing about robo calls from AI agents into hotels and restaurants or as my orthopedic surgeon was telling me, he just gets long thousand words, 10,000 word emails from patients who've clearly chat GPT'd their medical notes. This is unfortunately going to happen and that cost has now been pushed into the system and we're going to have to figure out how to deal with it. I now pay, it's not a lot, I pay about a buck a month to my phone operator to screen out spoof phone calls that are essentially robo calls. And that's what's going to happen and we're going to call that GDP. Those additional costs.

Ezra Klein: I think though one of the things that's—your comment Ezra really reminded me of this is that society we have this idea in sort of Western capitalism that it's always about efficiency. In a weird way, societies require friction.

Azeem Azhar: Mhm.

Ezra Klein: And when you take away frictions, even frictions that you didn't realize existed, all of a sudden it has consequences. You know, there've been examples of people using large language models to launch claims in small claims court, for example. Do we want to have less friction? Do we want more of that? Do we want this, you know, your example of the restaurants? And we have all of these frictions that exist in society and blindly removing all of them is kind of creating an—in tech parlance—almost a denial of service attack on modernity where all of a sudden we're having stuff launching at us from every angle, almost like, you know, Russian bots attacking a website, except it's coming from all angles because of the increasing numbers of bots and other things. And I think—

Host: That's what people are responding to, is it's this kind of out-of-control feeling as all of the frictions drop away and suddenly, you know, I can't get access to things or nothing works the way I'd expect. So, simply dealing with the reduction in frictions has huge consequences across society. And I think people are reacting to that.

Host: >> going to happen is that it's going to be these two AIs bouncing things back and forth and...

Azeem Azhar: Right.

Host: And it's just going to be a more and more and more unpleasant economy, which is why like I do think there are going to be large areas where people will say maybe not. So, I want to end on this question, which is: first for you, Azeem, what would you need to see to start thinking, "Oh no, this is getting pretty bubbly?" And then to you, Paul, the reverse. What would you need to see to think, "Oh, maybe I do have this wrong and these investors have it right and there's a lot of money to be made here"?

Azeem Azhar: Mhm. Well, there are so many risks that we have talked about, and there's all the complexity and fragility in the financing. The thing that keeps this on the other side of healthy is whether real customers are spending real money and getting real results. So, the thing that would make me uneasy would be if I saw a couple of quarters where the amount that American enterprises were spending on AI was—the growth rate was halving, say, two quarters in a row. That would suggest that they filled their appetite and there's not going to be the revenue to keep this on the right side of healthy.

Host: Paul.

Paul Kedrosky: Uh, there's so many—I mean, I'm convinced I'm wrong about most things, so I'm always happy to find more things I'm wrong about. So I just look at economic data. So, if we began to see, I don't know, productivity growth go well beyond trend, that would be striking, right? So far, we're sitting at around trend data with respect to US productivity growth. There's no sign of, you know, AI improving productivity. If we saw participation rates sharply increasing, that would be something that would help to tell me, "Okay, AI is an inclusive technology and bringing people back into the economy." That would really help. If we saw GDP growth going back above trend—we're still sitting around the levels of, you know, sort of the post-World War trend at the low end of it. Any of those things would tell me that something big was going on, like a true productivity shock in the economy that warranted the kind of expenditures. So, you know, if any of those—and probably the three of them at once—happened, that would tell me I was wrong. You know, if those productivity numbers come through, then fine, everything changes. But if it doesn't, you're looking at something that looks more like the 1930s than it does like, I don't know, 2000s.

Host: All right, I think that's a good place to end. So, always our final question, what are three books you'd recommend to the audience? Paul, why don't we begin with you?

Paul Kedrosky: Sure. Um, Middlemarch is sort of my go-to on thinking about normal people being swept along by kind of the winds of change as a major technology revolution taking place. It took place just as the Industrial Revolution was sort of taking hold, and I think it's a tremendous book about that. So, the other one that I like to recommend is Liberty's Dawn by Emma Griffin, which is a collection of the average workers' biographies and notes from during the Industrial Revolution. And it's just a wonderful, wonderful look at what it was like in the middle of that wrenching transition. And as people talk about economic data and everything else, what it was like to live through it. What did it feel like? What was the pain I felt? And, you know, what were my hopes for the future? It's just like a wonderful book about that. And the last one that I was going to—Samuel Insull, largely forgotten figure. He was probably one of the most important entrepreneurs of the 1920s, whose empire in electricity shows what happens that you can build—building an empire on top of a GPT and even doing that successfully can still end up in collapse and calamity. But it's a wonderful story. So, Insull: The Rise and Fall of a Billionaire Utility Tycoon by Forrest McDonald.

Host: Azeem?

Azeem Azhar: So, my three books. The first is The Day the Universe Changed. It's written by a BBC science journalist back in the early '80s. I read it as a teenager and I've just reread it, and it is all about how as we develop new types of knowledge, we start to rethink our role in the world, what matters to us, and how unsettling that is. And he walks us through a thousand years plus of that through Western and Asian history. The second book is Fathers and Sons by Turgenev, which is a book written in the 19th century. It's at that moment where the ideas of science and modernity are making their way into Russia and you have the generational conflict between the more traditional generation and the younger generation, and what stays as a bond between them, what even the most recent breakthroughs in science can't tell you about what it is to be human. I think it's very relevant. And then the final one which I'm just rereading again is Candide by Voltaire.

Host: Azeem Azhar, Paul Kedrosky, thank you very much.

Azeem Azhar / Paul Kedrosky: Thanks very much.