"Yes, AI Is a Bubble. There Is No Question."

There's a theory about technology and how it changes the world that I find myself going back to over and over again. Carlotta Perez, in her book Technological Revolutions and Financial Capital, proposed that whether it's the railroads or the canals or radio or the internet, What tends to happen is this. Companies get so excited about the prospect of the next big thing that they always overbuild. Spending always races ahead of revenue. What follows is a crash and then a golden age.

It's important to say, of course, that the railroads and the internet were both bubbles and life altering inventions. The bubble and the golden age are not one thing or the other. In fact, they tend to follow in a particular sequence.

So here's where we are today with artificial intelligence. The biggest companies investing in artificial intelligence are spending about 700 billion dollars on AI every year. That's on the chips and the data centers, the power, adjusted for inflation. That's approximately one Manhattan project every three to four weeks. It's one Apollo program every five months. This is an insane amount of money for the private sector to put into anything. Six months ago I was quite certain that AI was a bubble.

And I was certain for the very simple reason that I thought history was just repeating itself. AI spending was rising faster, I thought, than revenue could possibly match. And now here I want to be very careful and clear because I think some people who don't like AI for all sorts of reasons also hope that it's a bubble because they don't want artificial intelligence to turn out to be something that's important.

But a bubble in AI would itself be one of the most important stories of the decade for the stock market, for the economy, for unemployment, for housing, for politics. In a bubble that blows up in 2028 would be more consequential, I think, than who the party's nominated for president. For most of last year, I was convinced about the AI bubble story, even when I spoke to industry insiders who fiercely disagreed with me. But in the last few weeks.

I've changed my mind. And I want to be very, very clear about what happened that changed. In late 2025, the AI companies Anthropic and OpenAI released new agents. These are AI that can work autonomously on complex projects that often require multiple stages of reasoning. You can use agents to write software, to spin up a new website, or to break down big complicated data sets. My own work requires a lot of laborious and

clicking around government databases to come up with stories. So in January, I tried a little experiment. I used Claude Code to dig into the American Time Use Survey. I said basically, hey, Claude, I'm a data journalist. Look through these hundreds, thousands of pages, and tell me how childcare has changed for fathers in the last two decades. Tell me, has childcare time gone up or down? By how much? For educated fathers versus less educated fathers, for dads of young kids versus old kids?

Make me ten graphs I can use to write an essay about, let's call it, the changing face of fatherhood in the twenty first century. Five minutes later, Claude Code came back with ten graphs. An important question. Were these graphs correct, or were they fully hallucinations? Well, I didn't know myself, so I sent all these graphs off to an economist that I know who does similar analysis. Just check my work, I said.

He did. And his results came back almost identical. Except this project took him, an expert, several hours. And it took me, a novice working with Claude, several minutes. Reading reports about software developers who use agents all day long to write code, and watching and thinking about my own use of this technology, I had an aha moment. Not aha, this thing is going to replace all economists.

But rather, wow, every economist is going to be working with this sort of thing constantly. And data journalists and coders, and consultants making decks, and bankers building models. My basic insight was this. If enough knowledge workers are constantly running agents throughout the day, demand for compute is going to soar.

This kind of AI agent work, I think it's important to say, uses a lot of tokens, the fundamental unit of AI labor. That means it requires a lot of chips, a lot of power, a lot of money. There are reports of software engineers who spend tens of thousands of dollars a year or even in a single month using agents. That's the salary of full-time employees now going effectively to AI.

And the revenue numbers are remarkable. Anthropic recently reported that it doubled its annual revenue in two months. OpenAI is growing almost as fast. According to the payment company Stripe, which has a god's eye view of thousands of companies using its platform, AI firms are now growing faster today than any generation of companies they have ever seen. In short, more agents doing work means more tokens. More tokens means more revenue and more demand for chips and data centers. And altogether,

If bubbles happen because spending races ahead of revenue, then the current trajectory of revenue, I thought, severely weakened the case for a bubble. But as the bubble case seemed to be falling apart, I wanted a gut check. Because at the same time that token usage was soaring, a lot of other things were happening in AI land that were starting to give me pause. Stocks for the biggest AI investors like Microsoft and NVIDIA are wobbling.

Private capital firms that have gotten into data center construction are coming under pressure. I wanted to know, am I seeing the full landscape here? What am I missing? Last year, we brought investor and writer Paul Kadrowski onto this show to talk about his case for AI being a bubble. And it was one of the most popular episodes of this show that we ever published. Paul convinced me at the time that he was right. So last week I called Paul and I said, Try to convince me again.

That is today's conversation. This is This episode of Plain English is presented by Audi. We all know that feeling: a change of plans, a new opportunity. Instead of overthinking, what if you just said yes? With the all-new Audi Q3, the answer is easy. It's made for the yes life, with the power and room to handle whatever pops up. Yes to adventure, yes to right now. Because saying yes without hesitation, that's real luxury. The all new Audi Q3, made for the yes life. Learn more at Audiusa.com.

Paul Kodrowski, welcome back to the show. Hey Derek, good to be back. You were s with six months. I can't quite believe that. I can't believe it's been six months. So in September, you came on the show and you explained to me why you think AI is a bubble. And you utterly convinced me. And I talked about this interview so often I drove people completely crazy.

I didn't talk for a while and then I changed my mind a little bit. And I recently thought, you know, maybe I'm getting a little bit too optimistic about this technology. Maybe it's time to bring fallback into my life. Right. I need a new dose. to set me straight on AI. So before we go back and forth on who's right and who's wrong and let's wind back the clock a little bit. I I would love for my own sake and for listeners' sake and maybe for yours too.

Remind me, what is in a nutshell the Paul Kodrowski thesis for why AI is a Because it's AS a bubble because it's one of the probably five largest CapEx bubbles when meaning that we're at this moment where we're building out infrastructure like canals, like railroads, like rural electrification, like fiber optics, where we're building out this huge new substrate on top of which a lot of economic activity happens.

And this is a particularly large example of that, to the point that it's, you know, and this is part of what got me interested, where it was it's a material fraction of GDP growth, like 50 to 80 percent, depending on the quarter and whose numbers you use. And these things, inevitably.

And with a series of sort of rotating crashes as the assets themselves we overbuild and the assets become well, they might be in the useful in the long run. They uh are unable to pay the to pay their way with respect to the debt that's used often to finance them. And then there's a big reset and then we maybe find uh another use for them. This happened with fiber, this happened with rural electricity, this happened with railroads, this happened with canals. And if it didn't happen this time.

it would truly be the first time in modern economic history, which isn't the same thing as saying that AI itself is somehow uh you know frivolous, useless or anything else. No, AI is an incredibly important technology. And I I really have to drive this point home with people that saying that we're in a an infrastructure bobble that will have a whole series of consequences is not the same thing as saying that, you know, these whole large language model things.

You may think they work, but they don't work, or they're just auto-complete or whatever else. These are two very different arguments. And so that's my argument. And in particular, because this this particular moment is unique.

Because it combines all of the things that we found in prior bubbles, meaning loose credit, real estate, technology, government policy. We've never had A moment like this with a huge infrastructure build out where we were at the intersection of all four of the things that have caused the most consequential infrastructure bubbles in US history. So you have all these independent actors in each of these spaces.

all feeling like I don't know what those guys are doing in technology, but we here in real estate, we know what we're doing. And so whenever we sign a lease contract to a hyperscaler, we know we're looking at a prime credit who's good for the debt and whatever else. So every one of the actors feels like they're acting in a rational way.

And the consequence of that is what the finance theorists call a rational bubble, where you have all the intersection of all these rational actors is something that's economically indefensible. And so that's at the core, that's basically the core of my argument, that if we didn't have something like this. happen now would be the first time in economic history in this particular moment, is particularly fragile because it's at the intersection of all of these importance.

And the analogy that you've given so often is the railroads. That's not the only infrastructure bubble that you listed. There are the canals in the eighteen twenties, eighteen thirties, there's the fiber optic build out in the nineteen nineties, early two thousands. But it's the railroads that I feel like you keep coming back to, not only because the Transcontinentals were such an enormous infrastructure project.

But also because it's an infrastructure project that utterly transformed American politics and economics in the back half of the nineteenth century. Can you before we s sort of zoom up to twenty twenty six, Just pause a little bit and what what was the lesson to your mind of the railroads?

That you can have uh an entirely defensible infrastructure build out, railroads were a very good idea. They were not uh you know, ephemeral, they weren't like beanie babies, they weren't some kind of uh you know, strange and transient thing. They were really important and on top of which we did a lot of important ine uh economic activity, not least of which was settling the western United States.

But that didn't prevent people from overfunding startup, if you will, railroads and such such that track miles that were built in the peak periods in the mid-19th century. Uh, roughly half of those track miles were eventually abandoned. Does that mean that railroads were a bad idea? No, we just wildly overbuilt because of the impulse to build. was so imperative that the everyone who was who was building up railroads felt like there's an opportunity to be an oligopolis.

There's an opportunity when all of this shakes out, I'll be the consolidator, which is a very similar to jump forward for one second impulse as we hear today. I mean Dario Amade said this the other day that we think there'll only be one or two, maybe three players at the end of all of this. Same Impulse And it leads to this kind of uh misaligned uh incentives with respect to I don't mind that it's what I'm doing right now isn't paying off because over time I plan to be

The consolidator and the oligopolis. So that's lesson number one. And the other lesson, and and I think it's particularly striking, is people have forgotten. Two things about the railroad build up. Not just that there was all this redundancy, but it led to a series of financial crashes in the eighteen seventies, crash of seventy three, crash of seventy eight, crash of eighty seven.

A whole series of crashes, which each of which killed off a significant number of companies, financial institutions, and others. So it had financial consequences, not just 50 years later, but in the period after build-out. But nevertheless, in 1900, the railroads were roughly 60, 62 percent of the uh index market capitalization in the United States. They were the technology company of their time. So building out that platform.

was rewarded, but was also consequential in terms of leading to various financial crises and it also played a secondary role in the Great Depression itself. None of these things mean that railroads were a bad idea. Railroads were a tremendous idea. But the carnage along the way was dramatic. And so I love the metaphor for me is that you can have a hugely valuable build-out.

But it is really consequential for decades in terms of both the productivity consequences and the economic and financial carnage. So strikingly today, technology uh writ large is around sixty percent of uh

all in of the all um all US index so the MCSI. So you were kind of in a similar situation which is sort of striking that this this industry has grown to remarkable dominance. of the broader equity indices, much like the railroads did. And much like the railroads did at the time, it became increas it's become increasingly capital intensive, which has consequences in terms of how investors are are looking at technology companies now versus how they looked at them.

in the Halcyon days of the nineteen seventies. So that's kind of the bridge for me. It's a platform. Many of the impulses are the same, many of the consequences are the same. And I feel, you know, it would be very hard historically not to see not to think it will play out in a similar fashion. For folks interested in going a little bit deeper on the railroad analogy, we did a podcast with uh Richard White, the Stanford historian who wrote a wonderful book called Railroaded. And his thesis

Is that it's a tremendous book. It was a fantasti I loved the interview, another one that I think about all the time. And his thesis is that the Transcontinental Railroad was a technology built by corrupt. Idiots working in concert with craven, horrible politicians, which led to one depression after depression after depression, panic after panic, inflation, deflation, crashes of the economy, and

It was a good idea and it completely transformed the country. And so this idea that one has to choose between, is it a bubble? Is it built by idiots? Do I like the folks building this technology? Is it important? You can have yes, no, no, yes answers to these questions because they are profoundly different questions. And sometimes American history doesn't give us the simple morality story of bad people build bad things, good people build bad things. Hundred percent.

Sometimes bad people um on in the process of crashing the US economy over and over again and ruining people's lives, nonetheless build technologies that in the long run we can't imagine modern life without. It's a just just it's a strange um a moral, uh anti moral uh feature of history.

Um a lot's changed since we spoke. Okay. Yeah. Um well maybe this might be the last agreement that we that we have. So maybe let's maybe let's savor it here. Um a lot's changed since twenty twenty five uh when we spoke. Uh and I think some of those changes validate your prediction.

And I think th some of those changes complicate your prediction that AI is a bubble. Um so first, before we talk about those changes, maybe I'll just like get your thesis on the record. Do you still think that AI is an industrial bubble right now? Yeah, so yes, AI's a a AI is a bubble, there's no no question.

Let's go through some of the events that you predicted. In twenty twenty four and twenty twenty five, the Mag Seven stocks, Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla. Completely dominated equity returns. I think at some point Michael Sembalist at JP Morgan was saying it was like 65 to 75 percent of the total SP was just riding on the back of these seven stocks. That's changed. Over the first two months of twenty twenty seven. Not only did he say it it was true.

Yes, right. Yes. Yeah, exactly. Right. This is this is someone representing objective fact. This is my last... two months it's it's a different picture. Um and and this is also math. Uh the S P five hundred was up about point five percent for the first two months of this year. The Max seven were down five percent. All seven are negative year to date as of this morning. Microsoft down thirteen percent, NVIDIA down three percent, everyone else down somewhere in between.

What happened to the hyperscalers and why might this hyperscalers and why might this be indicative of your thesis? Uh okay. So the the important point is too. One, obviously there's been this reversal. Two, these when you talk about their contribution to something like the S P five hundred, it's market cap weighted, obviously. So you need to think about it that these are very large market cap companies.

So it's even more consequential when they decline. So Microsoft itself has probably cost the SP, you know, like eight to ten points. No, but probably more than that by now. So these are very consequential. So it's not just that a basket I have somehow magically found a basket of stocks that are declining, because who honestly cares? These are some of the most consequential companies.

at the vanguard of what's going on, who are truly barometers of this build out in both the sense of being builders and and and at the other side users of what's being done. And the market has had a change of heart. The market's change of heart is that

And I've shown this. I was I have a a graph that I often show people that two years ago, uh the analogy I make is it's kind of like crypto treasury companies. Early on in crypto treasury, you were rewarded. If I bought a dollar of crypto, I added two dollars to my market cap. It was like a magic money machine. Exact same phenomenon happened with AI CapEx two years ago. If I add a dollar of AI capex, the market rewarded me with two dollars of market cap. It was like

There is no reason to stop. This is a flywheel. I'm I'm gonna get rich. I've I'm a I'm a hyperscaler. I continue to spend on this stuff, and yet the market remote rewards me for more than I spent, and yet there's no cash flows associated with it yet. And that's fine, the market's entitled to do that. About six months ago, around the time we spoke, probably, it became neutral.

So in other words, whenever a dollar of market of AI cap expanding was no longer rewarded with an a dollar of market cap. That's n that's interesting. Now the markets, the market's sort of this. unthinking machine kind of wandering around and trying to make a discount information decide how consequential it is. And it came to this kind of c organic conclusion, you know what? I'm not as excited about that anymore.

And then in around October, November of last year, it reversed completely. So when you announced AI CapEx, that I'm Oracle and I announced that I'm, you know, I'm making huge investments in Some of the most massive projects on earth, suddenly the market started taking away market cap. So this is a very natural process you see in capital markets where they go from initial euphoria.

to okay I'm increasingly neutral about this to this reversal. So this is what you're seeing. So this this mag 7 reversal is a capital market manifestation of this gestalt shift, this change of heart in the markets.

about how we should value AI capex going forward in terms of the multiplier we should associate with it. And that multiplier is increasingly negative. And so just because Mag 7 is was coined after an old Western, the Magnificent Seven, I flipped it and I said, okay, then now they hate the eight.

Um so the Hateful Eight the Quentin Tarantino film. So if you add in Oracle, who can't, you know, bear to stay away from any bad financial ideas, they've added themselves to this. So now Oracle is part of the mix. And so the group is the Hateful Eight. And they're dragging down the SP. And now that can sound really, really glib, but the point is it's consequential because these are the same group who are being celebrated all the way up as.

Avatars of this new build-out and look how the market's rewarding them. And that is all now reversed, which was exactly what we talked about.

six months ago, which I said was in inevitable was going to happen. And you know, there's all lots of other pieces too, like I was pointing to at the time. The increasing role of debt in this process. So this idea that more and I want you to get to dead. I want you to get to dead in a second. Yep. Just taking everything that you just said. I think someone hearing that might say, Okay, first the hyperscalers were rewarded

for their AI investments. You put a dollar in, you get two dollars out, and AI investments went up and up and up. And then the logic shifted and you were punished for your investments. And therefore, I think someone might think, AI infrastructure spending went down. That doesn't seem to be happening.

What seems to be happening right now is that the infrastructure spending is continuing to rise, even in the face of the market suggesting that they're going to punish your equity value based on that rise. What sense does that make to you?

Complete. I mean that's capital markets at work. The idea that I'm no longer being rewarded for it, think of it in game theoretic terms. If I'm the first one to announce that I'm no longer spending on AI CapEx, what's gonna happen to me in the market? The market's gonna say,

Well, you just fold it up. You've decided you're not going to be a player going forward in this possible future oligopoly that Dario thinks he's helping to instigate. So you will not be part of that. You've essentially said, I'm out. The market's not going to reward you for that.

In the way that they rewarded you for say the old Capex spending, but it will punish you for that because it'll say you no longer are a player. So you have to think of it in almost game theory terms that you have no choice, even though it has negative value. The next piece is obviously this idea that more and more of so the argument that people make is the hyperscalers are big boys and girls.

They're very profitable. Who are you to tell them how to spend their money? Right? I get that all the time. And it's like, well, yeah, that's true, but in and it's an increasing fraction of their free cash flow is being devoted now To this thing we call AI CapEx, depending on who you look at, it's more than half of free cash flow for some of the largest hyperscalers. Now, they say, well, that still leaves a huge amount of money on the table that it there's lots more spending to come from.

People have aren't being very sophisticated in terms of the way they think about what's happening. Most of that, a great deal of that money is now goes to other places. Like for example, tech companies love to pay stock-based compensation, SBC is the term of art. And the way you do stock-based compensation, obviously, is I give employees stock.

My investors aren't very happy about that because that dilutes them. It increases the number of shares outstanding and suddenly my earnings per share look worse, stock might decline. So the way they neutralize that, they neutralize stock-based compensation is through buybacks. We hear about buybacks all the time.

So, buybacks historically were a huge chunk of free cash flow as a way of neutralizing stock-based compensation. Well, that's no longer possible because you can't spend at the same levels to do these buybacks. So now we have a double whammy of problems. Not only is the free cash flow increasingly being eaten by AA CapEx, I'm being diluted as an investor because the company is no longer spending as aggressively.

To buy back shares and make sure that stock-based compensation is no longer diluted. So the rational answer would be, well, if I was a technology company, I would just stop paying so much stock-based compensation. That's not happening.

Oracle, uh, Microsoft, Google, they're all paying as much stock-based compensation as ever. And as a matter of fact, we're increasingly using debt to do some of the buyback of the stock at a very much more modest level where they might have been doing six or eight billion dollars a year.

in stock buybacks a decade ago, Oracle last quarter we did like, I don't know, eight hundred million. I mean just a fraction, a pittance compared to how much is going out the door with respect to dilution. So this idea that these are just magic money printing machines and they can do whatever they want because it has no consequence is profoundly misleading because the the cash flow, the cash is not just sitting there burning a hole in their pockets. It's being used for all kinds of purposes.

Not least of which is in a sense inoculating their shares against all of this stock based compensation. So all of this is interlocking. And so what's one of the reasons why you're seeing these companies share prices now. The their multiples declined. There's a convergence between the forward multiple of the Mag 7 and the S P 500. They've largely converged. Microsoft's forward multiple is now roughly equivalent to Exxon.

What's that telling you? It's telling you that the market increasingly looks at these companies as very CapEC and intensive companies, and most of their cash flow is spoken for. And so this is constant, so this this is what we're really seeing. So it's not it's glib to just say, well, they're very profitable. They should be able, they should they they they're good for it.

It's not like that anymore. So now what's happened is an increasing fraction of the spending is coming from debt. A big chunk of that is coming from private credit. Uh and private credit in turn is now at r is now under pressure for a bunch of different reasons, not least of which is their exposure to technology.

Now, it's a backdoor exposure because what's happened is that private credit was very aggressive in the 2021 to 23 period, buying up shares and software as a services companies. Well, guess what's happened? We've had this SaaSpocalypse as this AI thing increasingly makes SaaS companies look at risk. And so now their holdings in these SAS, the private credit holdings in SaaS.

Look much less valuable at the exact same time as their exposure to data centers looks over large. So now it's we've had this race for the exits in private credit. Blue Al briefly had to shut down redemptions. Um I saw that today was at Apollo.

Um, or Blackstone was was was gating and would uh making it less less easy and less frequent for people to do to reach. These are all the things that were like six months ago when I said this, I was like, here's what's coming. People like, ah, it's it doesn't matter, it's not that much debt. Uh these companies have huge cash flow.

And I was giving a talk to a group of investors two weeks ago. And like, yeah, yeah, yeah, this was obvious. I'm like, dude, where were you six months ago, Mr. Obvious? So this is the way these things tend to play out. It goes from being completely impossible to to obvious.

Yeah, you're you're gonna have an opportunity to talk a little bit more about the SAS pocalypse in just a second, because I think I have a a a a frame to go deeper there. I want to reflect back on what you just said and put it in historical terms, people understanding maybe a distinction between the railroad age and the AI age. In the railroad age, the companies building the transcontinental railroad were often essentially startups.

They would take on enormous amounts of debt, often government debt. They would build and build and sometimes they would build these roads, railroads, to nowhere and then they would go belly up and they would take down a bank and we'd have the panic of, you know, whatever, eighteen seventy three, eighteen ninety three.

So the one thing that folks who defend AI in the twenty twenties pointed out is that AI in this respect is the exact opposite, because the companies building AI are not startups taking on government debt. They are the richest companies in the history of capitalism, right? Meta and Alphabet and Amazon or incredibly well-funded companies like the Frontier Labs, OpenAI, and Anthropic.

These are not in any way equivalent to the railroad startups of the 1870s. So I don't want to hear it about that analogy. You're pointing out that over time, the amount of AI spending that is necessary to keep up with the Joneses in this hyperscaler space.

is getting higher and higher and it's eating a larger and larger share of free cash flow. And we're arriving at the point now where something has to change. You simply cannot continue to increase AI infrastructure spending twenty percent year after year after year. And that's why you feel like, among other reasons, we're in this inflection point. It's not that we're going all the way back to today's AI builders or just like the railroad builders of the eighteen seventies. It's that the advantage

that they had over the last two years is not the advantage they'll have for the next few years. Is something like that a fair recapitulation? Yeah, I think I think that's a big piece of it, but it's even more insidious than that. So because they're not like the startups of the days of your of the old railroads, people feel like they're justified in extending them even more credit because these are prime credit.

They're like, wait a minute, they aren't gonna go broke. So the reaction is the exact opposite. Because they're not gonna go broke, people extend them more credit, and they're extending more credit to some of the most aggressive and strategic negotiators on earth. So guess what's gonna happen whenever prices change or I decide that you know, I I actually the efficiency gain suggests I don't need nearly as much um inference capacity as I once did.

They're going to be some of the most aggressive counterparties you've ever met because they have the cash flow, the balance sheet, and the resources to fight you to the death. So the idea that just because they won't gonna go broke means that I now have an asymmetric advantage in my contractual relationship with these people is boggles my mind because it's the exact reverse problem. It's because they're such prime credits that we create a new class of problems.

And that new class of problems is will play out because you're negotiating with a small cohort. of incredibly aggressive strategic counterparties who are perfectly happy to to to bleed you deb bleed you out while they s continue to fight you about what their, you know, four year renewal rate is on a data center. So it's it's the exact opposite problem.

Mm-hmm. I want to drill down a little deeper on who is losing at this stage in the business cycle and where value seems to be flowing. So in the losers column, and I think this is a door that you opened, you know, in big fat letters, you've got the SAS pocalyp. Go a little bit deeper in explaining what is the SAS pocalypse and why do you think it's relevant to the bubble question?

Okay, so one of the ways to think about what's happening right now is that we're having the emergence for the first time in a hundred years of a new industrial commodity. Okay. And this we think back to copper, coal, wood. These are all fundamental industrial commodities on top of which all this economic activity gets built.

The new industrial commodity that's emerging on the back of what's going on here isn't like large language models. That's not a tradable commodity. It's this thing called tokens, right? So one way to think about what's going on in terms of trying to s establish the playing field of you know who gets hurt and who doesn't is to think about this new commodity called tokens and what it means for people whose businesses essentially ingest tokens, digital information, symbols.

And produce tokens, right? Software code, both both. So think about companies that are in that business. Now this fundamental commodity that they uh that they ride on top of, it's price. is in a deflationary spiral. Okay, this is what's unusual. There's a token emerging called or a token emerging, a commodity emerging called tokens. Tokens are in a deflationary spiral. They decline on the order of

Last five years, 70 to 90 percent year over year. That's not like copper, that's not like uranium, that's not like wood, that's not like coal. This is a very unusual commodity. So what the market is sensing with the Saspoon.

And it's in this very blind, unthinking way, because that's the way markets work, is that when you brush up against this deflationary force called tokens, on top of which rides AI, you're in big trouble because suddenly the economics of your business were just transformed. And that's why We've seen this in asset management, another digital business. We've seen this in insurance. We've seen this in some aspects of e commerce.

It's not that the market's anointing winners. It's saying that these are all industries that are brushing up against a deflationary commodity, a hyper deflationary commodity called tokens. And I'm smart enough to realize, says the market, that that's not good. And that's not good because both it lowers the barriers to entry, people can compete with you more easily, and your economics are radically changed in terms of your moat and your ability to defend your margins. And so the SAS pocalyp.

Is the market in an unthinking way wandering around and saying, I have this new thing. It's called a token. Who's it gonna hurt? Who's it gonna help? I have no idea who it's gonna help. They've kind of given up on the whole helping thing. They've really focused in on I see now all the people this could hurt because it low reduces moats, lowers barrier to entry, and changes the economics of the business. So the SASPOCLYP.

In a sense, you can think about it in that in those terms, the emergence of this new deflationary commodity. So people get hung up on the idea that like. You know, no one's gonna build their own Salesforce.com. That's a pain in the ass. I don't wanna manage that. The reason why we have SaaS companies, I always argue, is that so that I don't have to do it and I have someone to shout at when it doesn't work, right? That's the whole reason that SaaS exists.

Which is sort of true, but it's beside the point that many companies are spending ten to twenty million dollars a year on renewal contracts for large SaaS companies and they're like, you know what? I can get away with a little less shouting if I can save ten or twenty million a year. So that's the pressure, but the pressure

to really be more thoughtful about it is driven by the emergence of this new commodity we call tokens. It's just I think it's important not to get hung up on the idea that someone's gonna vibe code their way to Salesforce. That's not the point.

So value is leaving software as a service. Right. Talk a little bit about where you see it flowing in markets. Because, you know, you I I take your point.

Aaron Ross Powell That markets are going around. I think the metaphor that you gave me in our email is they're like slime mold exploring energy gradients. They're going around trying to sniff out who can possibly be hurt by the commoditization of this new thing called tokens. I understand that You think that this phenomenon, you s understand why investors are crushing software as a service company.

But I wanted to spend a little bit of time talking about where value is flowing. So one place where value might be flowing, at least as a dumb look. Yeah. uh markets is energy. If you look at various sort of, you know, spider funds, the energy select sector spider fund dramatically outperformed the tech select spider fund by like twenty, thirty points in the last few months. So

I wonder maybe we can start with energy and go to other places where you see value flowing. Cause I know that I know that you have some places to go that aren't just energy. But what does it say about where we are in the data center AI cycle that energy stocks are on this tear? So you have to be careful with energy because obviously it's endogenous to the system, so it's both

benefiting from it as well as fueling it. So the notion that you can somehow separate the recent run of energy from this incredible demand to build out more energy supply to feed data centers is very tricky, right? Because we're seeing the same thing in with respect to

um transmission systems, we're saying the same thing with respect to transformers. All of these are benefiting and we'll tell you this on earnings calls that they're benefiting in a huge way. So having said that, we've underinvested historically dramatically in energy systems in the United States.

um, probably on the order of two to three X underspend in terms of what we should have historically. So this has been a trigger to spend more. Unfortunately, because so much of it is happening so fast, a lot of it's becoming

behind the meter spending where it's like, I don't care what you do, but you gotta bring your own power. Well you tell someone you don't care what they do and they have to bring their own power, they're not gonna build a river. They're not gonna suddenly have a hydroelectric dam because that's not available. They show up with natural gas. Right.

And they here just slow down a little bit on behind the meter.'Cause when you say they, you're talking in many cases about the hyperscalers essentially. Right. Right. Using overall. Right. Using more private energy generating technology and energy generating facilities to power their data centers rather than hooking into the grid. That's that's the idea.

Which is bec in part because A the utilities can't keep up with them because the demands are growing so quickly, but also because there's been a kind of consumer uprising. Uh the pitchforks are out about increasing

uh at consumer energy prices across the United States and the data's very clear. I mean the uh the EIA was out with a report yesterday, the day before, showing the dramatic spike since twenty twenty two. This is not coincidental and consumers see it. And so it's like I don't, you know, this data center thing.

make it go away because I can I'm smart enough to see that it has a connection to my higher utility bills. Don't try to tell me that it doesn't. And of course the answer is it does. And so then of course the reaction in the industry is to say, fine, I'll bring my own power, this idea behind the meter or whatever else, and I'll show up uh with uh natural gas plants. Okay, and the those of course have long cycles associated with them.

Um, they also have long lifespans, so they risk obsolescence and being stranded. But nevertheless, that's the secondary unanticipated consequences of what of what's happening. So it energy has been a beneficiary and we're accidentally building out parts of the grid that should have long ago been improved given in the increasing electrification of the US economy. But you have to be careful.

to not sort of turn the inputs back into outputs again and say, oh, well, energy is doing really well. And that's how separate from what's happening in data centers. And it's not, it's largely 60 to 70% of growth in the last couple of years and earnings per share in energy can be directly attached to the data center buildup.

Couple questions before we get to some places of potential disagreement. Um I wrote a column a few days ago, maybe a week ago. Um the title of which was Nobody Knows Anything, which is a famous William Goldman line from an autobiography that he wrote um about his career in Hollywood and the ideas that, you know, people who try to predict hit

pick hits in Hollywood. They don't know what they're talking about. No one knows what the next hit's going to be. And I said that nobody knows anything is my mantra for the effect of artificial intelligence on the macroeconomy. In fact, you know, you look at the the the the objective fact that the Citrini essay, that famous science fiction story about the future, moved markets by several hundred billion to up to a trillion dollars.

To me, it is a sign of just how starved investors are for any actual data on the macroeconomics of AI. They are so starved. for actual information that they will let a science fiction story lead them to buy and sell equities to move markets at at this rate. This is true. True. You you told me in in our back and forth over email something similar. You said you think this moment, this moment in the business cycle, is badly misunderstood.

What do you think is so badly misunderstood about this moment by investors or commentators?

How long do you have? Let's do five minutes. This is like a well you're in you you're down you you've flown into the buzz saw. So there's a host of things. I mean one of course is we just saw it yesterday. The new productivity data came out. And I'm sure you saw it. So it showed uh I we're up around two, two point eight percent. We're seeing productivity growth, we're seeing levels of productivity we haven't seen in a decade. And of course the immediate reaction is

you go AI, right? That this is a untirely an AI phenomenon. And this sort of thing a combination of motivated motivated reasoning and enumeracy drives me nuts. So the of course, the reason why And I blame a host of people for getting this wrong. The reason why, of course, productivity is rising has nothing to do with AI. AI may cause productivity to go 10% in five years. I have no idea. But it's not doing it now. What's happening now is mathematically GDP is is an is uh

Uh, the additive product of consumption, government spending, investment, and exports minus imports. Most of those are flat lending, except for one item. One item's gone parabolic. And you know, I'll give you five guesses. Yeah, it's the eye. The eye's going straight up and almost all of the eye is one thing.

AI CapEx data centers. So that's causing GDP to increase because of all of this CapEx spending, but hours worked aren't increasing. So what presto magic, guess what happens to productivity? Productivity soars. I saw a report yesterday from a cell site analyst thing. It's a mystery. I said it's only a mystery if you can't do math. I mean, there's no mystery here. This is exactly what's happening. And we're doing ourselves a disservice because when you start promoting the idea

That it's all AI, you're you're misunderstanding the nature of how economies even work. You're have you're turning this thing around in such a way that The more we spend on on AI, the better off we're doing. We should just spend everything on AI. Take government spending to zero, consumer spending. It's just an arithmetic equality that as I increase investment, if hours work stay constant, productivity increases.

But doesn't mean we're at all more productive. It just means we're spending a lot of money on AI, right? So these kinds of things are endemic right now because of this combination of sort of enumeracy and motivated reasoning. Another example I'll give you is that

Mm there's been this idea in people's heads that we had an inflection point in terms of the capability of AI, that a year or so ago in the dark days, models were coming out, and you know, it was a little bit better, but it wasn't that much better. And I could kind of not tell the difference between the models before and the models since. And there's this feeling in the last four or five months that it's all changed.

Right. That everything all of a sudden I have people tell me all the time there's like an unlock key that all of a sudden I've unlocked all this productivity. And of course what they're really referring to are these things sometimes called orchestration layers or harnesses, Claude Code, Codecs, these kinds of tools.

And they sit on top of the models and in a sense abstract them away. The analogy I make is they're like a really effective nanny managing bratty kids, right? So the bratty kids are the models.

They're kind of yapping, but they're really smart. They're just but they're doing their own thing and you have a really effective nanny sitting on top, orchestrating them, organizing them, and saying, you know, sit down and be quiet, do this, do that next, whatever else. And that orchestration was a piece that was missing.

So having said all of that, does that mean we've restarted the growth of large language models in terms of them becoming much more effective again? Has the argument that the scaling laws were beginning to fall apart dropped away? Of course not.

We're operating in a different layer of abstraction. The layer of abstraction is we're now organizing those units better, but the models themselves are still asymptoting and not improving as quickly as they were, say, you know, four or five years ago. It's not that they're not improving, it's just we're reaching a natural This is one place that I wanted to push back a little bit and hear what you have to say.

You I saw you've written in several places that you think model improvement is slowing down. Um Well not just think it's empirically true. The the the composite benchmarks show it. So I'm interested to know what benchmarks you're looking at. Because if you look at, say, the nonprofit organization meter, which measures the kind of tasks that large language models can complete. Yeah.

According to that's not that's not just according to a reading of the graph. That's what the the engineers there say. And so, you know, meter is not i meter is not God. Um but it seems Also a very squishy benchmark. I have a real I have real trouble with it because people misunderstand the nature of what the meter benchmarks are actually uh showing, which is to say

how long of a task a t humans estimate how long it will take them to do a task and we estimate how long it will take. This is squishy at both ends, right? It's squishy at both ends. This is a very difficult benchmark. To build your house on top of. I'm not saying it's useless, but I'm also not saying it's particularly useful because of the nature of the subjectivity at both ends. This is and this writ large is the trouble with most of the benchmarks, right? Most of these benchmarks.

have a huge number of problems. And so one of the things I rely on the most, I mean Epic and others does this, EPOCH dot AI, those guys, but there are others, this idea of composite benchmarks that I'll I'll I'll construct my benchmarks from a basket of other benchmarks rather than relying on a single one.

And I think that's better and a little more defensible, but a deep problem right now is that an increasing fraction of benchmarks have been ingested by the models themselves. And so the trouble is that almost all of the benchmarks are just a decent test of how effectively they've ingested the benchmark. And that doesn't matter.

Broadly known as Goodhart's Law or a version of Goodhart's Law. When a measure becomes a target it ceases to be a measure. Uh right. Right. I I take that I I take that as a problem of of all benchmarks, to be clear.

I just I wasn't as familiar with the argument that the benchmarks clearly showed a slowdown in model progress because most of the benchmarks that I spend the most time looking at just I I I I don't I don't know if if they're objectively the best ones, though just the ones that I look at.

Most of them seem to show a speed up since the summer of twenty twenty five rather than a slowdown. And so I I was surprised that you seemed as confident as you are that that there's been a um uh a stagnation in model improvement. It's just maybe it's just a difference of opinion.

Well, it's a difference of opinion, but there's actually data. So I I for me, if I let's pick the epic composite benchmark as an example, in the period there's been a series For folks at home, that's epic A E-P-O-C-H, in case you want to do that. Right. Yeah. And again, this is that none of this is there's lots of places you can go for you know, pick your own benchmark and come up with your own argument, but at least it's a composite benchmark, so it's defensible in the sense that

it works across a number of them. And so what I like to show in that was that there's been a series of regimes, if you will, that in the 2022 to 23 period, we were seeing roughly on composite benchmarks 12 plus percent year over year improvement, actually even a little higher than that. That dropped to around five to six percent in the twenty three-24 period and has since fallen to around two to three percent. So

This is this this is that. This is the idea. The models haven't stopped improving. It's just that the sharp pace of improvement has has has dropped off at the exact same time as training cycles are longer longer and costs of increase. So, from an economic standpoint, how much should I be willing to spend for a 2% improvement? How much should I be willing to spend for a 12% improvement? If consumers can't tell the difference,

Users can't tell the difference. How much should I be willing to spend on that versus orchestration layers versus an improvement in clawed code or codecs? And all I'm suggesting is we have this. And it's a very classic human idea that just because something's continuing to improve, that that means that it's it's improving in as consequential a way as it was a da uh five or six years ago. And that's simply not the case.

That we don't see of emergent properties. No one's all of a sudden saying, Oh, look, it's learned how to speak French or something else. These were the things that were falling out of the model as emergent properties can arrived because the models were improving so quickly. Now it's, you know.

A couple of gallons of gasoline and a you know billowing dust as they move along. And the thing that's made models much more effective are these harness layers, which are incredibly consequential, to the point that I have argued that it wouldn't surprise me over the next few years.

To see a model vendor say, I'm going to stop spending on models and spend exclusively on these orchestration layers as they move up the software stack, much like moving from semiconductors, large language models, to operating systems. Windows versus cloud code and whatever else, up towards applications. The value is gonna move increasingly up the stack given how incredible the spending has to be at the bottom of the stack for modest improvement.

Yeah. So let's talk a little bit about why I started to change my mind in the last few months. Um I remember when Cloud Code came out, uh which is an anthropic product and Codex, which is uh a similar product from OpenAI. When it came out and, you know, I was on parental leave. So, you know, I'm going on like three and a half hours of sleep. So it's possible I wasn't operating in my absolute uh maximal capacity.

But the gut feeling that I had in that moment, really even before I saw the amount of money that Anthropic has made in the last six months uh from this technology. The gut feeling that I had in my moment was These new autonomous agents are really useful. They're really useful at manipulating information for a broad range of knowledge workers. And I can imagine now.

that the introduction of autonomous agents is going to lead to an enormous increase in inference or usage. More usage is going to mean more tokens. More tokens means more revenue. More revenue means a smaller gap between levels of spending and levels of revenue. The more revenue you make, the less likely you are to be in a bubble. And that the shift toward agents might even do

what you were just suggesting, which is maybe moderate some of the spending on new models and move some of the use of this AI infrastructure toward inference in a way that makes it less likely that we're in a bubble. Might even create the need to build more data centers, to have more GPUs, to run more inference, because so many knowledge workers, so many folks in the bicoller economy are gonna be using these tools essentially like

Steroidal Excel programs, the same way that everyone's got Excel open, like in some tab, if you're a certain kind of worker, that every the one hundred million knowledge economy workers are just going to have inference uh running constantly. And if that's thousands of dollars worth um for these workers, you multiply that by a hundred million uh knowledge workers and that is a lot of money and that's making me less confident.

that we're looking at something as clear as the railroads to nowhere in the middle of Kansas, rural Kansas. So you so th that's that's honestly that's not a particularly like sophisticated argument. It exists, I know, at the level of vibes, but um I've got a more technical argument for you in a second, and I'm interested sort of how you feel about that vibe-based argument that um Agents introdu put us in a new world of AI revenue that makes a bubble less likely.

So um I've said already that I think Clyde Code and Codecs are probably some of the most important technologies that have emerged because the orchestration layer was hugely important given what's the inability of models to do these. Agentic or serial tasks, because these are the tools that are managing all of this. So I agree on that level.

Here's the difference though, is that I think there's a profound misunderstanding of the nature of what causes tokens to be both demanded and produced. The typical white collar work, if you think about uh I'm gonna go XY axis on you. So if you think about an XY axis where coding's sort of sitting out here in the top right corner.

And its unique characteristic is it's it's deterministic, meaning there are right and wrong answers. I can't just screw around and say, oh, well, this subroutine is more aesthetically pleasing. I'll go with it. It's blue. There is no blue subroutine. It either works or it doesn't work, right? So it's more deterministic. But software has that characteristic, but it also has the characteristic of being more expansive.

Meaning that as I write code, it generates not just one for one more code, it writes a nonlinear amount of code. So in other words, it's got test routines, build routines, the software itself gets bigger. So coding sits in an unusual place in the market of

producing an ingesting tokens. It sits out here in the top right corner. White collar work sits in the bottom left corner, meaning that there often isn't a right answer. I'm in marketing. Is this PowerPoint better than that one? I don't know. What do my boss like, right? There's no it's not deterministic.

And and it's not particularly expansive. Think of the most common white collar applications you see, at least in the early days of AI, they tend to be what I call compressive applications, meaning that I got a giant report, I don't want to read it. Tell me what the bullets are. Tell me what the most important points are. This is compressive. I'm taking a large number of tokens and reducing to a small number of tokens. So the naive idea that because

Like coin. I've never heard that before. Yeah. No, keep going. Sorry, this is great. So the the idea that by bringing AI to white-collar workers, I can duplicate the rampant token consumption that we've seen happen out there in the land of agentic AI with coders is a misunderstanding of the nature of how this again, this commodity works.

And the nature of deterministic versus non-deterministic applications and compressive versus expansionary applications. White collar work is down here in this bottom corner and coding is way out there. So even if we have universal adoption. of AI tools among my colored workers, it will look nothing like what we're seeing already in terms of the incredible inference demand from coders themselves because of this fundamental difference.

Right, right. And an another way to put that or or another way that I'm putting it to myself is that if you look at a graph of um the increase in token usage um in the l in the months Since clawed code was invented, you're looking at this enormous run-up that is disproportionately concentrated among an occupation that uses so many more tokens than other occupations. So rather than

See that trend line and say now extrapolate over the one hundred million knowledge economy workers. That's the future of token usage. You're saying we might see that the early adopters of this technology. are simply a different kind of user of this technology, and you can't simply extend that line across the whole knowledge economy. That's a vi that's a that's a compelling point. I I like that. But I th that's a that's a nice that's a that's a nice argument.

Yeah, and I just think it's not only that, they're profoundly unrepresentative, would be the way I would put it. They're profoundly unrepresentative of the rest of the quote, real world of work. And not just because of the the personalities of the people doing these things, because of the nature of their work that it sits way out there in terms of both being deterministic. And determinism is really important. Why is it important in agentic flow?

It's important because if I unleash an agent, there has to be what engineers would call a gradient descent. There has to be this idea that as it flows, it knows truth from falsehood, that it knows, oh, that's working and that's not. Well, that works in coding. There is a gradient descent, there is determinism.

That doesn't work in marketing. That doesn't work in almost most non-trivial white collar applications. There's lots where it does. You know, I'm let's say I'm trying to balance books or I'm trying to deal with ingesting supplier data. But again, those, while they're more deterministic.

They're compressive. They're not expansive. I'm reducing the data and not producing more. I'm not, there's no build test cycles where I'm producing more and more code. So even in those areas that are more deterministic in white-collar work.

It again, the white uh software engineering and this thing that's drove driven this incredible inference explosion is so unrepresentative that it actually is misleading. Doesn't mean it's not going to work well in white-collar work, it just means it's the wrong metal. Right.

Who's more wedded to the idea that we're not in a bubble? Because even in my my comment that I saw you um you found uh before the this show, I said, you know, my my odds are down from like, you know, eighty to forty or something. I think someone who's more at at like ten percent. what we're going to see is a Jevons Paradox phenomenon. Jevons Paradox is famous idea that's banded about all the time in AI spaces that basically when a commodity

re is reduced in price, you don't have that sector lose value, you just use it more. I think it was invented about coal in England that when coal gets cheaper, you just don't use the same amount of coal the you know, um you you use much much for coal because it's cheaper.

I think what they would say, and I'm not particularly interested in in litigating this infinitely, but what they would say is in a world where tokens, as you say, are a commodity and they get cheap, cheap, cheap, we're going to come up with Task. that are incredibly token intensive because of the cheapness of those tokens. And so the knowledge workers of the future are going to be engaged in activities

that might be different than their current activities, different than t take this memo and make it one sentence. Um might be more token intensive than simply that kind of sort of Token cheap compression. It's a possibility, but at that point I'm in the realm of utter speculation of AI isn't a bubble because we we haven't imagined 10,000 jobs to the Right.

Assuming what you're trying to prove. I don't know. It doesn't matter. But I but I don't I I have more self respect to to hinge my entire argument on on that kind of case. Um here's here's here's a second thing that I read that I'd love to get your your brain on. So One of your more compelling ideas when we spoke six months ago had to do with the durable value of AI chips. And I'm just gonna sort of try to run through your your

case as I understand it to get to the meat of of of the contention here. You said, look, these companies are spending billions of dollars on GPUs and they're behaving as if these chips are going to retain their value for years.

But these chips aren't like we're back to the railroads. They're not like bars of steel. They're not something that you can run a train on for decades. They're closer to bananas. And this was a joke. But you were saying, you know, bananas brown. You have to buy constant constantly buy new ones. They're a little bit closer to bananas than steel. And so the problem these companies are gonna have

is that they're gonna have to keep replenishing their data centers with new expensive chips faster than their official depreciation schedules. And that means that in a in two years It's gonna or or so, it's gonna eat into earning so dramatically and destroy operating income and eventually they'll just have to slow way down. And if you slow this down and NVIDIA income starts to decline, you've got a huge So investors like Gavin Baker have written that he thinks we're seeing the opposite.

He said if you look at NVIDIA's H one hundred chips, which are about, I think, four years old at this point. Yes, they are. They've actually gotten more expensive to run over the last two months because of this phenomenon you and I are talking about, this this explosion in AI agents that are in such hot demand.

He said even six year old A one hundred chips, which were the previous generation, are still running at full capacity for some companies like Amazon, AWS, and those rental prices don't seem to have moved much. So that doesn't sound like bananas to me. That sounds like industrial machinery. It sounds a little more like steel. And it suggests that these companies might get a little more value or a lot more value from these chips. than we anticipated six months ago. How do you feel about that idea?

So um there's t at least three responses and so I'll try to bridge them all together. One is that it depends what the chip was used for in the power. So a chip with a longer lifespan tends not to have been used for training frontier models, or at least having that done for very long. The analogy I make is I can buy the same car with the same mileage. One was used only to drive a grandmother to church on Sundays. One was used for 24 hours of racing at Le Mans. Which one do I want? Same mileage.

That's a GP, that's a GPUs for training versus a GPU used for inference. So lifespan, mean time between failure and thermal load is very dependent on how the chip was used. So I'll say that straight up. that we're mixing apples and oranges to bring the fruit back. Um, whenever you just uniformly say the lifespan of chips is longer than expected. That's just a profoundly misleading way of thinking and characterizing about the population of GPUs out there.

It depends almost entirely on how it was previously used. Ones that were used largely for inference have longer lifespans. Second point is that NVIDIA has done a very good job of restraining or at least limiting the amount of new product coming to market, which is protecting of their 73% gross margins.

And so there's a kind of artificial scarcity to a limited degree going on here in part because of CUDA and part because of Nvidia's actions, but that's changing very, very fast. I'd encourage people to look. There's a company called Talas in Toronto, T-A-A-L-A-S. There's lots of others like this.

That are absolutely transformative in terms of sharply reduced power consumption. And like I think in the case of Talos, it's about 75X increasing tokens per second. And NVIDIA becomes roadkill when these kinds of companies are more broadly used. So we've got an incumbent who's kind of like a mini computer manufacturer in the old days of DEC and Data General, who's protecting their position as a host of new sh upstars show up.

That are going to price them into obsolescence. So the notion that they've briefly been able to protect their position through artificial scarcity and relying on uh the lower obsolescence and inference chips. That's a transient phenomenon. It has nothing to do with the likelihood of this of of uh H-100 chips having a lifespan of fifteen years versus two. No, that's just wrong.

Paul, can you just go a little bit deeper into the story? W why specifically do you think NVIDIA is is at risk of getting, as you say, priced into obsolescence? So NVIDIA is the beneficiary of this this training centric world that's emerged as we're training new models for at AI data centers.

Its chips are uniquely and it in in its recent move show this, uniquely w poorly positioned for inference. They aren't cheap. They often require cooling. And this is the reason why we're seeing a host of new companies, Celestra and others, uh Talas, I talked about earlier.

who are all emerging because they see that as the soft underbelly, that they can do all of these do inference faster with lower cost and with less cooling requirement. And it's very different architect very difficult architecturally for NVIDIA to make that shift. And so this is no different. We've seen this play out many, many times in the history of technology. We saw with mini computer mainframes to minis, minis to PCs, PCs to phones. This is what happens.

It's, you know, uh Clay Christensen talked about it endlessly in the innovator's dilemma. This is the innovator's dilemma writ large, is that there's all of these toy products coming along very, very fast. That NVIDIA can't sell and make its revenue nut, but these companies can sell and grow up into being very consequential threats with entirely different economics. They won't have 73% gross margins like NVIDIA does. They'll have much narrower gross margins.

but they'll eat an increasing share of the inference market. And so this is the dilemma that they put. the innovators dilemma that they put NVIDIA into. And then as a secondary effect, which is that because of the dynamics of how high the demand has been for GPUs and people try to protect their position in the market, you get this phenomenon of over ordering. Uh and early ordering to try and protect your position in the queue. If you want delivery in 2027, you better place an order with me now.

Um you want an order in 2028, you better pay two orders now, that kind of thing. And that's we know from the history of these kinds of moments, that unwinds ferociously whenever the wheels start to whenever the wheels come off, which is usually just a slowdown is all it takes.

And you could get that slow down just from the rise of these cheap inference uh silicons like like talus and others, because suddenly I don't have as much demand for NVIDIA to provide that. It's being eaten up in lots of other ways. And so there's all of these threats around NVIDIA right now.

It doesn't it sounds to me like you don't like NVIDIA. You think NVIDIA might be priced into obsolescence. I want to use that as a bridge to talk a little bit about um what you still think is coming down the pike. So what I'm taking from our conversation is, you know, you see spending rising as a share of free cash flow, which is gonna put pressure on earnings. You see tokens being commoditized. Roll the story out for me. What does a popping of a bubble look like to you right now?

I think it plays out in credit with companies like Blue Allen and others, which we've already begun to see happening. So that's part of what the popping looks like. I think it just is a sharp repricing of companies like NVIDIA because their order book turns out to be two and three X times overordering because people had to protect their position in line. So it's not just that they suddenly are growing 10% slower than people thought.

is that there's a having. We've seen this over and over again, back to fiber, back to electrification, that people pret buyers protect their positions in the order books by overordering. That is almost certainly happening inside of NVIDIA. And so when all of that unwinds, you see a halving in the companies, these suppliers, and that in turn plays out into power. If I'm a transmission supplier, transformer supplier, any of the uh people in HVAC, air conditioning, all of that begins to unwind.

Um, because it's all leveraged this one thing that's at the core of it, which is AI uh AI CapEx. And I think it plays out in that way. I also think it plays out in a perverse way. that token prices collapse even faster because this is people then want to sell token prices at a level that covers their fixed costs, even if they can't cover anything else. So they'll sell, they'll just dump them on the market, which will be even more consequential in terms of

the incursions that um this token commodity makes on the white collar workforce because they're gonna become cheaper faster than anyone expected, which will probably lead to more, you know, economic dislocation faster than people expected, but for a perverse reason, which is gonna be the equivalent of

you know, dumping oil on the market. Well, we're gonna be we're gonna be dumping tokens on the market. So I think it's gonna play out in all of these different pieces. And you know, Anthropic's done a good job. They had a piece out yesterday showing the most threatened work slices of the workforce in terms of where

You know, the biggest impact will be. And of course, the largest one is software itself. So I think that software itself is among the most threatened occupations as we look forward because AI is effective enough to do real damage to an occupation that employs more than what two and a half, three million people in the United States.

In the dot com bubble, there were all sorts of companies that were famously laid waste to that are part of the graveyard of two thousand one, two thousand two. The the pets dot com, the web vans. We talked a little bit about the companies that you see being most endangered by a bubble that you consider inevitable. The dot com bubble also created a whole host of enormous winners.

whose existence wouldn't have been made possible without the fiber optic build out. Amazon being a great example of a of a company that just clung on by the skin of its teeth, made it through the the crash and then became a multi trillion dollar firm. Netflix, great example of a company that could not have existed in nineteen ninety three but is worth

however many hundreds of billions of dollars now because its data is flowing through all that fiber optic cable whose uh constructors are, you know, now belly up in many cases. Um We've talked about who the losers are of the near future as you see it, but there's gonna be huge winners. You know, you believe as I do that this might be a bubble, but it's also a a significant and useful and potentially transformative technology. I'm not asking you to make...

specific stock picks here. I'm more asking what kind of the same way that someone might have said in two thousand one, yeah, it looks like the internet is effed for now, but trust me, a whole host of companies that essentially act I'll use the Ben Thompson word here, like aggregators of data and media are going to absolutely crush in the next 15 years. What are the kind of companies where you think value is going to flow to in the next few years or decades?

So I like the Halo idea, which is this notion of uh and it's a it's a ridiculous term, but heavy asset low obsolescence. I actually think it's gonna flow out of technology entirely. Um and the reason is, and this is we have this backwards way of thinking that because technology has always had new beneficiaries, it must have another cohort this time. No, it doesn't have to. It could just be eating itself.

The incumbent tech, Mag Seven, are becoming utilities. They're being valued accordingly. That's why this contraction in their price earnings multiples because they're being CapEx intensive with maintenance capital requirements. These are old companies who are becoming utilities, right? In the same way.

that we're also eating away with some of the the SaaS companies, that they're being eaten by this deflationary force. So it doesn't have to be the technology companies themselves that benefit. That could have been the anomaly. Just like railroads were sixty percent of the S of the indices back at the turn of the century. The anomaly was the tech got to this size, this outsized share of the economy, and things begin to flow back.

into, you know, manufacturing, into um I'm uh uh production of of tools for the grid and uh into things like even waste disposal. There's a host of different way places that are very asset intensive industries that have low obsolescence that I actually think and if you look at it year to date, if you took a cohort of of heavy asset low obsolescence companies and compared it to S P five hundred or mag seven, they're actually radically outperforming already.

Money's already flowing and saying, you know what? I want out. I all I see is this incredible deflationary force destroying the margins in this business called technology, while simultaneously the incumbents are busy turning themselves into utilities. I'm out. Software eats the world. Software becomes the world. Software eats itself. Jurassic Park. No, no, no, that's exactly right. I wish I'd said it, but that's exactly right.

All right, we'll we'll we'll leave it there. When when when I come up with a with a haiku to summarize a complicated point, I think that's when it's time to to quit while I'm ahead. Uh Paul Kodrowski, it's it's always a pleasure to talk to you. I really appreciate it. And um I'll talk to you in six months when I change my mind again and you have to talk to me off that ledge. Sleep before then.