Why OpenAI Won’t Survive an AI Crash
Summary
Paul Kudrowski, a venture capitalist and researcher, joins Henry Blodget to analyze the current AI investment frenzy. He frames it as a classic capital expenditure bubble, similar to historical infrastructure build-outs like railroads, rural electrification, and the dot-com fiber boom. The core dynamic is the massive spending on a foundational ‘fabric’—data centers and frontier AI models—which typically ends poorly for the builders but eventually enables tremendous innovation on top.
Kudrowski makes several bold predictions: he believes most major frontier AI models, including OpenAI, will be ‘commoditized’ and disappear, doubting there will be a ChatGPT-8. He argues NVIDIA is not a long-term winner, and the ‘Magnificent Seven’ tech stocks are unlikely beneficiaries. The central economic problem is collapsing ‘token’ prices—the atomic units of AI computation—which are falling 50-70% annually. This creates an impossible growth requirement for companies whose revenues are tied to this collapsing commodity, while their debt-funded capital expenditures remain fixed.
The discussion details how the bubble has metastasized through the economy. Data center-related capex constituted over half of U.S. GDP growth in early 2026, drawing in insurers, sovereign wealth funds, and regional governments offering subsidies. Kudrowski explains that the perceived demand driving this build-out is largely illusory: consumer chat (like ChatGPT) uses less than 5% of global inference capacity, while most traffic is for training models that are hitting architectural dead ends and becoming indistinguishable commodities.
Despite the predicted crash, Kudrowski is profoundly optimistic about the underlying technology. The collapse of token prices means cognition will become essentially free and ubiquitous. This will unlock breakthroughs in medicine, materials science, and robotics, acting as a ‘cheat code’ for human progress. It will also help offset severe demographic headwinds in developed economies. The episode concludes with a vision of the 2030s where AI compensates for an aging workforce and makes specialized knowledge accessible to all, though it will require a painful societal transition away from certain white-collar professions.
Topic Timeline
- 00:00:01 — Introduction: The AI Boom is a Bubble — Host Henry Blodget introduces guest Paul Kudrowski, a venture capitalist and independent analyst who believes the AI boom is a classic bubble. Kudrowski predicts OpenAI will not survive, NVIDIA is not a long-term winner, and the data center boom is drawing in a wide range of investors who will get hurt. He remains optimistic about the underlying technology and the societal benefits of cheap, ubiquitous cognition.
- 00:01:32 — Historical Parallels: AI as an Infrastructure Build-Out — Kudrowski explains this moment is a predictable infrastructure frenzy, similar to canals, railroads, rural electrification, and the dot-com fiber build-out. The key difference is the unprecedented speed and capital intensity. He notes this is the first U.S. ‘bubble’ at the intersection of four drivers: technology, real estate, loose credit, and government policy, creating a ‘rational bubble’ where all actors think they are behaving rationally.
- 00:04:50 — Predicted Winners and Losers in the AI Shakeout — Kudrowski identifies the likely losers: most frontier AI model companies, including OpenAI, which he believes will be commoditized and disappear. He argues the winners will be those building on top of the subsidized, collapsing-cost infrastructure, similar to how Netflix emerged after the telecom build-out. He is skeptical that any current ‘Mag Seven’ company, including NVIDIA, will be a long-term winner from this cycle.
- 00:08:07 — The Collapsing Economics of AI Tokens — The conversation delves into ‘tokens’ as the atomic units of AI computation. Kudrowski explains that while training costs were high, the cost of inference (producing tokens) is in ‘full outright collapse,’ falling 50-70% per year. This is catastrophic for business models tied to token sales, as companies must achieve impossible growth rates (e.g., 10,000% over five years) just to stand still in revenue terms, creating a massive transfer of wealth to consumers.
- 00:17:44 — Macro Impact: Data Centers Drove Half of U.S. GDP Growth — Kudrowski reveals a critical macroeconomic insight: in the first half of 2026, data center-related capital expenditure was responsible for over half of U.S. GDP growth. This anomalous spending explains the disconnect between tepid job growth and strong GDP figures. The build-out affects a wide swath of the economy, from chips and cement to architectural design, creating a ‘super cycle’ narrative that has spread far beyond tech companies.
- 00:26:05 — The Illusion of Demand and Overbuilt Capacity — Kudrowski dismantles the perceived demand for AI infrastructure. He states that all global consumer chat inference could be satisfied from a single data center, highlighting massive overcapacity. Most data center traffic is for training, but progress has stalled, and models are becoming commodities. He cites Microsoft CEO Satya Nadella’s comments about a ‘multimodal world’ as an admission that large investments like OpenAI are now commoditized.
- 00:34:27 — How the Bubble Will End: Multiple Failure Modes — Asked how the frenzy will end, Kudrowski cites Herb Stein’s adage: ‘things that can’t go on must stop.’ It could end via a debt market freeze, a flood of speculative IPOs soaking up risk capital, or simply collapsing returns on investment. He believes the sentiment shift began in October 2025 when markets stopped rewarding excessive capex. The fallout will hurt insurers, regional economies with subsidized data centers, and frontier AI companies, with OpenAI in a particularly weak position.
- 00:42:55 — The Optimistic Long-Term Vision: Free Cognition — Despite predicting a crash, Kudrowski is extremely bullish on the long-term impact. He compares AI to antibiotics or air conditioning—a foundational technology that makes cognition cheap and ubiquitous. This will unlock progress in medicine, materials, and science. It will also help offset demographic decline in developed nations, potentially leading to a more robot-intensive economy. The key societal challenge will be managing the transition for white-collar workers whose roles are automated.
Episode Info
- Podcast: Solutions with Henry Blodget
- Author: Vox Media Podcast Network
- Category: Technology Business
- Published: 2026-01-26T09:00:00Z
- Duration: 00:54:18
References
- URL PocketCasts: https://pocketcasts.com/podcast/9c941ed0-56cc-013e-8b75-0e680d801ff9/episode/fe2ac50c-7d00-4664-bacd-87e01e05ee38/
- Episode UUID: fe2ac50c-7d00-4664-bacd-87e01e05ee38
Podcast Info
- Name: Solutions with Henry Blodget
- Type: episodic
- UUID: 9c941ed0-56cc-013e-8b75-0e680d801ff9
Transcript
[00:00:01] Is the AI boom a bubble?
[00:00:03] Of course!
[00:00:05] That’s what our guest, Paul Kudrowski, thinks.
[00:00:07] Paul’s a venture capitalist writer and researcher.
[00:00:10] He’s one of the smartest independent analysts on the planet.
[00:00:14] Like me, Paul has lived through several of these boom and bust cycles,
[00:00:17] and he’s not afraid to make bold predictions about where we’re headed.
[00:00:22] For example, OpenAI, Paul thinks they’re toast.
[00:00:25] NVIDIA, the AI chipmaker whose valuation just crossed $5 trillion?
[00:00:31] Not toast, but also not a long-term winner.
[00:00:34] Paul also notes that the data center boom at the heart of the AI economy
[00:00:38] is drawing investment from a wide range of investors,
[00:00:42] including insurance companies and sovereign wealth funds.
[00:00:45] If and when the bus comes, all of these folks and the economy itself will get hurt.
[00:00:51] Happily, Paul is very optimistic about the underlying technology
[00:00:54] The same trend that will create problems for the current AI leaders,
[00:01:00] the rapid drop of prices, will unleash cheap cognition for everyone.
[00:01:05] This will allow us to progress faster than we ever have before
[00:01:08] in medicine, materials, robotics, and many other fields.
[00:01:12] So it’s not all a bust.
[00:01:14] Here’s Paul Kudrowski.
[00:01:16] Welcome, Paul. So great to see you.
[00:01:18] So you have been reading a fair bit of your stuff,
[00:01:20] and it sounds like you think that AI,
[00:01:24] a rerun of the dot-coms and railroads and the financial crisis and everything else,
[00:01:29] are at least headed that way. So tell us about that.
[00:01:32] Yeah. So it’s simultaneously a really unusual moment and a really predictable moment.
[00:01:37] So we’ve got this lunge towards capital expenditure
[00:01:41] to create this almost like a fabric underneath society,
[00:01:46] much like what happened with the canals back in the 18th century
[00:01:49] or like the railroads in the 19th or like rail electrification in the 1920s.
[00:01:54] Or like telecom and the fiber build-out around 2000.
[00:01:58] These were all fabrics that underlay a bunch of economic activity,
[00:02:03] and this is very similar to that.
[00:02:04] The only difference, of course, is that it’s happening much faster
[00:02:07] and the capital intensity is much higher.
[00:02:09] And those are the key differences, but the moment is very much the same.
[00:02:14] This build-out of an underlying fabric on top of which
[00:02:16] a host of different kinds of economic activity take place.
[00:02:20] And for the most part, the build-out of the fabric never works out very well
[00:02:24] for the people behind it.
[00:02:24] People building out the underlying fabric,
[00:02:26] and it works out wonderfully often for the people who build things on top.
[00:02:30] And that was true whether it was dot-com or the telecom bubble
[00:02:33] or rural electrification or whether it was, as I say, railroads or anyone else.
[00:02:37] So that’s the story that gets replayed all the time.
[00:02:40] So I’m reluctant to make it, you know, like we’re just replaying the dot-com bubble.
[00:02:43] That’s to miss everything that’s interesting about this moment,
[00:02:46] that it’s really about building out this incredibly capital-intensive substrate,
[00:02:51] fabric, whatever you want to call it,
[00:02:53] that’s sitting underneath something that’s really important to us.
[00:02:54] I don’t know much of what’s going on right now
[00:02:56] and the capital intensity of doing that.
[00:02:58] The other piece that I often do to try to characterize
[00:03:01] why this is such an unusual moment is that most U.S.,
[00:03:05] I’ll call them bubbles, but whatever you want to call them,
[00:03:07] enthusiasms, most enthusiasms usually have a couple of key characteristics.
[00:03:15] It’s about real estate.
[00:03:16] We love speculative bubbles in real estate in the U.S.
[00:03:18] It doesn’t matter whether it’s Florida real estate
[00:03:19] or whether it’s U.S.-wide in the financial crisis.
[00:03:21] We love technology.
[00:03:23] It’s a great story.
[00:03:24] It’s a great story that often helps drive this kind of thing.
[00:03:26] And that’s been true back to the Industrial Revolution.
[00:03:28] Loose credit really helps.
[00:03:30] That’s a bubble we love.
[00:03:31] It doesn’t have to do with any of the others.
[00:03:32] It can happen on its own.
[00:03:33] And government policy can really drive bubbles in the U.S.
[00:03:36] This is the first bubble in U.S. history that’s at the intersection of all four.
[00:03:40] So we’ve got all of those pieces conspiring and kind of reinforcing each other this moment.
[00:03:45] And that’s what makes this so unusual because you have all of these actors,
[00:03:50] people in tech, people in credit, people on the government,
[00:03:54] government side, who are talking about this as an existential crisis
[00:03:57] and a battle with China or whatever country you’d like to fill in the blank with.
[00:04:00] All of these pieces, real estate in particular, all coming together.
[00:04:05] And each of them think they’re acting rationally.
[00:04:07] But of course, at the intersection of all these rational actors,
[00:04:10] you get, you know, you finance folks call it a rational bubble.
[00:04:12] You get a rational bubble.
[00:04:13] You get all of these elements coming together because each person thinks
[00:04:17] they’re acting in their own interest and the combination of them all is kind of crap.
[00:04:21] It’s not a great moment, to use the technical terms.
[00:04:24] All right.
[00:04:24] So I want to talk about the rationality of it.
[00:04:26] I have participated in one of these bubbles, as you know well,
[00:04:29] and certainly a lot of the decision making then to me did seem rational.
[00:04:33] And then it ended in this huge collapse.
[00:04:37] But before we do that, you did say that in these build outs or
[00:04:40] enthusiasms, there are some winners and then some losers.
[00:04:45] So within the AI build out, who are who are the winners and losers?
[00:04:50] So the losers are almost certainly most of the major Frontier AI models.
[00:04:54] They’re almost certainly going to be losers.
[00:04:58] Most of them will probably disappear,
[00:05:00] being replaced by small language models, consolidate something else.
[00:05:05] I’m on the record as saying I don’t think there’ll be a chat GPT-8.
[00:05:08] I don’t think OpenAI will be around long enough to do that.
[00:05:12] But nevertheless, so those companies are going to be commoditized.
[00:05:15] It’s already happening. They have no mode.
[00:05:16] It’s highly capital intensive.
[00:05:18] The pace of developments is slowing and we can get into that.
[00:05:21] But these these are the unlikely ones.
[00:05:22] And then everyone who’s essentially
[00:05:24] subsidized and building on top of that stack, if you will.
[00:05:27] So this collapsing and I think this is an important element here,
[00:05:30] but the collapsing token prices that are the underlying sort of atomic units
[00:05:34] of this new economy, if you will, anyone who’s building on top of that
[00:05:38] and the recipient of that subsidy is kind of is an interesting place.
[00:05:42] So there’s it’s not clear to me yet who those people will be.
[00:05:45] It may turn out to be robotics companies.
[00:05:47] I don’t think it’s likely to be some of these thin wrapper companies we see,
[00:05:50] like people who are building lawyers in a box.
[00:05:53] I think that’s commoditized nonsense.
[00:05:55] And again, I don’t think it’ll be anyone.
[00:05:56] Almost no one in the current mag seven
[00:05:58] strikes me as a likely winner here, not Nvidia, not OpenAI, possibly Anthropic only
[00:06:03] because they’re oriented much more towards companies.
[00:06:05] But that’s the thing that I think it is very typical of this moment is that you’re
[00:06:10] at the build out and it’s not wasn’t obvious at the build out of, say,
[00:06:13] railroads who were going to be the big shipping companies.
[00:06:16] No, it wasn’t at all obvious.
[00:06:17] And instead it just led to a series
[00:06:19] of economic crashes and eventually out of that emerged some really interesting
[00:06:23] companies that were transporting goods over large distances.
[00:06:25] And the same thing was true with rural
[00:06:26] electrification and the same thing was true with the telecom build out.
[00:06:28] It wasn’t obvious that Netflix was going to be a big winner.
[00:06:31] No, it took time for that to happen.
[00:06:33] And so the mistake people make is they
[00:06:34] want to anoint someone who’s a sort of mag seven company to say, oh,
[00:06:38] they’re the winner and I’ll take the other side of that bet all day long.
[00:06:41] But I mean, you’re connected with all of the folks who are putting just astronomical
[00:06:47] amounts of money still into OpenAI and Anthropic and all the others.
[00:06:51] There’s this big race.
[00:06:52] And what I hear is,
[00:06:53] like, one of these guys is going to be the Google.
[00:06:54] We got to own it now and the rest are toast.
[00:06:57] You just don’t think so.
[00:06:58] No, I don’t. And I think that that’s it’s a classic mistake.
[00:07:01] You and I have been around these cycles so many times.
[00:07:03] People always want to fight the last
[00:07:05] battle when it comes to geopolitics and the same thing is true in tech.
[00:07:09] They want to anoint the next Google
[00:07:12] because that was the thing they missed in the last cycle.
[00:07:14] And I swore to myself I would never let that happen again.
[00:07:17] And it just doesn’t play out like that.
[00:07:19] These mirror echoes partly don’t play out in the same way because it’s
[00:07:23] not the same the same businesses, but an infrastructure build out like this
[00:07:28] is not analogous to what happened in the dot com period.
[00:07:30] Not at all.
[00:07:31] It’s as I say, it’s much more like rural
[00:07:33] electrification in the 1920s, where the main beneficiaries were sellers
[00:07:36] of like refrigeration and air conditioning decades later.
[00:07:39] Now, that’s not to say that whoever benefits this time around won’t show up
[00:07:42] for decades, it’ll all happen much faster, but it won’t look like General Electric
[00:07:46] or it won’t look like someone who was building out telephone poles in North
[00:07:50] Dakota. That’s the wrong way to think about what
[00:07:52] happens when you build out a new substrate on top of which people build
[00:07:56] companies based on a collapsing commodity price.
[00:08:00] And in this particular case, then it was power in this particular time.
[00:08:03] It’s this these these units we call tokens, not the crypto kind.
[00:08:07] All right.
[00:08:07] So for those of us who are not totally
[00:08:09] embedded in the jargon, don’t really know what tokens mean and that kind of thing.
[00:08:13] I think I know what you’re describing, which is basically revenue units.
[00:08:18] Ultimately, I pay chat GPT the same as when I first subscribed
[00:08:22] two years ago or whatever it was.
[00:08:24] What is it that’s declining in that?
[00:08:27] Because all I hear is that they’re losing more and more money, in fact,
[00:08:30] more and more money on every search or every question and so forth.
[00:08:33] So what if what’s what’s actually declining?
[00:08:36] So tokens are the representation of the underlying data that dates back to I mean,
[00:08:42] we’ve had tokens for probably 15 or 20 years now in machine learning,
[00:08:46] but they’re the atomic unit of how you capture the data.
[00:08:49] So think about them in a sense as words.
[00:08:51] Right. So loosely, they’re words.
[00:08:52] So loosely, they’re all of these units of information that’s in my training data
[00:08:57] that in turn, whenever you ask a question and if you want to think of these things
[00:09:01] as giant autocomplete engines, the tokens are the things that are being generated.
[00:09:05] So they consume them.
[00:09:06] They’re trained on them to generate frontier models, whether it’s open AI or
[00:09:10] whether it’s a cloud or whatever else, but they also produce them.
[00:09:13] So they consume and produce these things called tokens.
[00:09:16] So what’s happening is initially the cost of training models was very, very high
[00:09:21] and it stayed high.
[00:09:22] But for kind of weird reasons.
[00:09:23] But the cost of inference of producing tokens is what’s in full, full outright collapse.
[00:09:30] And that’s for a bunch of different reasons, partly competition,
[00:09:33] partly improved efficiency in data centers, partly because of technological changes,
[00:09:38] partly because this is this particular technology went from lab bench to billion
[00:09:43] users faster than any technology in history.
[00:09:45] So as a result, our ability to produce these these autocompletes, if you will,
[00:09:50] to your prompts, will continue
[00:09:52] to get cheaper and cheaper and cheaper.
[00:09:53] Isn’t that good for the companies, though?
[00:09:55] Well, it’s horrific.
[00:09:56] It’s good for the company if you can grow commensurately.
[00:09:59] So to literally just to stand still, they have to grow at more than 3000%
[00:10:05] because of this thing.
[00:10:06] Imagine being General Motors and the cost of cars is falling 60% year over year.
[00:10:10] I better find some new planets to sell
[00:10:11] cars on because I’m in big trouble selling on this planet.
[00:10:14] That’s the problem.
[00:10:15] So it’s me, the ChatGPT user who’s getting way more tokens for the same amount of money every year.
[00:10:20] That’s right.
[00:10:21] That’s right.
[00:10:21] OK.
[00:10:22] So it’s not that their costs to produce them are plummeting and therefore their
[00:10:25] profit margins are rising or lack of profit margin.
[00:10:28] No.
[00:10:29] So you’re you’re there’s this tremendous benefit to you as a consumer, which is great.
[00:10:32] It’s a transfer of wealth, right?
[00:10:34] It’s being transferred to you and you’re able to do more with them.
[00:10:36] But it’s a lousy business from the standpoint of being the producer of these
[00:10:40] things because they’re facing this exponential declining cost with respect
[00:10:44] to the units on top of which they generate their revenues and sell it.
[00:10:48] And so now now let’s go out a level and think about frontier models.
[00:10:52] They don’t run in isolation.
[00:10:53] They run on these things we call data centers, these great big buildings full
[00:10:56] of, you know, liquid cooling and humming NVIDIA chips and rack planes and so on.
[00:11:00] And most of those are funded from equity or funding from cash flow from large tech
[00:11:04] companies, but increasingly and it’s really took off last year.
[00:11:06] More and more of it’s being funded by debt.
[00:11:08] Well, if that’s the case, think about what happens with respect to my debt payments
[00:11:13] in the future that are tied to this center, the revenues of which are driven by token economics.
[00:11:19] I don’t get to shrink my interest payment
[00:11:21] in terms of what I’m paying back to Blue Owl or whoever my private whoever
[00:11:25] extended the loan to me that I’m using to pay for the data center
[00:11:28] on the basis of collapsing token prices that you don’t get to do that.
[00:11:32] So the economics of data centers are were bad in the first place in first place
[00:11:37] and getting worse because of this collapsing token economics problem.
[00:11:41] So all of this stuff going on under
[00:11:43] the surface makes the business of being in the data center side of things,
[00:11:47] the frontier model side of things difficult, if not impossible, because
[00:11:51] you’re trying to build at this kind of pace with high capex on top of a commodity whose prices are collapsing.
[00:11:57] And so where are we in this whole process?
[00:11:59] If you look at some of the eras that you’re talking about,
[00:12:02] and usually is anywhere between five and seven years of this huge enthusiasm where
[00:12:08] the early folks make tons of money and grow really fast and then everybody piles
[00:12:12] in, feel like it’s going to go on forever, then ultimately it crashes and then we rebuild from there.
[00:12:16] Like like where are we now?
[00:12:18] We in 1997 for the telecom build out?
[00:12:21] Are we? So this is going to happen?
[00:12:23] Yeah, that’s a great question.
[00:12:24] So that’s going to happen.
[00:12:24] Even this is going to happen even faster because we just don’t know how to do
[00:12:27] things slow anymore and in part because the nature of this thing is that it
[00:12:31] connects things up and accelerates the pace at which people produce new products
[00:12:34] and companies and all you have to do is look at things like cloud code or cloud
[00:12:38] co-work to see how people are this kind of acceleration, this phase we’re going through.
[00:12:41] So I think it’ll all play out very quickly.
[00:12:43] And I think my guess is we’re probably in early 99 in terms of my guess is that
[00:12:49] this year we’ll probably see and we’ve seen this already in China,
[00:12:51] to a degree, the first few truly speculative AI IPOs just came out over the last three
[00:12:56] months in China and both of them did, you know, moon rockets, moonshots.
[00:13:00] And so you will see this year the first non-frontier company IPOs.
[00:13:05] I think you’ll probably see at least a dozen.
[00:13:07] It could be two dozen.
[00:13:08] And the retail the retail investor appetite for those is immense,
[00:13:11] in part because of what’s happened recently with Mag7 stocks,
[00:13:15] but also because that’s the nature of the beast.
[00:13:17] People become really fatigued with quote to use the Wall Street parlance you and I know too
[00:13:21] well with old names.
[00:13:23] They want new names.
[00:13:24] Very exciting.
[00:13:25] Get me new names and everyone wants new names.
[00:13:27] And so the stock market, the equity sides,
[00:13:29] the sell side is nothing if not terrific at providing new names when asked.
[00:13:33] So you want new names?
[00:13:34] I will get you new names.
[00:13:36] And so we’ll see a host of those new names emerge this year,
[00:13:39] which will be very speculative, most of which will fail.
[00:13:41] But that will also begin to soak up some of the risk appetite in here,
[00:13:45] much like happened in 99 when we had dozens and dozens,
[00:13:48] more than 200, 300 IPOs at one point.
[00:13:50] Forget the numbers now.
[00:13:51] Huge numbers.
[00:13:52] And so that’ll begin to soak up some of the risk appetite at the margin,
[00:13:55] which will pull some of the capital out of the Mag7,
[00:13:58] which will make it more obvious that there’s going to be real winners and losers among them.
[00:14:01] And so that natural evolution is already beginning.
[00:14:04] There was a little bit of a regime change,
[00:14:05] a sea change in sentiment about the Mag7 stocks in October.
[00:14:09] And that’s really continued up to now.
[00:14:11] Like if you look at the XMAG, like the S&P 493 and compare it to the Mag7.
[00:14:16] I love this is the S&P 493.
[00:14:19] Compare the two of them.
[00:14:20] The XMAG
[00:14:21] is up about two percent a year today, a little more.
[00:14:24] The Mag7 is off more than five percent.
[00:14:26] And you can trace that back to not just this year,
[00:14:28] but to the October change with respect to how people began thinking about how should
[00:14:33] I, quote, reward companies that are spending this much, Oracle, others, Meta,
[00:14:38] in terms of the capital spending that they’re doing.
[00:14:40] Because we went through this phase, right?
[00:14:41] It was much like the crypto treasury companies where you were
[00:14:45] rewarded for spending it, for doing something silly.
[00:14:48] And the crypto treasury companies, my valuation went up by two.
[00:14:51] So I spent a lot of dollars if I put a dollar’s worth of crypto in treasury.
[00:14:56] The same perverse phenomenon was happening early on with capex.
[00:15:00] The more I promised to spend and dilute you as a shareholder and soak up cash flow,
[00:15:04] the more you rewarded me. So I know what to do.
[00:15:06] Whatever that’s the case, if I’m a board of directors, I spend more money.
[00:15:10] So they began spending much more money.
[00:15:11] But that moment changed in October when people stopped being rewarded for excessive capex.
[00:15:16] And so there’s a these things are all coming together in answer to your question.
[00:15:19] I think it’s cyclical. And again, going back to the dotcom bubble,
[00:15:23] there were pretty much every year from 1995 to 2000, you had lots of incredibly
[00:15:28] smart people like you coming out and saying, oh, it’s ridiculous.
[00:15:31] A bubble. It’s going to end in disaster.
[00:15:33] This is also silly.
[00:15:34] And often we would get a 30 percent correction in the big stocks and then
[00:15:38] they double and triple from there and then we get another one.
[00:15:41] So do you feel like we’re done?
[00:15:42] Was that the turn?
[00:15:44] Yeah, I think that was the beginning of the beginning of the turn.
[00:15:46] And this year will mark the change when we rotate out
[00:15:49] to the people who are building the infrastructure and there’ll be a second
[00:15:52] like an echo bubble, if you will, in the app related services that sit on top
[00:15:57] of the this fabric that we’ve built out.
[00:15:59] But I think the days of being rewarded for a grotesque amounts of capex and how
[00:16:03] many gigawatts you promised to build in like New Mexico or whatever the case may
[00:16:07] be, the days of being rewarded for that are largely over.
[00:16:10] And there’s some very good reasons for it, not least of which is that the other
[00:16:14] thing that’s unusual about this bubble is this moment is that we’re investing
[00:16:19] into a collapsing price commodity that wasn’t true in the dot com period wasn’t
[00:16:23] true in telecom. So think about that.
[00:16:25] The token prices on average have been
[00:16:27] declining 70 percent plus year over year for at least three, four years now.
[00:16:32] The math is pretty simple in terms of what that implies.
[00:16:34] If I want to produce a make Wall Street happy growth percentage, percentage growth
[00:16:39] over the next few years, let’s say I want to produce 30 percent growth on a
[00:16:41] compounding annual basis over the next five years to keep Wall Street happy.
[00:16:45] Leaving aside what happens with earnings, just revenues.
[00:16:48] If I’m doing that
[00:16:49] as a frontier company or I say a data center, a public data center provider,
[00:16:52] think about what I have to do if my underlying commodity, the thing that I’m
[00:16:57] selling that’s driving revenues in one form or another, is declining 50 to 70
[00:17:01] percent year over year, that’s unprecedented.
[00:17:04] So the implication is, let’s say I want to do 30 percent year over year in the
[00:17:07] in the underlying token prices are falling 60 percent year over year.
[00:17:11] The five year growth required is around 10,000 percent.
[00:17:14] Hello. That’s a faster growth profile than Facebook
[00:17:18] in its best five years than than cloud services ever than mobile phones ever.
[00:17:24] So while it’s possible, it’s so wildly unlikely that you at least should be a
[00:17:31] little more skeptical about the likelihood of producing that kind of unit growth
[00:17:34] that would justify the capex that requires that amount of growth.
[00:17:44] So you’ve written and talked a lot about just how deep this extended
[00:17:48] into the economy in terms of construction and the percentage of GDP growth and everything else.
[00:17:53] So give us a snapshot of that.
[00:17:56] And and what happens if that starts to reverse or if we start to max out here,
[00:18:01] which is, I assume, what ultimately ends the cycle?
[00:18:04] Yeah, well, one would think, but that’s being rational.
[00:18:07] I’ve learned not to be rational.
[00:18:10] So, yes, I’ll go with that.
[00:18:11] Let’s take that as a given.
[00:18:13] So, yeah. So in the first half of this year, data center related capex,
[00:18:18] at large, was about a little more than half of U.S. GDP growth, which is a really
[00:18:22] big deal because in the first half of the year and certainly in the second half as
[00:18:26] well, there’s been so much speculation in the analyst community, especially with
[00:18:29] respect to strategists about why the U.S. economy is behaving so strangely.
[00:18:32] Job growth is tap is tepid and yet GDP is holding up.
[00:18:37] And so what’s the disconnect?
[00:18:38] A huge chunk of the disconnect was this
[00:18:41] misunderstanding of the nature of what was driving GDP growth.
[00:18:44] More than half of it was coming from this
[00:18:45] anomalous spending on this single thing that we call data centers.
[00:18:48] And so my the point the reason why I bring
[00:18:50] this up is not just because it’s like good party talk.
[00:18:53] It’s because if you don’t understand the things that are driving GDP growth,
[00:18:57] you’re likely to come up with any random idea about what’s driving GDP growth.
[00:19:00] Maybe you’ll decide it was tariffs.
[00:19:02] Maybe you’ll decide it was my own personal charm.
[00:19:05] Maybe you’ll decide it was something else.
[00:19:06] When the reality was it was this one anomalous thing.
[00:19:09] This disconnect in our understanding of what’s driving the current macro cycle
[00:19:13] is a huge source of risk because you’re likely to double down on things that are
[00:19:17] actually.
[00:19:18] Releasing like, say, for example, tariffs or whatever else,
[00:19:21] not realizing that what’s driving this is what is actually this anomalous spending on GDP growth.
[00:19:26] So and again, now driving into that,
[00:19:28] you can say this is like 60 percent of the cost of data centers is chips.
[00:19:33] But there’s lots of other things going on, right?
[00:19:34] There’s architectural design firms.
[00:19:36] There’s construction.
[00:19:37] Roughly 30 percent, give or take, is cement now, which is ridiculous.
[00:19:43] I mean, so there’s people drive.
[00:19:44] There’s people making super cycle arguments that we’re entering the era
[00:19:47] of perpetually growing cement demand.
[00:19:49] I mean, it’s really becomes madness, right?
[00:19:52] So you get you people are finding their own favorite bull cases across all the
[00:19:55] underlying commodities that create these things we call data centers and their derivative products.
[00:19:59] And so so what ultimately stops it?
[00:20:02] So in other words, you’re painting a picture where if the demand drops and if
[00:20:06] the whatever that real spending driver here drops and it doesn’t, it stops
[00:20:09] filtering out into all other areas of the economy, it’s going to hit a lot of the economy.
[00:20:14] What stops it?
[00:20:15] What have you learned from other?
[00:20:17] Enthusiasm cycles like it’s the ROI.
[00:20:20] What happens is the market eventually says you’re making a huge investment that’s no
[00:20:24] longer just being funded out of cash flow.
[00:20:26] And I didn’t like that in the first place because that was dilutive to me and EPS was declining.
[00:20:31] And so fine, you wandered off and began doing all the spending using increasingly
[00:20:34] amounts of debt and that’s being provided by private credit providers and others.
[00:20:38] But that in turn also increases the risk here because now you’ve got some risk
[00:20:41] with respect to the stability of your cash flow, because that in turn is being driven
[00:20:44] by this collapsing price, your obligations are tied to this
[00:20:47] collapsing price commodity.
[00:20:48] So the bottom line is the ROI is now getting huge focus in terms of here’s the capex and
[00:20:55] here’s the likelihood of me being able to deliver a competitive return on that, which
[00:20:59] gets back to the point I made earlier that, say, 30 percent year over year compound five
[00:21:03] year growth would require this sort of like around 10,000 percent unit growth greater
[00:21:08] than the growth that we’ve seen in any under any comparable tech product in the history
[00:21:12] of such tech product doesn’t mean it can’t happen.
[00:21:15] It just means that suddenly people are
[00:21:17] actually being much more, I’ll say, careful about it.
[00:21:20] And I see this increasingly in data center spend.
[00:21:22] I see this increasingly as I talk to hedge funds and others who are who are active in
[00:21:26] this area and are actually looking at it, there’s a lot more attention being paid to.
[00:21:30] How does this generate a competitive
[00:21:33] ROI that justifies the valuations that we’re seeing?
[00:21:36] And the answer is, I don’t see it.
[00:21:38] And for a very simple example, I’ve been watching for years the open
[00:21:42] AI financing numbers go up in terms of the amount they’re raising and the valuations.
[00:21:46] Yeah.
[00:21:47] Amazon IPO was very controversial.
[00:21:49] Everybody’s, oh, it’s a money losing internet bookstore.
[00:21:51] You know, who cares? It’s worth zero.
[00:21:53] But when it went public, it was only 400 million.
[00:21:56] It’s like this tiny little number.
[00:21:58] And so if they got a lot of things right, you had a huge upside from that.
[00:22:01] I look at open AI now.
[00:22:03] A last headline I saw was they’re raising money at eight hundred billion dollar
[00:22:08] valuation as dropping similarly. Yeah.
[00:22:10] Yeah. As a professional investor, which you are, what kind of return do you need to see to justify?
[00:22:16] Justify that. What are you expecting?
[00:22:18] So there’s two different answers to that.
[00:22:20] One is that I don’t care as long as retail investors give me forex on the IPO.
[00:22:25] So I mean, there’s a very cynical answer to this, which is if I can unload it at
[00:22:30] the IPO at a significant premium to what I paid for it in the last post fund,
[00:22:34] the last pre IPO financing round, then I don’t really care what their
[00:22:39] underlying unit economics look like. That’s a problem for retail investors in future.
[00:22:43] And I think it’s really important to keep that in mind because that’s one of the
[00:22:46] reasons why people got this all wrong during the dotcom episode was that, oh,
[00:22:50] you know, Amazon will never generate a competitive return.
[00:22:53] That mattered later and that mattered later for holders post IPO.
[00:22:57] But early on, what mattered more was how does this perform and can I flip my shares?
[00:23:02] So we’re going to go through that phase first.
[00:23:04] And so the fact that calculating what a reasonable rate of return that they’ll
[00:23:07] have to generate in terms of a cash flow that would justify the valuation I pay
[00:23:11] in a series double Z round on cloud on open AI is kind of irrelevant.
[00:23:16] But in the sense that most of the people buying here are not are buying to try
[00:23:20] and get rid of it, to unload it post IPO.
[00:23:22] So you said a few minutes ago, which I think is the key to all of this,
[00:23:25] that most of the folks making the decisions think that they are rational decisions.
[00:23:30] Is that true? Oh, absolutely.
[00:23:32] That’s the thing that’s really funny to me is that I get asked all the time by
[00:23:36] some of the hedge funds I consult. It’s like they must know inside.
[00:23:38] This is ridiculous. It makes no sense.
[00:23:40] Who can we find someone like a whistleblower who says it’s fraud?
[00:23:44] I’m like, dude, you don’t understand how
[00:23:46] people drink the Kool-Aid inside these companies.
[00:23:48] Not only do they not think it’s fraud, they actually think they’re changing the world.
[00:23:52] So you’ll never find that whistleblower who says, you know, last week in a meeting
[00:23:57] at OpenAI, I heard Sam say, boy, oh, boy, are we ever pulling one over on those
[00:24:01] of Rubes at Blue Owl? That’s not happening.
[00:24:04] No one’s doing that. These are true believers.
[00:24:06] That’s the nature of tech.
[00:24:07] Sure, you’ll find people who are screwing around with really speculative data
[00:24:11] centers. There’s a few out there that are like politically driven that make
[00:24:14] absolutely no sense or probably very cynical.
[00:24:16] But inside of the large frontier companies,
[00:24:18] these are uniformly true believers in terms of what they’re doing.
[00:24:22] So you will not find somebody who’s going to tell you that, oh, you know,
[00:24:25] this is all just dodgy nonsense that we’re doing to screw people over.
[00:24:29] And so if the decisions are rational, because I think you can go to lenders,
[00:24:33] too, you can certainly go to the investors.
[00:24:35] You explained the logic there, which is, oh, my goodness, I missed Google last time.
[00:24:39] I got to get to Google this time.
[00:24:40] It’s not going to be me who misses it.
[00:24:43] They all make sense.
[00:24:45] And I think you can look
[00:24:46] at OpenAI and say, look, you know, guys, if we have any chance of doing this,
[00:24:52] we’ve got to be the one that crashes through here and makes it so spend
[00:24:56] absolutely everything we can, we’ve got the ability to raise money.
[00:24:59] So it looks very rational.
[00:25:01] So given that,
[00:25:03] can it happen any other way?
[00:25:04] I mean, you’ve been you’ve seen these cycles.
[00:25:06] You’re an amazing writer.
[00:25:08] So you make fun of everybody as you go.
[00:25:10] You’re so sardonic about it.
[00:25:11] I mean, it’s like I look at all this stupidity.
[00:25:13] But is there another way that we could be
[00:25:16] doing this, these innovation cycles?
[00:25:19] Oh, I don’t think so.
[00:25:20] That’s that’s one of the and that’s Carlotta
[00:25:22] Perez and lots of other people’s arguments that these kinds of cycles
[00:25:26] require grotesque overspending at this point in the process.
[00:25:31] And all you can do is try to either stay
[00:25:33] out of the way or caution people that, you know, this here’s how this part of it is
[00:25:36] likely to end, but you don’t know if you’ve built enough until you’ve built too much.
[00:25:41] And I think that’s really key to understanding.
[00:25:43] And people who come into this thinking that there’s some kind of rational,
[00:25:46] clean way, clean way of doing this sort of thing are naive and not because of,
[00:25:53] you know, the animal spirits or whatever else, but it’s because you just don’t know.
[00:25:57] You don’t know when you’ve built enough till you’ve built too much.
[00:26:00] You don’t know when you’ve got enough supply of tokens until you’ve got too much.
[00:26:03] I mean, I’ll give you an example.
[00:26:05] People often say, you know, oh, I’m hugely bullish about AI because chat GPT is super
[00:26:10] useful to me and I’m like, that’s terrific news.
[00:26:12] I’m glad it’s really working out for you.
[00:26:13] But I said, how important do you
[00:26:16] think consumer chat is to the to the AI story?
[00:26:19] And they’re like, well, it’s huge.
[00:26:21] Everyone I know is now using, you know, Claude or chat GPT or whatever else.
[00:26:25] And I said, so that’s inference traffic.
[00:26:27] So that piece of inference traffic you think is really important to the story
[00:26:30] and it’s driving your bullishness and they’re like, yes, I said, OK, fine.
[00:26:33] I said, how much of current global
[00:26:36] inference capacity is required to satisfy consumer chat?
[00:26:40] And they are usually off by a factor of ten.
[00:26:42] They say, oh, you know, 60, 70 percent of all of the data centers out there, it’s five.
[00:26:46] It’s less. Well, it’s actually a little less than five percent.
[00:26:48] You could satisfy all of global inference,
[00:26:51] consumer chat related global inference from a single data center in northern Virginia.
[00:26:54] All of global inference.
[00:26:56] We are not the story here.
[00:26:57] You and I, as much as we want to be part of this as good narcissists,
[00:27:01] I want to be part of the story. I’m not part of the story.
[00:27:03] The story is, is this build out for training
[00:27:07] and for something else that they’re hoping will come along.
[00:27:10] And so we’re already at a point where the level of overbuild is titanic in service of things
[00:27:16] that have yet to be identified. So we’re already at that moment.
[00:27:19] If you think about it in terms of
[00:27:21] the relevance of the thing that makes most people bullish,
[00:27:24] which is the utility of consumer of consumer chat with respect to these apps.
[00:27:28] OK, so so I want to actually restate that because I think that people it’s so easy
[00:27:32] to get lost in the jargon and the training and inference and everything else.
[00:27:35] So what you’re saying is the actual product that I’m consuming when I use
[00:27:40] ChatGPT or Claude, when I ask a question or I ask it to do something,
[00:27:44] that’s the inference piece and that.
[00:27:46] That actually uses a single data center in Virginia and that’s it.
[00:27:49] It could is my point.
[00:27:50] It’s it uses it not.
[00:27:52] My point is that we’ve got like five thousand plus data centers out there.
[00:27:56] We could satisfy it all from the racks of a single day in Northern Virginia.
[00:27:59] It doesn’t happen that way.
[00:28:01] But the point is to drive home how much
[00:28:03] excessive supply we have already if the world is moving towards inference.
[00:28:07] So then the question is,
[00:28:08] is what the hell is going on inside of all those other data centers?
[00:28:11] Is it like laser tag or, you know, hide and seek or whatever else?
[00:28:15] And
[00:28:15] it’s about 60 percent of the traffic in those data centers in terms of the workload
[00:28:19] is training and you’re like, oh, that’s terrific news.
[00:28:21] So you’re you’re spending all that money
[00:28:23] to make better things better and shinier things for me.
[00:28:26] That makes me very happy.
[00:28:28] And of course, then you get into the thing.
[00:28:29] Well, no, that’s actually not the problem, not what’s happening.
[00:28:32] It’s training loosely, but most of training is just experiments,
[00:28:35] which is to say the large frontier companies are trying to want to create
[00:28:39] better models, but they’re at a bit of a log jam where we’ve kind of hit
[00:28:43] an architectural cul-de-sac, a dead end,
[00:28:45] with respect to the progress of these models.
[00:28:47] And you can see that in how people react to new models like, oh, this GPT 5.2.
[00:28:52] I can’t tell the difference. It’s not that much better than five.
[00:28:55] We’ve all become like Yelp reviewers of these things, right?
[00:28:57] We were like, oh, this is only four and a half stars.
[00:28:59] This isn’t as good as I expected.
[00:29:01] And that’s but if you look under the hood, what’s really going on is we’re
[00:29:05] reaching a natural point with respect to the the the richness of the underlying
[00:29:10] training data that’s used to produce these models, that it’s taking longer, producing
[00:29:15] more marginal improvements and costing more to produce the things that are being created.
[00:29:20] And so that’s the moment that we’re at now.
[00:29:22] And those models are running at these data centers, but that are scaled up
[00:29:26] in anticipation as if we will be able to do this forever.
[00:29:30] That’s a huge error. That’s not going to happen.
[00:29:32] And it’s already it’s already we’re already
[00:29:33] seeing that with respect to the to the models themselves.
[00:29:36] And this is something I was very struck a few months ago when Gemini,
[00:29:41] Google’s thing came out and suddenly got better ratings than ChatGPT.
[00:29:45] And all I can say is that never happened in the 1990s.
[00:29:49] Barnes and Noble never caught Amazon.
[00:29:52] You could say what you want about Amazon, but they were way ahead and accelerating.
[00:29:56] It’s just simply not happened here.
[00:29:58] And you’ve got really smart people like you and others in the tech industry are
[00:30:02] saying, look, the LLMs are commodities now and you can use you can use them
[00:30:06] interchangeably, which is very striking, really striking.
[00:30:10] And you even heard yesterday was that yesterday Davos Sachin Adabo, Microsoft
[00:30:15] CEO, was making the point that we live in a multimodal world and we need to build
[00:30:20] fabrics that can allow our users to mix and match and so on.
[00:30:24] And that’s all really great.
[00:30:25] But it’s like, dude, you’re like a huge investor at OpenAI.
[00:30:28] You just said one of your largest investments is kind of meh to you.
[00:30:33] What the hell? This is this is a really unusual moment
[00:30:37] because you’re essentially saying that something in which you’ve got a huge
[00:30:40] stake is commoditized, so you’re you’re placing bets across the landscape because
[00:30:45] you don’t know whether or not any of them will actually matter.
[00:30:47] And that went by like people said, huh, nice to see Sachin Adabo.
[00:30:51] So no one was saying, wait a minute.
[00:30:53] This is a remarkable statement that this
[00:30:55] is so thoroughly commoditized that one of the largest investors is diversifying
[00:30:58] across. Yes, that’s your point.
[00:31:00] Yeah. And you’ve got to say for Microsoft, what an incredible investment.
[00:31:03] They figured out a way to take tens of
[00:31:05] billions of dollars of cash sitting on their balance sheet, doing nothing and
[00:31:09] turn it into revenue because they get a huge royalty of all the revenue that
[00:31:13] ChatGPG generates.
[00:31:14] So it’s tremendous, although that expires in the early 2030s.
[00:31:18] But nevertheless, yes, yes.
[00:31:19] Well, we’re worried about the short term here.
[00:31:21] But Satya Nadella, he sort of he was asked on the last call,
[00:31:26] you know, hey, dude, like everybody’s talking about a bubble.
[00:31:29] Like, what are you doing spending all this money?
[00:31:31] And his response was, we cannot keep up with demand.
[00:31:36] We it’s not just about chips.
[00:31:38] We’ve got racks of chips.
[00:31:39] We can’t even get power to them.
[00:31:41] There’s an incredible demand from what you’ve just said, which I think is incredibly
[00:31:44] important for everybody to understand is most of that demand is for training new
[00:31:51] models that we may not need or where they are not just needs that are actually
[00:31:55] incremental in commodities like I defy you in a Coke Pepsi sense to tell the difference
[00:31:59] between most of the models other than rock, which is regularly fairly offensive.
[00:32:03] So of the non offensive models, I defy you
[00:32:06] to tell the difference between them in a blind taste test.
[00:32:08] That’s a really big deal.
[00:32:10] So, yeah. OK, fine.
[00:32:11] They have no choice. So they continue to do it because this was
[00:32:14] Oscar Wilde, the writer, had this tremendous line years ago where he said,
[00:32:17] a man’s second marriage is the triumph of hope over experience.
[00:32:20] This is a little bit of that, right?
[00:32:23] This is the triumph of hope over experience.
[00:32:24] I know my last training round didn’t
[00:32:26] produce a model with all kinds of wild new functionality, but maybe this time it will.
[00:32:31] Well, that’s back to the Oscar Wilde line, right?
[00:32:32] So it doesn’t really work that way.
[00:32:34] And the second point I’d make is that one
[00:32:36] of the illusions underlying the current wild expansion that Satya is citing in terms
[00:32:40] of demand is that, OK, fine, five percent of traffic is
[00:32:44] is consumer chat and 60 percent of overall traffic is training.
[00:32:48] That’s still a lot of unaccounted for traffic.
[00:32:50] What’s the biggest piece of that? It’s coding.
[00:32:52] So about 85 percent of inference traffic is coding.
[00:32:56] So and that’s exploding because developers increasingly are like, ah, screw it.
[00:33:00] I’ll just throw all my code in here and say, you know, do this and I’ll check
[00:33:03] that periodically on the tab and see how it’s doing with respect to producing code.
[00:33:07] Coding is two things.
[00:33:08] One, models are very, very good at doing
[00:33:11] code, but they’re also incredibly wasteful at doing code because you have to ingest
[00:33:14] an entire code base every time to make incremental improvements.
[00:33:17] So it’s like having to ingest an entire
[00:33:19] book every time you want to read the next paragraph.
[00:33:21] So this is really important.
[00:33:22] But that’s as a signal with respect to demand in the marketplace.
[00:33:28] It’s not exogenous.
[00:33:29] It’s not external to the market.
[00:33:31] It’s internal to the market because the
[00:33:32] largest users of coding tools are inside the frontier AI companies themselves.
[00:33:36] So the notion that they’re radically expanding their use of AI,
[00:33:40] IDEs, development environments, and that in turn is a very wasteful use of tokens.
[00:33:44] And that’s creating the impression that token demand is growing.
[00:33:48] That’s not an external signal that doesn’t validate anything you’re doing.
[00:33:51] And then more fundamentally, and you know this, because we’ve both been around this
[00:33:55] stuff for so long, I defy you to name a large cap tech company in the software
[00:34:01] development tools business. And there’s a reason for that.
[00:34:03] Selling to developers is a really crappy business.
[00:34:06] So the notion that I should therefore extrapolate a giant multi-trillion dollar
[00:34:11] market on the back of it, of huge inference traffic flowing
[00:34:14] from developers is wrongheaded on at least two, on at least two bases.
[00:34:24] All right. So, Paul, how is this going to end?
[00:34:27] And I will, I will sort of provide context of in the end of 1999,
[00:34:32] I was very aware that it was probably a bubble.
[00:34:36] Yeah, sure. And what I couldn’t do in my head was
[00:34:39] figure out how we were going to get from this moment of total euphoria where a business plan
[00:34:43] could raise whatever they needed to do, they would go out and spend all the money
[00:34:47] to this wreckage that everybody was talking about.
[00:34:50] I couldn’t figure it out. Then it became obvious.
[00:34:53] The financial interest rates went up, the financial markets closed.
[00:34:56] Suddenly there was no money that everybody could spend anymore.
[00:34:59] And then all of these different things
[00:35:01] that were supporting the fundamentals just collapsed and we had the wreckage.
[00:35:04] So what happens here?
[00:35:06] So I always fall back and there’s an old line from a member of the Council
[00:35:10] of Economic Advisers during the Nixon era, I think it was, Herb Stein, who said
[00:35:13] that things that can’t go on must stop, which is like a ridiculously trite
[00:35:18] statement, but actually applies in all of these contexts.
[00:35:20] And in particular, in this one,
[00:35:22] it applies because there’s so many ways in which it can fail.
[00:35:25] It can fail because we have a moment where debt markets are no longer as fluid in
[00:35:29] terms of providing financing for the more speculative data center providers.
[00:35:34] It can fail on that basis.
[00:35:35] It can fail because we have a hundred IPOs this year or over the next two years
[00:35:40] that dilute the risk capital from retail investors and suck all
[00:35:43] the air out of the frontier companies.
[00:35:46] So the way I look at this is, A, it must end, because that’s what happens
[00:35:50] with all of these capex frenzies, they always end the same way,
[00:35:53] which is to say by oversupply, something like 50 percent of the rail lines ever
[00:35:58] built were abandoned, similar was true with canals,
[00:36:00] similar was true with rural electrification.
[00:36:03] Most of the dot com fiber sat fallow for a better part of a decade.
[00:36:06] That will be true here as well.
[00:36:09] So knowing that, trying to time it and say, oh, you know, it’s going to end in six
[00:36:13] months because of this and this and this, all I can say is we’re closer to the end
[00:36:17] than the beginning. Could it go on for two more years?
[00:36:19] Sure, it could go on for two more years.
[00:36:21] And is it possible that in the middle of all of this, we have a kind of, I had
[00:36:24] someone argue to me the other day that we’ll have like the Uber of tokens emerge
[00:36:28] that will be using, you know, trillions of tokens and all of the arguments about
[00:36:31] declining, you know, collapsing token prices and that these frontier companies
[00:36:34] will never get enough appetite don’t matter because there’ll be autonomous
[00:36:38] robots exploring the world and consuming it on the basis of tokens.
[00:36:41] And they’ll be streaming it all back.
[00:36:43] And that’ll require.
[00:36:43] You know, quintillions of tokens flowing back and forth.
[00:36:46] And I’m like, you’re like that old joke about there’s like an economist and a
[00:36:49] physicist and somebody else who gets stuck on a desert island.
[00:36:52] And the physicist says, I’m going to drop and they want to open up a can of food
[00:36:55] and physicists, I’ll drop it out of a tree.
[00:36:56] And I forget what the chemist says, but the economist says, I’ll just assume a can
[00:37:00] opener. I’m like, dude, you just assumed a can opener.
[00:37:02] You just assumed away the problem.
[00:37:04] This is a really bad way to approach these kinds of moments.
[00:37:07] So the way I look at it is, roughly speaking, we’re in this moment.
[00:37:11] It always turns out the same way in these
[00:37:12] CapEx frenzies and it’ll end with declining ROI or at least an appetite for
[00:37:17] finding out what the actual ROI is on this stuff that’s already begun to happen.
[00:37:21] As I was saying with, you know, the XMAG and MAG7 stuff and the sentiment change
[00:37:24] in the fall and even look at the credit default swap
[00:37:27] numbers for some of the proxies, like, for example, CoreWeave or look at Oracle.
[00:37:31] And you’ll seeing all of this, the sentiment’s already begun to change.
[00:37:35] And so there’ll be a quick flip of it to retail that’ll soak up a lot of the risk
[00:37:38] capital and that gets you to maybe, you know, 18 months from now at most.
[00:37:42] And who’s going to get hurt?
[00:37:46] So that’s an interesting question.
[00:37:47] So the weird thing in this particular moment is that it’s really metastasized,
[00:37:53] to use the, you know, the oncology term, it’s spread across the entire marketplace.
[00:37:57] This little this cancerous cell, if you will.
[00:38:00] So, you know, there are insurers who are
[00:38:02] providers of capital with respect to data centers that are going to be heard.
[00:38:05] And then on the other side, you have the obvious candidates.
[00:38:07] Right.
[00:38:07] I think I think OpenAI is probably in the weakest position with respect to being
[00:38:12] one of the people
[00:38:12] that provides the great shakeout with respect to frontier models.
[00:38:15] I think obviously Google and Microsoft are in much better positions.
[00:38:17] Anthropic is to a greater to a larger
[00:38:19] degree because I think they’re more focused on the enterprise market,
[00:38:22] which is probably a better place to be than trying to farm engagement and
[00:38:25] convince people that your model loves you and wants to take you home
[00:38:29] if you’re in the more retail side of the chat market.
[00:38:31] So these are so these are some of the
[00:38:32] people I think are sort of obvious winners and losers.
[00:38:35] I think in the longer the longer run, Nvidia is no winner here at all.
[00:38:38] I mean, they’re
[00:38:39] they’re kind of a mini computer provider as personal computers emerged.
[00:38:42] If you
[00:38:42] want to think about it in historical computing terms, they’re they’re very
[00:38:45] monolithic, expensive, power hungry chips.
[00:38:48] And people misunderstand the role of Nvidia chips here.
[00:38:53] I use like a highway tollbooth analogy here.
[00:38:57] So Nvidia is the tollbooth provider that’s trying to chew through this line
[00:39:02] of people who are coming across a bridge and all the lanes are funneling into the
[00:39:04] tollbooth.
[00:39:05] But if I have 100 lanes coming down to two or three,
[00:39:08] it doesn’t matter how fast I make the tollbooth provider.
[00:39:11] So Nvidia, you’re seeing this already.
[00:39:12] We’re Nvidia will announce that it’s got, you know,
[00:39:14] I have petaflops of performance and people are just like, yeah, I don’t really care.
[00:39:19] And in the end, the reason is because that’s no longer the gating factor.
[00:39:23] The world has changed from 2023.
[00:39:25] The gating factor are the number of lanes, if you will, feeding the tollbooth.
[00:39:29] And that’s memory.
[00:39:30] And that’s why companies like SK Hynix and SanDisk and others are doing so well.
[00:39:34] It’s because it’s already changing the nature as we move towards an inference
[00:39:39] world about where the blockages are.
[00:39:41] So I don’t think in the long run.
[00:39:42] Nvidia, which kind of stumbled into this in weird sorts of ways because it was this was
[00:39:46] just happened to be really well attuned to what they were doing.
[00:39:48] So I don’t think they’re a long term winner here either.
[00:39:50] And for some of those reasons that increasingly it’s about memory storage and high bandwidth,
[00:39:55] high bandwidth memory in particular, not about whether or not my tollbooth runs faster
[00:39:58] than yours.
[00:39:59] And so from what you’re saying, it goes way beyond just the enormous, incredibly cash
[00:40:06] risk, rich tech companies like Microsoft and Facebook and Google that are spending hundreds
[00:40:11] of billions of dollars.
[00:40:12] Collectively, Amazon, too, is I think for a while people did feel like, hey, you know, the money’s not coming from individual investors.
[00:40:20] It’s coming from these companies that make $40 billion a year.
[00:40:23] They don’t care.
[00:40:23] They can lose it.
[00:40:24] And we do know they can lose it.
[00:40:26] You look at Mark Zuckerberg burned tens of billions of dollars a year on the metaverse, which was a hallucination.
[00:40:32] Nobody cared.
[00:40:34] Ultimately, he said, ah, shut it down.
[00:40:36] We’ll write off 80 billion or whatever it is.
[00:40:38] So I assume that that’s what Microsoft will do.
[00:40:40] Just write it off.
[00:40:41] And nobody cares.
[00:40:41] The stock will go up.
[00:40:42] Because they finally got the bad news out.
[00:40:45] But what you’re saying is it goes way, way beyond that.
[00:40:47] Oh, way beyond that.
[00:40:48] Into the economy.
[00:40:49] OK.
[00:40:49] Right.
[00:40:50] And even in more actually somewhat sadder ways, which is I often talk to regional economic development people who are offered the opportunity, in air quotes, to have data centers built in their region.
[00:41:02] And so these, you think about it in terms, I’m a small town.
[00:41:04] I’ve got a, I’m a regional economic development official.
[00:41:07] I bid on the Hyundai battery factory.
[00:41:09] I bid on the assembly plant for Honda.
[00:41:11] Didn’t get any of those.
[00:41:12] I’ve got, you know, 40%, 30% unemployment.
[00:41:14] And along comes fill-in-the-blank company who wants to build a data center.
[00:41:18] And they make this great story about how these are the factories of the new industrial revolution.
[00:41:22] And all I need from you is a tax abatement and subsidized water, subsidized power, and I’ll be there for 20 years with high-quality tenants, and it’ll create permanent jobs.
[00:41:32] But when you do the math, almost inevitably on these things, the cost on a per-job basis is usually in the millions, given the subsidies inherent in constructing these things.
[00:41:41] And yet they’re being built all over the country in this kind of speculative fervor.
[00:41:45] And so there’ll be a lot of regional economies hurt by this that end up with these, you know, white elephants, data centers that are either underused or unused that were built at the peak of the frenzy of building out the cycle.
[00:41:56] And we’re done for good reasons.
[00:41:57] Regional economic development officials are having an impossible job.
[00:42:00] And when given this and told that this will be, you know, in use for decades, how do you say no to that?
[00:42:06] And so these are, so the damage goes much more broad than that.
[00:42:08] And that’s just why I think it’s important to get away from San Francisco.
[00:42:11] I’m saying, oh, you know, Microsoft’s good for it.
[00:42:13] This is a mistake.
[00:42:15] All right.
[00:42:15] Let’s assume that this vision of the future is right, that there’s going to be this big crash wreckage.
[00:42:21] The press is going to pounce.
[00:42:22] All these books are going to be written about how stupid everybody was and how could they believe these things and so forth.
[00:42:28] And yet from that, there will be green shoots.
[00:42:31] And then hopefully a massive AI-powered economy will be built just the way the internet was, just the way railroads were.
[00:42:39] Yeah.
[00:42:39] And you have built.
[00:42:40] Yeah.
[00:42:40] Been very optimistic about that.
[00:42:42] And we’ve been painting you into this, you know, oh, I’m going to be the curmudgeon and I’m going to tell you that all this stuff that looks so great, it actually sucks and people are being stupid and all that stuff.
[00:42:51] So tell us the positive vision of after this collapse.
[00:42:55] Oh, yeah.
[00:42:55] This is literally the single most important technological change in my lifetime and probably one of the most important, probably dating back to maybe like air conditioning or something else in terms of really changing the way people work and live.
[00:43:07] If you think about it in those terms, it’s like antibiotics.
[00:43:10] It’s like antibiotics in healthcare.
[00:43:11] It’s like air conditioning in urban life.
[00:43:13] It’s like that.
[00:43:14] It’s this fabric that really radically changes the nature and cost of everything we do.
[00:43:19] So in that sense, turning from, you know, what’s going to happen to these random, you know, fools who are building giant data centers in locations that it can’t justify, that’s kind of irrelevant in the longer run.
[00:43:30] In the longer run, what’s going to be much more important is that cognition loosely is going to become free.
[00:43:37] That’s what we’re really.
[00:43:38] So the atomic units of cognition.
[00:43:40] And I sense are these tokens.
[00:43:42] Token prices are collapsing.
[00:43:44] They will be ubiquitous everywhere and essentially costless.
[00:43:48] So they’ll underlie almost everything you do that will have immense job transformation consequences.
[00:43:52] And it could cause a huge disruption in the white collar workforce.
[00:43:55] But on the other hand, the opportunities inside of science, inside of research, inside of medicine, inside of all these areas that were previously very difficult to plumb because the complexity and the interactions inside of it made it hard.
[00:44:09] You know, all you have to do is look at the cost of drug discovery.
[00:44:10] The cost of launching almost anything in medicine.
[00:44:13] But across a host of these fields, suddenly all of these open up because the data and cognition essentially became costless and free.
[00:44:19] And one of the ways my partner and I like to look at the world in terms of our own investing is what happens if something that was previously scarce and expensive becomes free.
[00:44:27] And that was true about bandwidth.
[00:44:29] That was true about CPUs and processing.
[00:44:31] It’s been true about physical distance back in the days of railroads.
[00:44:36] It was true about power back in the days of electrification.
[00:44:38] And it’s true now in the sense of cognition.
[00:44:40] What happens?
[00:44:40] What happens if cognition becomes free?
[00:44:43] And again, something that was expensive and scarce becomes cheap and ubiquitous.
[00:44:47] It’s lousy for people in grad school.
[00:44:49] I will concede that right away because these are, you know, again, predicated on cognition remaining scarce and expensive.
[00:44:55] But it’s tremendous society-wide in terms of unlocking.
[00:44:59] It’s like a cheat code for unlocking cognition for a huge population of people who might otherwise never have access to all of the tools that were otherwise, you know, locked up in seven years of grad school or locked up inside of years of working in a particular field.
[00:45:10] And so, you know, I think it’s really important for us to be able to do that.
[00:45:12] I talked to my plumber the other day, and he often has his phone sitting on his shoulder, filming things as he does them and getting feedback on a continual basis from a vision model.
[00:45:21] He uses OpenChat GPT for this.
[00:45:23] It’s not built for that, but it’s a general-purpose technology that’s providing costless cognition, and that’s incredibly valuable.
[00:45:29] And so what about all the folks in grad school and the job transition?
[00:45:32] I mean, there have been some very dire predictions, often out of Silicon Valley, about, oh, nobody’s going to have a job anymore and so forth, which would lead to society.
[00:45:40] It will collapse and so forth.
[00:45:42] What do people do?
[00:45:43] So I think people will do what they’ve always done, which is they recognize that the thing that I’ve been training myself to do no longer is a thing.
[00:45:49] And that’s unfortunate, and it’s often sad, and it’s often multigenerationally consequential for those people.
[00:45:55] And I take nothing away from that.
[00:45:57] After the Industrial Revolution, the U.K. got richer on GDP per capita basis, but individuals in general were poorer for almost 30 years.
[00:46:06] So you wave that away at your peril, because think about that in the U.S. terms.
[00:46:10] The U.S. gets richer.
[00:46:11] And most people get poorer.
[00:46:11] Are people going to say, you know, up with AI, and I love everything that’s going on?
[00:46:15] Or are they going to say, you know, let’s burn it all down?
[00:46:18] So we have to be very careful to say, even though there are gains coming, there have to be mechanisms for making it easier for people to gain access to this.
[00:46:25] And I’m not a believer that you do that by, like, training people to become, you know, AI impresarios.
[00:46:31] I think you do that in part by just dodging this altogether.
[00:46:34] You say, why aren’t you in trades and vocational schools?
[00:46:37] Why aren’t you in things that are less in the—
[00:46:40] Bullseye of what’s going on right now, and more importantly, have huge numbers of openings.
[00:46:45] I had a water commissioner in a relatively small town the other time.
[00:46:48] He had more than 100—no.
[00:46:50] He had 25 openings, none of which paid more than 100—paid less than $100,000 for people who were going—for technicians, vocational jobs within his town.
[00:46:59] And he’s like, I can’t get anyone to apply for any of these.
[00:47:01] And this is the problem, is that we’ve got this huge misallocation of people towards occupations that are shrinking and even going to go away,
[00:47:08] and away from things that have—
[00:47:10] large numbers of openings, stable incomes, and a huge appetite for people to do them.
[00:47:16] And for whatever reason, societally, we’ve turned up our noses at those things.
[00:47:19] And I think that’s where everything has to change, is recognizing that, you know, that’s where the opportunities are,
[00:47:25] not in the old world of saying, if I just train long enough, I can eventually become a managing director at Goldman Sachs.
[00:47:30] I’m going to ask you to end with a very optimistic picture of what humanity and progress and everything else looks like in the 2030s.
[00:47:37] So if I understand you correctly, what you think’s happening is that—
[00:47:40] basically, things that it took human beings five to ten years in grad school to learn are now going to become available to everybody at their fingertips.
[00:47:50] And that will create lots of opportunities in medicine and so forth.
[00:47:53] But paint that picture for us.
[00:47:55] What does that mean?
[00:47:55] Does that mean, as some optimists are saying, that suddenly GDP growth is going to radically accelerate and we’re going to cure all diseases and so forth?
[00:48:03] What does it actually mean?
[00:48:05] Yeah, I don’t think it means that.
[00:48:06] So I think there’s a lot of relatively low-hanging—
[00:48:10] low-hanging fruit in terms of things in materials and medicine and drugs and so on
[00:48:14] that’ll suddenly turn out to be tractable, like solvable problems that weren’t previously solvable
[00:48:19] because we can deploy these tools that are very good at analyzing these giant amounts of data
[00:48:26] and finding trends and patterns and helping us develop things that deal with our problems.
[00:48:30] So I think we’ll see a step-function improvement in those things.
[00:48:33] But at the exact same time, there are other countervailing forces that I think will hold back economic growth.
[00:48:39] The biggest one is demographics.
[00:48:40] Most developed economies are shrinking, and so population is going to be one of the factors.
[00:48:46] And there’s been some good research on this that, well, say, AI-related productivity improvements might take us out over the next, say, 10 years,
[00:48:53] maybe add maybe a little bit, maybe like 2% to productivity growth.
[00:48:58] On the other side, we have a shrinking and aging population, which will probably subtract 1% to 1.5% from that.
[00:49:03] So the gains are there to be had, but in some sense, they’re being offset by changes in demography,
[00:49:09] in the underlying makeup of the—
[00:49:10] And that’s not just true in the U.S.
[00:49:12] It’s true in Japan.
[00:49:13] It’s true in Western Europe and so on.
[00:49:15] But that doesn’t make this less important.
[00:49:17] It makes it even more important because we’re trying to compensate for changes in the nature of the workforce and the population of the United States.
[00:49:23] And there’s some great data showing that there’s huge advantages to jumping to some technologies earlier.
[00:49:29] Like, for example, Japan has one of the highest levels of robots per capita in the world.
[00:49:35] And that’s in part because it’s one of the oldest countries in the world.
[00:49:38] So weirdly, an aging society—
[00:49:41] —sometimes leaps to new technologies that help solve the problems faster than others would be.
[00:49:45] So I think a U.S. in the early 2030s will be one that’s much more robot-intensive for that reason,
[00:49:50] as a compensation for the changing demography of the country,
[00:49:53] that rather than having mass immigration or whatever else, we’ll have an increasingly number of robots.
[00:49:58] And that’s—I think you’re seeing the very beginnings of that right now, and it’s scary to people.
[00:50:02] But if you think it through, it’s actually hugely important because unless you’re willing to have, like, 10 babies over the next few years,
[00:50:08] this is not—it’s not going to change the—
[00:50:10] —it’s not going to change the underlying problem in terms of the nature of the U.S. workforce.
[00:50:14] So I think that’s incredibly important, that this is an offsetting force to huge headwinds that were already in place
[00:50:20] and we’re not going to be able to do anything about.
[00:50:22] So I’m really optimistic about some of the developments coming,
[00:50:25] but I’m also optimistic that it’ll help offset the demographic headwinds
[00:50:28] that might have otherwise been just dire for the U.S. over the next 10 years.
[00:50:33] Paul, it’s always fascinating and a pleasure to talk to you.
[00:50:36] Thank you so much for joining us.
[00:50:37] Yeah, thanks, Henry. It was great.
[00:50:40] Thank you.