Title: Paul Kedrosky: AI is the First Bubble With Every Ingredient at Once
Show: The Meb Faber Show #648 (host Meb Faber, Cambria Investment Management)
Guest: Paul Kedrosky (partner, SK Ventures; research fellow, MIT Initiative on the Digital Economy; ex-sell-side analyst)
Date: 2026-AUG-28
URL: https://youtu.be/xcftQ5HtDoU
Length: ~45 min
Note: Auto-transcript, lightly cleaned — pure fillers (um/uh/"you know"/"I mean" as interjections) and stutters removed; wording otherwise verbatim, timestamps unchanged. Garbles corrected: "Met Favor/Faber" = Meb Faber; "Camry/Kimri" = Cambria Investment Management; "Paul Kadski/Cadoski" = Paul Kedrosky; "Dario Amade" = Dario Amodei; "Barry Riddholds" = Barry Ritholtz; "DFC" = GFC (global financial crisis); "Michael Bur" = Michael Burry; "skyh highix" = SK Hynix; "Quen" = Qwen; "Toltoy" = Tolstoy; "Jeieves" = Ask Jeeves; "A6" = ASICs; "video GPU prices" = Nvidia GPU prices; "situational awareness" = Situational Awareness (the fund whose implosion is referenced); "Owen Lamont" is correct as spoken. "Edged" (the inference-ASIC startup in a Wall Street Journal story) is left as spoken — spelling unconfirmed.

(00:00) This moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in US history. I joke all the time that from the standpoint of most of the lenders that I talk to, there could be hide-and-go-seek competitions going on inside the data centers and they wouldn't give a — tokens are the first hyper-deflationary commodity in the history of modern economies.
(00:20) Good God, I hope I don't go crashing down because I need to grow at this speed just to stand still. >> [music] >> Welcome to the Meb Faber Show, where the focus is on helping you grow and preserve your wealth. Join us as we discuss the craft of investing and uncover new and profitable ideas, all to help you grow wealthier [music] and wiser. Better investing starts here.
(00:44) Meb Faber is the co-founder and chief investment [music] officer at Cambria Investment Management. Due to industry regulations, he will not discuss any of Cambria's funds on this podcast. All opinions expressed by podcast participants are solely their own opinions and do not reflect the opinion of Cambria Investment Management or its [music] affiliates.
(00:56) For more information, visit cambriainvestments.com. >> Summertime's over, everybody. So, we got an awesome guest today. Paul Kedrosky, a fellow at the MIT Initiative on the Digital Economy, partner at SK Ventures, weather ski nerd. I remember him from back in the RealMoney/TheStreet.com days, former sell-side analyst, a little bit of everything. Paul, welcome to the show.
(01:18) >> Hey, thanks, man. I'm glad to be here. >> I want to kick it off with a quote. He says, "It's going to take a lot of AI to solve the problems of AI." What do you mean by that? >> One of the things that interests me in general, whether it's this particular episode or the global financial crisis or the dot-com meltdown, take your pick, your favorite implosion historically all the way back to canals and railroads, is scale.
(01:42) Big things that move in ways that are both unexpected and more consequential than people realize. And there's a tremendous line from the physicist Albert Bartlett, goes something like the greatest failing of the human species is an inability to understand the exponential function, which is another way of saying that things that get scale quickly really confuse humans.
(02:04) Like the old idea that three periods before some algae covers an entire pond, right? So then in the last period it was only half covered and before that it was a quarter covered. And so that doubling notion of algae in a pond is another example of this scale. So AI is the canonical current example of prodigious scale and exponentials that are completely misunderstood by people, both by bulls and by bears.
(02:30) And I spend a ridiculous amount of time talking to people about this in all the different ways that the underlying exponentials for example are misunderstood. That there's an exponential adoption curve but there's also an exponential deflation curve. It's the fastest deflating commodity in the history of quasi-industrial commodities, falling something like 70 or 80% a year.
(02:52) And all of these sorts of intersections are really interesting to me. And the one that specifically triggered that comment was because the scale is growing so quickly and the interconnection into the grid can't keep up. People are increasingly being applauded for bringing their own — bring your own beer, bring your own power — adding natural gas turbines behind the meter.
(03:11) And that of course has consequences in terms of specifically energy and emissions. And so the joke always is Dario Amodei, Sam and lots of others love to talk about how AI is going to solve all these problems whether it's drugs, longevity or carbon emissions and climate. And it's kind of hilarious because one of the greatest contributors at the margin right now is this prodigious addition of behind-the-meter power usually in the guise of natural gas powering all of these new data centers.
(03:41) So long answer to a short question. It's going to take a lot of AI to solve the problems created by AI in climate alone. If you go back through the history of the largest financial and economic bubbles, paroxysm moments, whatever you want to call them over the last 200 years, they tend to have a bunch of things in common which is really interesting, right? That they tend to be related to maybe loose credit.
(04:00) They tend to have a great technology story behind them. They sometimes have a real estate component behind them. And they sometimes have a policy angle. You take that back to the South Sea bubble and there's a policy angle. To my way of thinking, this moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in US history.
(04:19) And that's what makes it distinct, unusual, and makes the scale really difficult for people to understand because they approach it through the lens of credit or they approach it through the lens of real estate or the geeks on X approach it through the lens of five prompts that'll change your life or whatever the case may be.
(04:34) And the reality is this moment sits at the intersection of all of these. So you can go backwards and think about the railroads. You can think about the canals. You can think about rural electrification in the 1920s. You can think about all these scale moments where massive amounts of capital were allocated towards something hugely consequential and then created a lot of breakage along the way.
(04:54) And I think there's lots to learn from those moments and increasingly people are interested in it. It is really tough for humans to wrap our hands around these giant numbers — a billion, 10 billion, 100 billion, trillion. One of our old podcast guests, Owen Lamont, had a great quote where he said, "Stocks are unreasonably high when prices rise to a valuation level that cannot be justified by rational forecasts of subsequent cash flows, then they double." You know what's funny
(05:24) though, and it's true, but the reality is humans are so myopic about this stuff. There was that famous bumper sticker at the end of the dot-com crisis you used to see around the Bay Area and it went something like, "Please God, give me just one more bubble, this time I'll know what to do." And it turns out they don't know what to do.
(05:40) And they didn't know what to do in the financial crisis. They didn't know what to do this time. So, as much as it's fun and you feel like people would learn from it or at least have a good time on the way up, it doesn't work out that way. Humans — and I'm stealing from our mutual friend Barry Ritholtz's playbook on this stuff — but the behavioral aspects of how people approach capital, money, manias, and fads doesn't change.
(06:05) They panic at all the wrong times. They chase things at the wrong times. And it just seems to be uniformly true about how people approach these things. And it's not because humans are stupid. They're just doing the best they can with the limited information they have when faced with these massive phenomena at scale that exceed their ability to think concretely about them.
(06:25) And this is such a good example of it. All the same phenomena that we saw in the GFC in the lead-up are all in place again now. They just have different names. And so that's what makes it really fun and, as you say, fun and interesting. >> This episode is brought to you by Upwork. Every business owner hits a point where they need more expertise than they can handle alone.
(06:44) But another full-time hire isn't always the answer. And that's where Upwork comes in. It's where growing businesses find highly skilled freelance specialists. Not just one-off tasks, but you could build an entire team without the commitment of permanent headcount. From software development and AI implementation to marketing design and operations, you can browse profiles, review past work, get help scoping out the role.
(07:06) Upwork handles contracts and payments all in one place. And with Business Plus, you can access the top 1% of talent. Visit upwork.com right now and post your job for free. That's upwork.com. >> When you look around, the hard part with these big numbers is often it just leads to storytelling and narratives.
(07:28) And unless you can describe something in terms that I feel is relatable to people. One example that I think you were talking about was these big AI companies, these stocks are now bigger than the entire energy sector combined in the stock market. >> Not just that. I'll give you another example just because the opportunities to tell better stories about this stuff are legion because, again, blind-man-and-the-elephant stuff, people approach it and from one person's standpoint this is all
(07:58) just a mad race to AGI. From another person's standpoint, this is just a better way to read my email or whatever else. And so people have these really idiosyncratic ways of looking at what's going on, which makes for confusion in terms of how they interpret these scale phenomena. And for example, one of my favorite examples in all of this stuff is that they sample from a very skewed sample when coming up with their projections of what, say, the future of token use is going to look like. Just to pick an AI
(08:27) related scale example and not crypto tokens, but AI tokens. In the early days of AI, one of the most effective places for large language models to operate was in this domain with a very strict grammar, a tight what's called gradient descent, meaning that you learn a lot from small errors.
(08:45) And that's how models in general improve, is through this process of iterative gradient descent, and a domain where there's a real consequence to being wrong. And that domain of course is software itself. Software has a tight gradient descent. If I start changing things inside of software with the idea that, you know what, I want this subroutine to be much more aesthetically pleasing — the equivalent of painting it pink — that doesn't really work very well in software.
(09:09) What you get instead is a non-functioning piece of software that may look aesthetically pleasing, but doesn't do anything. That's not the same with a grammar, with an essay about Tolstoy or the history of Lolita or something like that. Small changes are just like, yeah, that might be an A also, I have no idea. So software is really unusual in that sense, but it's also really unusual because most white collar applications in AI are what I call compressive, meaning that you, for example, receive a bunch of sell-side research like I do
(09:36) probably. And you look at it and you're like, "Oh my god, shoot me. All of this stuff is coming flying in. How do I know if anything in here has any signal whatsoever that I need to pay attention to?" And more importantly, is there any differential signal? Not just is it valuable, but how is it different from what I saw yesterday? I need to see, because the markets move on differences in expectations.
(09:54) So I want to understand that. And so the applications for white collar workers, unlike coders, tend to be compressive in the sense that I take a lot of information to turn it into a little bit. I take a huge sell-side document 40 pages long and say give me five bullets. Software is exactly the opposite. It's expansive.
(10:14) I write a prompt to say give me a fitness app that does this, and what comes out the other end? A million lines of code. I'm not trying to compress. I'm trying to expand. So all of this is a long way of saying that the first domain where AI was applied could hardly be less representative of AI's future if you tried. Meaning that it was expansive.
(10:33) It had a tight grammar and had a rapid gradient descent. Almost no other domain in which we operate in economic life has those properties. And yet that's the domain from which we're extrapolating our futures. >> Where does it progress from here? >> Well, it progresses in the sense that token usage no longer rides this geometric curve that was created by all of these early expansive applications that had this tight grammar and the gradient descent.
(11:00) So the point is that the projections that were made in the first four years of large language model adoption are largely useless. It doesn't mean large language models are useless. It just means that no different than — we say this a lot in venture capital — that you have to be very careful who your first customers are because early adopters aren't like anybody else.
(11:18) They're crazy people who try half-baked products, and if you try to build things for them, in all likelihood you'll never find the true valuable customers. And I say that all the time in venture capital when people come to me and they'll say, "Hey Paul, I have this application, it would be perfect for you."
(11:33) And I always respond with, "Well, then you're — because I'm completely weird and nothing that was perfect for me is probably good for five other people on Earth." In a weird way, AI has started in that strange world of building for the strangest early adopters possible, coders, making a perfect tool for them and then extrapolating naively about what that means in terms of what kind of data center capacity we need to build and what that means for internal versus external financing.
(12:01) Can we do it from cash flows? Do I have to do external financing? What should the hurdle rate be? Are these commercial real estate projects? Should I be syndicating the debt? Blah blah blah. So hugely consequential, this really one simple decision about I built a great thing for coders and it started there. >> And so you've got to answer all those questions you just asked.
(12:19) What are the answers to those? >> Well, the answer, and this is what we've learned from prior financial episodes, is you reach a point of escape velocity where it doesn't matter anymore whether or not the applications match the initial indication, right? No different than in a drug context.
(12:37) And so it creates a flywheel. The flywheel in the '07-'08 financial crisis was this idea of being able to push debt off my balance sheet such that I no longer face the liabilities that might come due if this stuff turned out to be toxic. And that created a big problem with asset-backed securities and everything else and all the different syndication things that went on.
(12:58) And so whenever you see, sometimes called a Minsky moment, whenever these things become financialized and it becomes divorced from the underlying application, you need to pay a lot of attention. And that's the moment we're in right now. So in answer to your question, we've now in 2026, the first half of '26, we crossed over from more than half of the financing for data centers specifically, but let's call it hyperscalers at large.
(13:19) It no longer comes from internal cash flows, because that was always the line, right? Is Paul, shut up, because these are smart companies. If they want to spend their cash flow this way and they have a lot of it, who are you to tell them not to? These are smart people. And I'm like, oh curses, you've got me there. And of course, what's happened was the predictable — it has now flipped.
(13:36) Now most of the financing as of mid-2026 is external financing coming from a constellation of things like ABS, private credit, sovereigns and everything else, and all of these SPVs and new structures, because of the rate at which these data center expenditures are growing, in part because of this flawed sampling at the original sin, extrapolating from coders' usage.
(13:58) And so now, of course, what's really funny to me is when I get into these discussions with people, the people who used to chide me and say, "Hey, this is all coming from cash flow. This isn't a problem." Now say, "Hey, it's all coming from outside investors and they're really smart. Why do you care?" And I'm like, "Wait, wait, wait, wait.
(14:12) We just had the exact same conversation 180 degrees turned around, where you told me before I should be really worried if the money's coming from the outside, not from cash flow. And now it's coming from outside, not from cash flow, and that turns out to be an indication that all is well." I said, "This is a narrative problem. This motivated reasoning is what it's usually called.
(14:32) This idea that I'm telling a story because it makes me feel good, not because it has any basis in making the world work better." So, we're in this moment now where this flywheel has begun. There's so much money flowing into data centers and the AI complex in general that it's distorting sovereign flows.
(14:46) It's one of the reasons behind the bond freakout we just saw because it's pushing up longer-term rates. It's much more appealing to loan to a data center and get a hyperscaler with a prime credit implicitly or explicitly backing it than giving it to this flawed credit known as the United States or the UK or whoever else.
(15:03) And that's one of the reasons why — we used to say, and you know this better than I do, that sovereigns being active in funding markets would squeeze out privates. You want to be careful how much debt-based financing you're doing as a sovereign because you'll push out the privates. And the funny thing is we're doing the exact opposite right now.
(15:21) Sovereigns are so flawed in terms of their own fiscal, their own balance sheets, that their fundraising is being challenged by the fundraising of the data centers, who in turn, their fundraising is increasingly divorced from what's actually going on inside the data center. I joke all the time that from the standpoint of most of the lenders that I talk to, there could be hide-and-go-seek competitions going on inside the data centers and they wouldn't give a — it just doesn't matter to them because they look through the data center and say there's a prime
(15:45) credit on the other side. It's a 12-year lease renewable and they're good for it and this is secure cash flow much better than holding a 10-year or anything else. And so I'll hold this and then at the end I look through whatever is going on inside the data center. And that of course just leads to much more data center construction and far more GPUs and a lot more SK Hynix high bandwidth memory, NANDs and everything else.
(16:10) So you have this incredible flow of money purchasing things that's increasingly divorced from what's going on inside the data centers, which again has no bearing on whether or not large language models are useful. They're wildly useful. In a weird way that's required setting in the system. That's what makes it into these kinds of moments.
(16:27) A really good story driven by technology that works. If it was shitty technology that made no sense, sort of the beanie baby level stuff, we wouldn't be having this conversation. What drives these incredible moments in history, like the role that the railroad buildout played in the golden era of the 1920s and in a sense into the crash of the late '20s and the Great Depression, was that electrification was a great story.
(16:50) It was incredibly consequential. One of the most important technology stories of the 20th century. And this is that just happening at much faster speed, with the capital allocation that happened over 60 years in electrification happening in four here. Well, less actually. And that's hugely consequential.
(17:08) I just find that really interesting, the way these processes get divorced from the underlying economic realities. I mean there was a great study came out the other day showing that in one warehouse — sort of think of it like an AWS of GPUs — that the GPU usage is only sitting at around 35-40%, which should be eye-popping in terms of utilization rates. Because you ask yourself, wait a minute, I was told a story about scarcity and rising prices and that A100 prices in the used market, Nvidia GPU prices, were
(17:40) being pushed higher because of their scarcity. And now I'm hearing this idea that they're all sitting there and being used at less than 50%. How do I square these things? And what you're seeing is the classic dynamic that exists at this point where there's hoarding going on, colossal amounts of hoarding going on, double and triple ordering going on because everyone's terrified that if something happens, some inflection, some parabolic moment, even more parabolic than what we've seen, that
(18:05) they'll be caught short, won't have it, or they won't be able to purchase it if they need it. So you have this incredible amount of hoarding going on and all of these things sitting unused even at peak loads, which is remarkable because it represents a kind of potential flux of product into the market at some point in the future. And yet there it is, totally predictable, and this is the way these moments tend to play out.
(18:29) >> Does that potentially end up benefiting the end consumer, and all these AI companies and sovereigns and private credit just nuked a bunch of money on oversupply? Or is that a short-term thing, or what do you think? >> No, no, no. That's exactly what happens. Tokens are the first hyper-deflationary commodity in the history of modern economies.
(18:50) Because they're under immense price pressure both because of the technology curve they're riding and because of the massive amounts of capital flowing in, we should expect a series of waves. The way I think about it is that as tokens brush up against other parts of the economy, they do huge damage to it because it's like being brushed up against by some magic deflationary force.
(19:10) So they enter it and all of a sudden you see huge deflationary waves entering that sector because you're brushing up against this structurally deflationary force that's deflating around 70 to 80% a year year-over-year, which we've never seen before, and it's not easy to see how that stops for a bunch of structural and technology and financial reasons. So yes, it'll probably hugely benefit consumers. It's toxic to the frontier companies. Because you can do the simple math. If I need to stand
(19:41) still in terms of my unit growth — let's just say stand still — and the price is falling 80% year-over-year, and this goes back to Albert Bartlett's line about exponential series. How fast does my unit growth have to be? Well, the answer of course is 400% minimum. So if it's falling 80% year-over-year just to stand still, leaving aside just trying to please Wall Street, and leaving aside whether or not the margins are sufficient to cover the debt load that I have to pay off because now all of the financing has become
(20:10) external, you're in a terrible situation, my friend. This is among the most difficult things in capitalism. Just to stand still I need to grow 400% year-over-year. So that's the lens through which you need to interpret the kinds of numbers you see from the frontier model companies, because they're in this Wile E. Coyote roadrunner thing where they're literally standing over thin air with their legs spinning saying, "Good God, I hope I don't go crashing down because I need to grow at this speed just to stand
(20:37) still," let alone producing. What is Wall Street going to want from these guys? 50% year-over-year growth, 100% year-over-year growth. OpenAI apparently in its last quarter did 18% quarter after quarter, and that was deemed a disappointment. So, this tells you what we're up against here. Not just the expectations but structurally the economics of riding this deflating exponential are so consequential in terms of both the kinds of top-line growth you can deliver but also in terms of paying the fixed obligation that comes from having more
(21:09) than half of the financing for these things coming increasingly from debt, not from cash flow, to the point that — I think I just saw this, I think it was a JP Morgan report — that something like 15 to 18% of the investment grade marketplace is now data center related. It's now, if it were a sector, which it isn't, it's now larger than the financial services sector in terms of its share of the investment grade marketplace, which is just — I'm old enough to remember that one of the prime attractions of technology after
(21:36) growth was that it had no debt. The balance sheet was pristine. These were cash flow monsters. We've completely reversed it where they're now encumbered by more debt and it's growing, to the point — and they also have maintenance capex, as Michael Burry likes to point out all the time — that this isn't a one-time phenomenon.
(21:55) I'm going to have continuing capex obligations into eternity to keep these data centers at the cutting edge. In a sense, turning orthodox technology companies into a kind of utility company, right? With similar obligations, but wildly overvalued compared to an orthodox utility. So, you have to ask yourself, who gets re-rated first? Do utilities all of a sudden see a huge spike in valuation, or do technology companies, especially the ones most tied to the capex wave, get re-rated to look more like utilities? And I obviously
(22:25) think the latter. >> You had a chart, I think that was one of the most jaw-dropping charts I've seen probably this year. And so, speaking of Wall Street, if you go back to the late '90s, crazy time, and you had all these things about this bubble bull market, high valuations, people doing crazy stuff, on and on, right? There's one that had been missing for a long time that was littered throughout the late '90s, which was supply and IPOs. And then you had a chart where it was like, if these three or four
(22:56) companies go public — SpaceX, Anthropic — they're going to be bigger than all the IPOs from the '90s combined. >> Well, not just that, it's all post-World War II combined [laughter]. So it's staggering, right? >> It's staggering. And I looked at that and I had to sit with that for a while, because for me, we talk a lot about supply and dilution and funding and IPOs in companies.
(23:24) Now it's even shifting from companies like Oracle, like buybacks versus share issuance and dilution. Historically, all those things are shifting, and shifting pretty hard in the other direction. >> And what's interesting is it goes back again to the early part of our conversation about scale and systems. Let's think about it.
(23:43) Okay, I said $4 trillion in new issuance from that narrow subset of companies alone. It's going to be more than that now. It's going to be more like 5.5 trillion, which makes the top of my head pop off. Okay, fine. Now, put that on a t-shirt and run down the street with your hair on fire. Okay, you happy? Good.
(24:01) Okay, so let's move on now and think about what are the consequences of that? I'm a large long-only manager. I'm running an index fund. I'm a hedge fund with a decent-sized short book. The consequences get really interesting because people, especially retail investors, they tend to act as if these large funds are just sitting on a printing press of cash in the basement.
(24:22) So when they want to make a purchase and buy a large allocation to one of these upcoming IPOs, they just turn to somebody down the hallway and say, "Hey, Mary, get me some money. I need a piece of this Anthropic thing." And of course, it doesn't work that way at all, right? Cash is a drag for the most part on performance, especially of long-only funds.
(24:39) And so the question then, where's the money coming from? Well, the money in general comes from selling other stocks. I have to free up cash to make the purchases. Let's put inflows aside. Okay, fine. What gets sold? The things that get sold, generally speaking, have a couple of characteristics that make them really interesting.
(24:58) The things that get sold tend to be the most liquid because I don't want to have a huge price impact. They tend to be, weirdly enough, overlapping, because I don't want to duplicate my exposure. I don't want to accidentally end up over-indexed to some factor that I didn't really realize I was getting over-indexed to.
(25:13) So, I tend to sell things that are liquid, that look like the thing I'm buying, and perversely some of the things that are performing the best, right? Because I want to lock in my gains in other things so that I can later on write my quarterly letter and tell everybody what a good boy or girl I am.
(25:29) And so you can model that out. You can model that out fairly straightforwardly, and I did back in March or April, and showed what the impact would be and that it would actually start long before then, because people needed to start freeing up that money much earlier, because you can't do it all. It's not like I need to borrow some money from my dad so I can buy skis or something the night before.
(25:50) I need to free up that money weeks and even months in advance and do it in a way that the market's not paying attention so I don't look quite as bigfooted and price distorting, right? And so you can anticipate that in the April, May, June time frame that that was going to happen, that there was going to be a lot of pressure on the winners, that some of the most liquid names were going to get sold.
(26:11) And in a weird way, and I was talking to one of the derivatives guys at JP Morgan about this, I think that played into the implosion at Situational Awareness, that they were hugely long some of the most liquid and best performing names for obvious reasons, but those things came under early pressure because they were going to become a source of funding inevitably for the upcoming spate of IPOs.
(26:33) So in a perverse way, anticipation of the upcoming flood of new issues became a pin that pricked at what was going on inside of Situational Awareness. Now granted, lots of other things began happening and people realized what his book was and he was under pressure and everything else, but there was more to it than that.
(26:52) This was structurally predictable in the context of a system in which money at an unprecedented scale had to be freed up from over here to put it over there. >> Today's show is sponsored by Cambria. Do you hold legacy investment positions with significant gains? What if you could transition into an ETF without facing a large tax bill? You can with a 351 ETF exchange. Here's how it works.
(27:15) Investors contribute stocks or other securities to a newly formed ETF in exchange for ETF shares. As long as the special rules and diversification requirements are met, the investor is essentially able to seed the launch of the ETF without an immediate taxable event. Because ETFs typically don't distribute any capital gains, investors don't face taxes until they sell their ETF shares, allowing for better control over the timing of the tax event.
(27:40) Are you ready to explore a 351 ETF exchange? Visit cambriafunds.com/351 to take the next step in innovative tax-savvy investing with Cambria today. Cambria Investment Management LP. Cambria is a registered investment adviser. The information set forth herein is for informational purposes only and does not constitute financial investment, tax, or legal advice.
(27:58) Past performance does not guarantee future results. All investments are subject to risk, including the risk of loss of principal. >> There was an interesting quote Derek Thompson had on Twitter the other day where he posted a chart that Nike is down 75% from the peak. And he's like, "What happened to Nike?" And there's hundreds of comments.
(28:13) And the comments almost universally are like, "Oh, the product, the competitors, or politics, Kaepernick, macro," whatever it may be, all the things you would expect people to talk about. And then I was like, none of them mentioned the fact that the P/E ratio at the peak was like 70 and now it's 20. >> No, I think that's spot on. And leaving aside the ex-post business of people saying, "Shoe XYZ sucked, man.
(28:36) This is why Nike's down." And I use this line a lot when I'm talking about AI because people always are on about, well, what's going to cause this moment to implode? And I said, here's the thing, and Nike is a good example. At high valuations, failure is overdetermined in a statistical sense, meaning that there are so many ways to fail, all of which are low likelihood.
(28:58) All of which are low likelihood, maybe 5% at most, that when you combine them all and turn it around and say, given 20 different ways this could fail because we're at this precipice, it's high valuations and everything else. And there's so many different ways that this moment could end. What an overdetermined system tells you is that let's say I've got 20 different ways this thing could fail, each of which has no more than a 5% chance of happening.
(29:22) If you flip it and do the math, it's greater than a 60% chance of failure in the period that the 5% applies. So what looks not particularly risky, or it looks like it's unpredictable, is actually highly predictable. Because as you say, when your P/E is 70, failure is overdetermined. Lots of ways things can go bad because the market increasingly looks at it and says, "How can this thing continue to outperform expectations?" And the answer is, whoops, it can't because of this, oh whoops, it can't because of that. And it starts
(29:52) pointing around. And the same thing is true in this AI moment, that there's so many different ways that the eventual implosion of a lot of what we see going on can end, because again it's overdetermined. It can be sovereigns. It could be rogue AIs. It could be — I've been pointing lately to the incredible and unprecedented cash inflows into the Taiwanese chip and Chinese chip manufacturers in terms of what looks like a tsunami of supply in early 2028.
(30:22) And this is a boom-bust industry. So once you lock in supply, my friend, prices are going to zero. Why? Because I've got to cover my fixed cost. So this stuff is all going out the door. It's not going to sit around. I'm not going to hang on to inventory. And we see it already being built for that.
(30:39) And the semiconductor industry has a history of exactly this. If you look back, it's probably the most capital intensive boom-bust industry on earth. But from a system standpoint and a scale standpoint, it's a classically overdetermined system where in 10 years people look back and say, "Oh, remember when Micron was at such and such a price, and here's what went wrong.
(30:57) " And they'll come up with a bunch of ad hoc explanations of what went wrong. And the reality is it's exactly what you just described in the context of Nike. It was just really expensive and anything at that point could have been seen as consequential enough to bring it down, and then it just got — what's the old line — kind of got pecked to death by ducks, man.
(31:15) And it just kept getting pecked all the way down and different things took it lower and lower and lower, and then eventually it finds itself a bottom. And that's what we're going to see happen in some of these areas where there's a this-time-is-different narrative, and the current one is in semiconductors, where the argument is we're in this kind of semiconductor super cycle, that RAM will never come back, that the GPUs are inevitably a duopoly and are probably more like a monopoly, and it's the
(31:40) most consequential computing industry on earth and all these things. And yet recently there was a story in the Wall Street Journal which made me laugh, about this new inference ASIC — application specific integrated circuit — company called Edged that had, I don't know, a couple of quadloopzillions of capital or something that they had raised, and it was a bunch of young guys and they did the usual story about telling the story of the brash young guys who create the thing that
(32:08) goes up against the incumbents and yada yada yada. But that's not very interesting because the semiconductor industry historically, as a venture capitalist, has been a faking elephant's graveyard. You just go there to die, my friend. So the question is, what's different about these guys that's gotten them to the point where they are?
(32:24) And what's different and interesting is that historically, there's a bunch of stages in terms of producing chips, where you might do the initial design specification and then you begin doing verification, then you do these test cycles and eventually if everything works, which it never does, you tape it out and then send it off to a fab and it gets manufactured.
(32:44) It was 42 days in the initial design verification stage versus what should have probably been six or seven months. And not only that, their first design worked, which is sort of like saying, I don't know, I walked down the street and married the first person I met. That just doesn't work out. It just doesn't work out. And so the real story is this is one of the first companies that's structurally using AI to produce chips in a really fundamental structural way, because code in the context of chips is exactly in the sweet spot of what large language
(33:13) models do really well, for the reasons I was talking about earlier. And so what you should expect is — and I was making this joke the other day to somebody that instead of having vibe coding, we're going to have kind of vibe chipping. People are going to be producing chips at rates that we've never seen before in terms of new designs showing up at fabs, and all of the old barriers are going to collapse.
(33:34) And so the notion that some kind of tribal knowledge protects you as a chip manufacturer is going away. And so all of these companies are going to be amazed at how many different designs of, say, low power ASICs that do on-chip large language models at the edge inside of security cameras.
(33:53) They're going to come out of nowhere. They're all going to be coming out of nowhere because the process of producing chips itself is being transformed by AI, which in a perverse sort of way will help prick the bubble itself. >> Is it a scenario where there's five, 10, 20, 50 of them that are all good, but it's hard to distinguish between one being slightly better than the other? Or is it a scenario where maybe Google bubbles up above, as with Ask Jeeves and all the other search engines, and there is kind of a
(34:28) couple winners? Do you have any guesses, predictions, thoughts there? >> You've hit one of my favorite crusades, kids, and I'm really sorry for you. So I go on about this all the time and here's the thing: we're kind of in the iPhone 7 moment when it comes to large language models, where weirdo geeks are showing up and lining up at the store and everyone else looks at them and says, "Dude, I can't tell the difference between that one and the two before.
(34:54) " Large language model performance really hit its maximum inflection in terms of year-over-year change almost four years ago, 2022, 2023. And if you look at things like composite indices — and I'm not talking about these really gamed benchmarks that get promoted all the time because the models increasingly ingest those benchmarks. And so the results mean nothing, no more than if you had seen the SAT before you'd done it.
(35:18) And so if you look at these composite benchmarks, what you see is essentially two things. One, the year-over-year gains in the models themselves has essentially flatlined over the last 6 months. Minor changes, not no change, but where we once might have seen 10 or 12% year-over-year changes in the composite indices, we're now seeing one or two% at most.
(35:38) And that's masked by a couple of things. It's masked by the emergence of this class of things we call harnesses. These are things like Claude Code and the codexes that wrap around the model and give the appearance of things changing much faster than they are. The analogy I use all the time is it's kind of like Sound of Music where the models are kind of like the bratty kids and the harnesses are kind of like Julie Andrews' character, whoever she was.
(36:00) She's able to take a bunch of bratty kids and make them sing. Harnesses do that with models. They essentially harness them and make them look more functional than they are. But they're masking that despite 18-month training cycles and billions of dollars, models increasingly are flatlining in terms of their performance.
(36:16) And then the second factor is that they're not just flatlining, they're also converging. So if you look at the variance across all models. So look at the performance of most of the models that are being shipped in any given year. There used to be immense variance from the best to the worst, or worst to the best, whichever way you like to think about it.
(36:33) And that variance has collapsed. There is very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek or somewhere else, in practical composite terms. People delude themselves and say there's a huge difference. But I do some testing every once in a while just to piss people off where I'll literally do like a Pepsi-Coke test and say, "Okay, you think you can tell the difference? I'll put them in front of you and see if you can.
(36:56) We'll hide it behind a harness." And inevitably no one can tell the difference. Everyone thinks they do and they can't. So this is a really startling thing, right? That goes to your point. Not only will that moment come, it's here. The moment is here where models are both converging, where it doesn't make much difference which one you choose behind whatever harness, and the year-over-year performance is really falling off asymptotically and even flatlining and being hidden by harnesses. So the game is almost over in
(37:24) terms of pretending that you can justify multi-billion dollar training runs to produce some new name model. And I joke all the time that the most successful frontier AI company will be the first one to stop pretending they can train new AI models, because it's basically a gift, right? I don't have to spend that money anymore.
(37:44) There was a great piece in the Economist the other day about this which was completely wrong but it was still interesting, arguing that chat models in general would break us out of the trap of the TikTok-ification of tourism. And this idea that people increasingly get the same signal and respond with the same action, which is then they show up and I'll take a picture of some fountain in Milan or something, and that's this TikTok convergence and herding which we see in financial markets, which we see in lots of other places when everybody's exposed
(38:11) to the same signal. And their argument was models are much more divergent and so we'll see less herding, and that is completely untrue and you just gave two good examples of it. But the reason why, of course, as soon as you step back and you look at the training data on top of which these things are constructed.
(38:27) And I use this line all the time, but the median data nugget inside of a large language model is a 37-year-old male on Reddit. So if you back-extract where the signal is coming from, the homogeneity of the result isn't very surprising because of the homogeneity of the convergent data underneath it in terms of what's producing the recommendation.
(38:49) So, if you ask yourself, what would the average 37-year-old dude on Reddit tell me? And do that and then compare it to what a large language model says, you're probably in pretty good shape. >> I saw a stat this week that said something like 75% of people said they wouldn't want a data center in their county, which was higher than nuclear.
(39:08) Are they being rational or is this a PR problem where people just aren't explaining data centers? What's going on here? Because it seems like it's kind of a mass universal distaste for data centers and this whole AI sort of umbrella. >> Yeah. So, I have a couple of takes on it.
(39:31) I feel like people increasingly feel as if they've lost agency in their own lives and data centers are both a physical and digital manifestation of that. It's a great hulking presence that I didn't ask for and it's like, okay, this is pissing me off and I don't want it. So there's this notion of lost agency in their physical lives, but also this idea that it's a representation of lost agency in their digital lives, in their relationships, and how they deal with things and whether or not I'll have a job and everything else. So I think some of this
(39:58) is you can wrap up in this ball of things that people feel as if they've increasingly lost agency, and it plays out in all sorts of perverse ways and particularly in the United States with people's ideas about vaccines and other things, right? Where people feel like they want agency and they resent when people take it away from them, even when it was a good idea.
(40:18) It doesn't matter. I still want to be able to make the decision myself. Stop telling me what to do. This is a uniquely American phenomenon. The other piece of this that I think is interesting is that historically the US is among the most aggressive and early adopters of technology in the world.
(40:32) Like we've never seen a digital toaster we didn't like or whatever. And so what I think is interesting is how the US is anomalously negative about AI compared to other G7 and OECD countries. And once you take the data and look out even more broadly, the OECD countries in general are more negative than, say, sub-Saharan Africa, than some developing countries.
(40:57) And so that's really interesting, that inside of developed countries, people increasingly see AI as a force for homogenization and lost opportunity, that they're going to take things away from me. And I blame Dario and others for this in part, for all of the "some large percentage of jobs are going to disappear."
(41:15) And you tell people that and they take you at your word and they're like, okay, I'm out. Whereas in sub-Saharan Africa and other places, it's seen as a leg up. It's like, okay, I'm never going to get to go to Harvard, but if this thing makes it such that I can feed my family and I can find a way to educate and don't have to go to college or whatever, I'm in, right? I'm in.
(41:32) These are things that they feel like a great leap forward for those regions. The other thing that's going on which is uniquely American, that makes it break from the OECD which was already somewhat negative, is the healthcare consequences. Uniquely, in the United States any threat to employment is a threat to health care, and a threat to health care is a threat to fiscal, your personal solvency. And that is not the case in almost any other Western country. People are like, yeah I don't like it, but it's not as if
(42:00) whenever my job goes away, if I have to go in for some dental work or some unexpected health care that suddenly I'm bankrupt. So the notion that the US has tied health care so tightly to employment makes loss of jobs highly consequential, and we've largely promoted AI on the basis of job losses. So big surprise, people look at that and say, "Yeah, no, this is not for me.
(42:22) " And that's a uniquely American problem. >> One of my favorite things that y'all do — and we're a subscriber of this — is put out these great charts. Are there any charts that really stick out to you that you're like, "Wow, this was a banger. This was super interesting." >> One of the things that got me into thinking about AI in terms of the system consequences was realizing that it was a large and growing share of GDP growth.
(42:43) Just that statistic. And when I put it out it was mostly to satisfy myself, because I figured if I put it out there someone will tell me I'm wrong and it'll be like, okay fine, I screwed that up and it's actually not as large of a share as I thought. Not only did it turn out that it wasn't wrong, that share grew and has persisted over the last six quarters since I first began writing about it.
(43:05) So those kinds of things, as opposed to novelty charts that X isn't as Y as you thought it was kind of stuff — that's okay. The stuff that makes you take your cognitive axes and the ways you think about the world and rotate them and say, "Oh, wait a minute. This is really unusual." Not just because it's this crazy moment, but because for probably the sixth time in Western history, we have some force that's non-governmental that's actually so large as to shift the tides of GDP growth in and of itself. And what does
(43:36) that mean? What does that mean? I use the line all the time, but it reminds me of my dog, right? My dog will bark when the mailman comes to the house and of course triumphantly the mailman leaves and the dog's like, "Dude, I did this." Right? But the problem of course is even if he didn't bark, the mailman would leave, right? So he has a problem with causality, right? Dog causality is really messed up because he feels as if he made the mailman leave.
(44:03) So turning it back to GDP growth and that chart, what it tells you is that if you thought your policies, tariffs or whatever else, were causing US GDP to grow in this way — it was because it was more than 50% of US growth — you have a messed up model of causality because you don't understand what's actually driving growth.
(44:20) And so you might be tempted to take actions that are actually really consequentially negative because you didn't understand that, like the dog barking making the mailman go away. Your tariffs did not cause the US GDP to grow in that period. What caused it was this remarkable thing called data centers and hyperscalers.
(44:37) >> Paul, best place for people to find you. Where do they go? >> paulkedrosky.com. It's all there. >> Perfect. We'll put it in the show notes. Paul, thanks so much for joining us today. >> Yeah, glad I could do it. >> Podcast listeners will post show notes to today's conversation at mebfaber.com/podcast. [music] If you love the show, if you hate it, shoot us feedback at themebfabershow.com.
(44:59) We love to read the reviews. [music] Please review us on iTunes and subscribe to the show anywhere good podcasts are found. Thanks for listening, friends, and good investing.