Paul Kedrosky warns of AI financial fragility, debt risks, and labor market tail risks
Summary
Paul Kedrosky argues that the current AI investment cycle is unique due to the simultaneous convergence of technology, real estate, credit, and government policy, creating systemic fragility. He compares private AI valuations to the dot-com peak and predicts a prolonged corporate deleveraging phase similar to Japan's balance sheet recession. Kedrosky highlights NVIDIA's role as a 'central bank' running a shadow banking system and warns of tail risks from labor market repricing and cybersecurity vulnerabilities. He advises investors to focus on inexpensive long-dated call options with high leverage potential.
Episode sections
- 0:00 – 1:00
AI Investment Framework: Four Pillars
Speaker 0 argues that the current AI investment cycle is unique because it simultaneously combines four historical drivers of economic booms: technology, real estate, credit, and government policy. He notes that while past bubbles typically featured one or two of these, the current moment sees all four active, creating a complex system prone to fragility.
- 1:00 – 11:07
Speaker Background and Systems Thinking
Speaker 1 introduces Paul Kedrosky, highlighting his PhD in complex systems and work at MIT. Speaker 0 outlines his background as a mechanical engineer turned finance professional, explaining that his interest in AI stems from studying systemic fragility, contagion, and tail risks in financial markets rather than just the technology itself.
- 11:07 – 13:17
Historical Parallels: Dot-com vs. GFC
The speakers compare the current AI boom to the dot-com bubble and the Global Financial Crisis (GFC). Speaker 0 distinguishes the dot-com era as an equity-focused bubble contained largely within public markets, whereas the GFC was a credit and balance sheet crisis involving the shadow banking system.
- 13:17 – 15:45
Private Market Valuations and Deleveraging Risk
Speaker 0 claims that the nominal valuations of private AI firms like SpaceX, Anthropic, and OpenAI are equivalent to the peak NASDAQ valuation during the dot-com bubble. He predicts a prolonged period of corporate deleveraging similar to Japan's 'balance sheet recession,' driven by companies prioritizing debt reduction over expansion despite low interest rates.
- 15:45 – 21:06
Technology vs. Financial Fragility
Speaker 0 rejects the dichotomy between believing in AI's technological merit and fearing its financial bubble. He asserts that while AI is a consequential technology, historical capital expenditure booms (railroads, electrification) inevitably lead to overbuilding and financial damage, regardless of the technology's ultimate utility.
- 21:06 – 23:29
Exponential Growth and Token Economics
Speaker 0 describes AI tokens as a hyper-deflationary commodity, citing a 13-fold year-over-year price decline. He argues this creates an exponential demand for volume (doubling every 95 days just to stand still) and highlights the fragility arising from the assumption that demand elasticity is infinite.
- 23:29 – 26:03
Cybersecurity and Systemic Fragility
Speaker 0 points to a surge in critical software vulnerabilities, noting that 8% of all historical critical vulnerabilities were found in the first three weeks of October and 50% since January. He argues that AI's expanding attack surface creates a precarious balance where defense struggles to keep pace with offensive capabilities.
- 26:03 – 32:09
Concentration of AI Spend and TAM Overestimation
Speaker 0 cites data from Ramp showing venture-backed companies are the largest per-employee spenders on AI. He warns against extrapolating Total Addressable Market (TAM) from these early adopters, as software engineers are uniquely suited for AI due to the clear 'gradient descent' signals in code, unlike other white-collar work.
- 32:09 – 33:04
Oracle's Debt-Fueled Transition and Fragility
Speaker 0 identifies Oracle as a 'destination of fragility' due to its heavy reliance on debt to fund a transition from declining monolithic database business to AI. The company faces potential credit rating downgrades as it balances this pivot.
- 33:04 – 34:04
Neo-Clouds and Crypto Dependency
Neo-clouds are described as 'recipients of fragility' that are using massive debt to pivot away from crypto mining, which became unprofitable, toward AI compute services. This structural shift requires significant capital expenditure.
- 34:04 – 37:14
NVIDIA as the 'Central Bank of AI'
Speaker 0 characterizes NVIDIA as the 'central bank of AI,' running a quasi-fractional reserve system through unsecured lending and extended repayment schedules. This creates a shadow banking system analogous to the Global Financial Crisis (GFC).
- 37:14 – 42:08
Export Policy Tightening and Texas Politics
The discussion highlights a tightening cycle driven by US export policies on GPUs to China and political shifts in Texas, the third-largest 'sovereign' for AI data center build-out. Texas announced a 'go slow' policy on data center construction.
- 42:08 – 44:27
GPU Rental Price Indices as Misleading Metrics
Speaker 0 argues that GPU rental price indices are 'completely useless' for assessing market health due to Simpson's Paradox. They average disparate spot prices for marginal buyers with long-term hyperscaler contracts, masking true market dynamics.
- 44:27 – 51:14
Engineered Scarcity and Margin Protection
NVIDIA's high gross margins are attributed to 'engineered scarcity,' a strategy used previously in gaming and crypto. This artificial constraint on supply allows NVIDIA to maintain premium pricing power.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| NVIDIANVDA | equity | mixed | And then I'll just throw one more post-the NVIDIA conference call, fornication. That's a risk factor, |
| Meta PlatformsMETA | equity | neutral | I think what Broadcom has done, something with Meta, I think, or a bunch of companies invested in Meta's equity round. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
The current AI investment cycle combines four pillars: technology, real estate, credit, and government policy, which historically drive economic booms but are usually not present simultaneously.
Unidentified speaker (speaker_0) ·0:00
How to verify: Review the speaker's framework presentation in the transcript; verify historical economic booms (railroads, GFC) against this four-pillar model using economic history sources. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Data centers are treated by investors as commercial real estate (CRE) yield instruments, analogous to apartment buildings.
Unidentified speaker (speaker_0) ·0:19
How to verify: Check investor presentations or prospectuses for AI data center REITs or infrastructure funds for CRE-style yield metrics. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
The US government is using EPA changes to permit broader sulfur dioxide emissions for behind-the-meter natural gas turbines to avoid grid interconnection.
Unidentified speaker (speaker_0) ·10:00
How to verify: Search US EPA regulatory filings and executive orders from 2026-2026 regarding sulfur dioxide emissions and natural gas turbine permits. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceThe combined nominal valuations of SpaceX, Anthropic, and OpenAI are roughly equivalent to the peak valuation of the NASDAQ at the dot-com bubble.
Unidentified speaker (speaker_0) ·13:37
How to verify: Sum the latest reported private market valuations of SpaceX, Anthropic, and OpenAI; compare to historical NASDAQ index valuation data from 2026. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
8% of all critical software vulnerabilities ever found were discovered in the first three weeks of October of the current year.
Unidentified speaker (speaker_0) ·24:09
How to verify: Consult CVE (Common Vulnerabilities and Exposures) databases and cybersecurity reports for the specific year mentioned to count critical vulnerabilities by month. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
50% of all critical software vulnerabilities in installed software were found since January of the current year.
Unidentified speaker (speaker_0) ·24:29
How to verify: Consult CVE databases and cybersecurity industry reports for the specific year to verify the cumulative count of critical vulnerabilities. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Venture-backed companies are the largest per-employee spenders on AI, with a ratio of approximately 99 to 1 compared to other firms.
Unidentified speaker (speaker_0) ·26:47
How to verify: Access Ramp's published AI spend data reports to verify per-employee AI spending by company funding status. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Meta accounted for approximately 18% of Anthropic's revenues a year prior to the interview.
Unidentified speaker (speaker_0) ·27:27
How to verify: Review Anthropic's S-1 filing or public financial disclosures for revenue breakdown by customer. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceAI tokens are a hyper-deflationary commodity with a 13-fold year-over-year price decline.
Unidentified speaker (speaker_0) ·21:26
How to verify: Analyze historical pricing data for AI compute tokens or API costs from major cloud providers over the last 12-24 months. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
The volume of tokens must double every 95 days just to stand still given the 13-fold price decline.
Unidentified speaker (speaker_0) ·21:46
How to verify: Perform the mathematical calculation: if price drops 13x, volume must increase 13x to maintain revenue. Calculate the doubling period for a 13x increase over a year. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceOracle is using prodigious amounts of debt to transition from its monolithic database business to AI, placing it on the edge of a ratings downgrade.
Unidentified speaker (speaker_0) ·32:49
How to verify: Check Oracle's latest SEC 10-K or 10-Q filings for debt levels and credit ratings from agencies like S&P or Moody's. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceNeo-clouds were previously in crypto mining and pivoted to AI because mining was no longer a good business.
Unidentified speaker (speaker_0) ·33:24
How to verify: Review industry reports on crypto mining profitability trends and neo-cloud business models (e.g., CoreWeave, Vast Data) from 2026-2026. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
NVIDIA engages in unsecured lending and extended repayment schedules with no cash exchange, creating a shadow banking system.
Unidentified speaker (speaker_0) ·34:04
How to verify: Analyze NVIDIA's SEC filings (10-K, 10-Q) for 'accounts receivable' and 'long-term receivables' disclosures and terms of customer financing. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceTexas is the third-largest 'sovereign' in AI data center build-out, behind the US and China.
Unidentified speaker (speaker_0) ·37:35
How to verify: Compare data center capacity announcements and construction permits in Texas vs. China vs. the rest of the US using industry reports (e.g., from JLL, CBRE, or specialized AI infrastructure trackers). For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Texas announced a 'go slow' policy on data center construction.
Unidentified speaker (speaker_0) ·37:54
How to verify: Search for news articles or official statements from the Governor-elect of Texas regarding data center moratoriums or permitting delays around November 2026. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Neo-clouds represent approximately 60% of token sales.
Unidentified speaker (speaker_0) ·39:53
How to verify: This is a specific claim requiring verification against industry reports on AI inference compute consumption by entity type (hyperscalers vs. neo-clouds vs. others). For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
NVIDIA attempted to backstop GPU prices with a $500 billion kitty involving Blackstone and Goldman.
Unidentified speaker (speaker_1) ·40:09
How to verify: Search for press releases or news reports from NVIDIA, Blackstone, or Goldman Sachs regarding a $500 billion GPU price support fund. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceNVIDIA's gross margins rose to 55-60% during the crypto boom due to engineered scarcity.
Unidentified speaker (speaker_0) ·45:06
How to verify: Check NVIDIA's historical quarterly earnings reports for gross margin percentages during the 2026-2022 crypto period. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Primary sourceAI capex is justified by targeting the Total Addressable Market (TAM) for global white-collar labor.
Unidentified speaker (speaker_0) ·46:48
How to verify: Review earnings calls and investor presentations from major AI companies (e.g., Microsoft, Google, Meta) for stated TAMs related to enterprise software and labor displacement. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
CDS on banks traded at 20-25 basis points in 2005-2007.
Unidentified speaker (speaker_1) ·48:56
How to verify: Consult historical CDS index data (e.g., CDX or iTraxx indices) from financial data providers like Markit or Bloomberg for the mid-2000s. For recent events, verify 2026-08-20 through 2026-11-20. Use another period only when the speaker explicitly names it.
Do your own research
- Verify private valuations of SpaceX, Anthropic, and OpenAI against historical NASDAQ peaks using recent funding round data and market comparables.
- Check Oracle's credit rating history and recent CDS spreads to assess the 'edge of downgrade' claim.
- Review Ramp's AI spend data to validate the claim that venture-backed companies are the largest per-employee spenders on AI.
- Analyze NVIDIA's financial filings for unsecured lending practices and extended repayment schedules to evaluate the 'shadow banking' analogy.
- Research the US export policy restrictions on GPUs to China and Texas's data center construction policies to assess supply constraints.
- Look for historical data on critical software vulnerabilities to verify the surge in security risks reported in October and January.
- Compare AI CapEx forecasts against potential labor market displacement scenarios to model the 'repricing' tail risk.
- Examine GPU rental price indices for Simpson's Paradox by separating spot market prices from long-term hyperscaler contracts.
Full transcriptexpand
0:00Unidentified speaker (speaker_0): As I think about the current moment and as I try to explain it to people, one of the things that makes it really unusual, in almost a rational bubble in financial in a financial economic sense, is that it combines each of those components, each of those pillars, not quite an equal way, but in a very comparable way. There's a real estate component to many people participating in the AI bubble, the AI moment, they think about it almost like a commercial real estate project, so a CRE. And so in that sense, they're they treat it as a US. yield instrument, an instrument on top of which they're trying to generate a yield. Data centers become like apartment buildings full of GPUs, these graphics processing units. Similarly, there's a terrific technology story, which is AI itself, but there's also a loose credit story, which I've all seen in lots of different forms already. But there's also an incredibly strong, and I even underestimated how strong it was, government component that the government sees this, especially in the United States, as an excess. existential battle for the future of capitalism, but also the future of AI. And they see themselves in particular in this life and death struggle with China over it.
1:00Unidentified speaker (speaker_1): Hello, this is Dean Kernutt, and welcome to the Alpha Exchange, where we explore topics in financial markets associated with managing risk, generating return, and the deployment of capital in the alternative investment industry. I'm joined today by Paul Kodrowski. He is the managing partner at SK Ventures, also someone doing a tremendous amount of independent research, also a fellow at MIT. Paul, it is a pleasure to welcome you to the podcast today.
1:33Unidentified speaker (UNKNOWN): Great,
1:33Unidentified speaker (speaker_0): thanks for having me. Glad I could do it.
1:35Unidentified speaker (speaker_1): One of the joys for me in hosting this podcast, and it's literally, this is the eighth-year anniversary. So November 18th of 2018 was the launch date. So you're on a special day for me. We're 275 episodes in. And the best part of it is just really learning. And I have definitely learned a lot in reading your work and following some of your podcast appearances on other shows. So thanks for making that contribution.
2:03Unidentified speaker (UNKNOWN): No,
2:03Unidentified speaker (speaker_0): no. I'm more than happy to. As my wife says, tell someone who cares. So I have to find other people who care.
2:09Unidentified speaker (speaker_2): Perfect.
2:10Unidentified speaker (speaker_1): So you got a PhD. You have described it as a study in complex systems and how complex systems can become fragile along the way, and the study of the consequences of that as well. So take us a little bit back to your background, and that is going. to lead us into a conversation around this incredibly intricate ecosystem called the AI chain. Tell us about your background.
2:36Unidentified speaker (speaker_0): Yeah. So I'm an engineer way back when I was actually a mechanical engineer of all things. And I got interested in systems as a result of mechanical engineering because mechanical engineering is really the engineering of systems, mechanical and otherwise. Increasingly, they were not just mechanical, but also computer. related and other things. So I spent a lot of time at the intersection of mechanical engineering and computer science. And so, as I say, that got me thinking in systems. And that in turn got me thinking about areas of systems where I was particularly interested in finance, was one of them. So I did an MBA in finance way back when, and uh, spent some time working on the cell side for HSBC HSPC securities as an equity analyst. And then I at coincidentally coincident with that, I decided I wanted to do a PhD. in the whole area of complexity, but more specifically in this idea of contagion and how this was long before the days of memes and other things, but the notion of how a combination of contagion and path dependency can lead to optimal and suboptimal outcomes, and in particular leads to different kinds of unexpected fragility. And so that was really the origins of it was coming from an engineer. engineering standpoint, thinking about systems, then in turn thinking from systems to fragility, and then in terms of what some of the predictors of fragility are. And along the way, Matt, all kinds of interesting people. I spent a ton of time with Nasim Talla back in the day, reading early galleys, some of Nasim's work and others who were doing
3:55work in this area, Bernstein and many others, but just the whole idea of complex systems, the predictors of fragility. and what, if anything, that tells us in terms of tail risk in the context of financial markets. So that in turn led to a lot of my investing activities. So, in large part, I'm in the business of buying long dated call options, which other people call early stage venture, which is really what it is. You're buying long dated call options in a private and In highly illiquid marketplace. The most you can lose is whatever you paid for them. So in that sense, you're buying the option. And obviously you're hoping to get many multiples of that if something works out. So that's one of my sort of day jobs. As well as doing that, I spent some time at UC San Diego, University of British Columbia, and now at MIT and to a lesser extent at Oxford. Also lecturing and researching on the this topic, in which is to say, in complex financial systems, what can we know in terms of the predictors of risk? So that led me into the dot-com dot com episode. the global financial crisis, where I spent a lot of time speaking, thinking, writing about what was going on there. So my interest in AI to bring us to the present really comes through that, that history and that path. And in particular, about how it's very much like a lot of these prior episodes in terms of the build-outs build outs of prior technology waves. But the separation I make from other people who talk about this is I'm not particularly
5:15interested in it just from a technology standpoint. I can be as geeky as the next guy, but I think in systems. And so I think in terms of systems, fragility, tail outcomes, and those are the things that interest me. It doesn't really matter whether it's mortgage bonds, credit default swaps, CDOs, AI CapEx, any of those things. So for me, these are all part and parcel of the same package of things that lead to an and a I'll say interesting outcomes, but in particular for me that lead to fragility. So if that's kind of the path that's gotten to me where I am is thinking in systems and long dated co-option terms.
5:48Unidentified speaker (speaker_1): Yeah, my own career is almost entirely in the derivatives market. And I would say if I've studied one thing, it's the time variation in the price of vol All and in insurance. You said on other podcasts that your thought process tends to live in the tails of the distribution. Very similar for me, obviously, Taleb, famous for that as well. And trying to think through whether the tails, because that's where all the leverage is, are just wickedly underpriced or overpriced. And as you mentioned, the GFC, boy, what an incredible underpricing of the tail for several years running. And that's really what I want to try to explore here is this idea, as you say, of fragility of these complex systems. You've talked about today's CapEx boom. It's a gigantic debt fueled trade at this point. It's a concentrated bet. And you've got this framework, which I would love for you to just open up the conversation with. You've talked about it on kind of four pillars: technology, real estate, credit. and government. And why don't you walk through how you arrived at that and then how today's trade is just unique in checking all the boxes?
7:01Unidentified speaker (speaker_0): Sure. So I'm a research guy, in at least in part. So one of the ways I like to think about the world is through frameworks. It's a way of describing it to myself, but describing it to other people as well. So at least at least I can understand it well enough to explain it to other people. So one of the frameworks that I approach this stuff through is the idea that in Western economic history, most of the most exuberant moments in terms of Western economic history, back to the South Sea Bowel, have had as predicates at least one of four things. Usually there's a great technology story. Often there is a great technology story. Sometimes there's loose credit. There's a credit story associated with it, which can be come from a lot of different perspectives and angles. There's often a real estate component as an asset. It's been particularly ficile with respect to being a precursor to these explosive events. And then there's often a government component where, not just in terms of government omissions, where there's some kind of policy failure that led to the crisis, but actually acts of commission, where the government's actually urging people to do things and in a sense creating what we used to call back in the 90s, like a greenspan put underneath the market, where it was this idea that I have your back. And so by my actions, I'm creating an incentive for you to take actions. It's not so much that you're running some kind of arbitrage and as a government, I'm missing something. It's that I'm actively encouraging you to do something. And the way I try to describe some of the biggest moments in Western economic history, back to the South Sea bubble and the railroads and canals in the 19th century, both in the
8:20US and in Europe, and then rural electrification in the 20s, twenties, or fiber in the later part of the 20th twentieth century, or the GFC. All of these had to a greater or lesser extent one or more of these components associated with it. And you can think specifically about the GFC as being kind of at the intersection of real estate government policy and credit. They were sitting at that component. It wasn't a strong technology story, even if there were aspects of it. Similarly, you can think about fiber in those terms as sitting at the intersection of technology and credit, not so much of a real estate story. And you can go back to the railroads. which really had a little bit of everything. It was a technology, there was a real estate story, there was loose credit, and there was a government component that was urging it on because of this idea of the expansion of the westward front, with the westward expansion of the frontier in the US. So this is all a long way of saying, as I think about the current moment and as I try to explain it to people, one of the things that makes it really unusual, an and almost irrational a rational bubble in financial financial in a financial economic sense, is that it combines each of those components, each of those pillars, not quite an equal way, but in a very comparable way. There's a real estate component to many people participating in the AI bubble, the AI moment, they think about it almost like a commercial real estate project, so a C R E. And so in that sense, they're they treat it as a yield instrument, an instrument on top of which they're trying to generate a yield. Data centers become like apartment buildings full of GPUs, these graphics processing units.
9:40Similarly, there's a terrific technology story, which is AI itself, but there's also a loose credit story, which I've all seen in lots of different forms already. But there's also an incredibly strong, and I even underestimated how strong it was, government component. The government sees this, especially in the United States, as an existential battle for the future of capitalism, but also the future of AI. And they see themselves in particular in this life and death. struggle with China over it. And if you look back at executive orders over the last two and a half years, the bevy of them, the number of them that are directly associated with trying to accelerate the build-out of AI, everything from EPA, EPA changes that permit even broader sulfur dioxide emissions for behind-the-meter power, natural gas turbines, so that you can avoid grid interconnection. And all of these sorts of things are cumulative and at create this incredible tailwind with respect to driving the moment. So what's unusual about this moment and what a lot of people miss is that it combines these four forces that historically have been the drivers of some of the largest such moments in Western economic history. And they combine it in a really unusual way, such that the participants in each of those silos say, Hey, I don't know what those other guys are doing. But from my standpoint, what I'm doing as a real estate investor. doing powered land or as a technology investor looking at the consequentiality of GPUs, they all see what they're doing as rational. But of course, it's the intersection and
10:59confluence of all of their discrete activities in this system of the economy that creates the fragility that's the thing that I end up on.
11:07Unidentified speaker (speaker_1): When I look back on let's just To say the precursor to the GFC and then also the, let's say the internet bubble. I see some similarities and then some differences. The internet bubble being just this almost absurd valuation bubble, just people paying ridiculous premiums for expected earnings. In the GFC, you've just got so much leverage in the wrong hands, right? It's on the balance sheet of Merrill Lynch at 35 to 1. one. It's just too much.
11:35Unidentified speaker (speaker_0): And it stayed there, obviously. This is the thing, is that unlike today where a lot of the debt is moving around, it was all they were still able to wear house a significant amount of the debt which created the problem that you're describing.
11:46Unidentified speaker (speaker_1): And then also just a lot of interconnectivity in the GFC. And that's definitely going to be a part of our conversation. When you look at those two episodes, and of course the GFC did a lot more damage. The internet bubble was a crazy crash, but it didn't really the economy. had a mild recession. You could say that some part of 9-11 might have been in there, but the obviously the GFC crash was catastrophic for the economy. What do you see if you look back on those two as analogs? What's similar, what's different from your standpoint today?
12:18Unidentified speaker (speaker_0): Yeah. So the dot-com episode was largely an equity bubble. And so in that sense, the consequences were felt most directly in equity markets and in other than some diminished spending. power and which may have led to some recep to a recession as you describe it, it wasn't as consequential in terms of its broader impact in the financial system because it was compartmentalized in broad terms inside of equity markets, even if it did lead to a recession and lower interest rates, which may have had a role in terms of ZERP and other things subsequently. So putting that aside for a second, the GFC, on the other hand, was obviously there was an equity component to it, it was much. more of a credit, much more of a credit problem, in particular in the sense that the entire financial system obviously became captives of this shadow banking system that was creating this vast expansion in the money supply and often leaving it levered and in the wrong hands. But also, even more insidiously creating the impression that it was safer than it was because of this insurance component with respect to what was happening to get the kinds of credit ratings of. that the bond rate agencies gave. And there was this fundamental arbitrage going on. So there's elements of both of this inside the current moment. There's obviously an incredible equity component to it, which we're seeing play out this week in the context of the anthropic IPO. But this has been going on for some time. The scale of this is unprecedented in nominal terms. SpaceX, Anthropic, and OpenAI, in nominal terms, the
13:37valuations in the private markets of those. three firms are roughly equivalent to the peak valuation of the NASDAQ at the peak of the dot com bubble. So to give you a sense, people often say, well, if this is a financial bubble, where are the dot-com stocks in this? Where are the frothy valuations? And of course, we've seen over the last 25 years the declining number of public equities on exchanges. What's happening increasingly is these things are now in the private markets. So the froth, if you will, and these valuations are increasingly in private markets. And we can see them in the incredible valuations being accorded to these three companies in particular. The way I look at this more broadly in terms of the systemic effects is very much like some of Richard Ku's work about balance sheet recessions with respect to what happened in China after the unwinding of the real estate boom of the 80s and 90s. And so this for me, in terms of as you think about the consequences of this, is more analogous to that than it is to specifically the GFC or to the dot-com episode, in the sense that we have all these private companies levering up dramatically, in part because they can, because they had such pristine balance sheets. The Piperscalers pre-AI CapEx
14:43had such pristine balance sheets. That what we will see on the other side of this is a prolonged deleveraging that will be largely insensitive. to interest rate declines because they are deleveraging, as Koo pointed out when he was analyzing the Japan event. It won't matter that rates continue to drop because they simply don't want to take on more debt. They're trying to deleverage. And that's will be the hallmark of this episode was the prolonged deleveraging that will be largely insensitive to lower rates. So I pulled pull together pieces of all of these moments. The Japan episode, which to many people will seem completely unrelated, but is very much related, has some real comparable aspects, then aspects of both the dot-com and the GFC.
15:21Unidentified speaker (speaker_1): Maybe just stepping back, you are a technologist, you are a big user of AI. And I'm going to guess someone who marvels at just the achievement of what this thing is and what its potential is. But you're looking at the playbook of past giant booms. You talk about railroads and electrification and capex booms just typically lead to busts.
15:45Unidentified speaker (UNKNOWN): Yeah.
15:45Unidentified speaker (speaker_0): And I think that's a really important point. And it's one that people often miss. They have this very narrow view of the system where they say, You're either with me or you're not, in the sense that if you believe that AI is this really important and consequential technology, then why are you saying the things you're saying? Or vice versa. And I'll get this from some of the more bearish financial types: that if you feel this way about the underpinnings in terms of the finance, how can you possibly believe that the technology works? And of course, this is a false dichotomy, right? This is a fundamentally unhelpful way of thinking about the moment, no different than saying in the 20s that we're seeing these incredible build-outs of these pyramided structures that led to all of this leverage into the crash of 29 that were predicated on rural electrification and autos. But it doesn't mean that either electrification or autos were not technologies that were incredibly consequential and important. And similarly, in the late 19th century, with the rolling series of financial crises during the railroad build-out and the collapse of the Erie Railroad, like, I don't know, five times over, this didn't mean that railroads were not an important technology. So you have to be somewhat, I think, sophisticated as is giving people too much credit. Just someone mature about this and say, we can have both. And in a sense, having a strong technology story is the ante to the game. If the technology wasn't so important and so dramatic, you wouldn't be able to attract this much capital expenditures to it. This is just the ante.
17:05So the idea that somehow, and I see some people out there making this argument that AI doesn't work or that this is all hyperbole, and these are nonsensical arguments. It's probably a among the most important technologies of the last hundred years from an IT standpoint. I'll grant you air conditioning. Some of the areas inside of medicine have been more important writ large, but inside of IT, it isn't among the most important things we've done. And so, but that's what's driving the moment. So you have to embrace both sides of it. And if you look through history, and this is the one of the ways I got interested in, this particular episode was the parallel. parallels to other AI CapEx build-outs in support of canals, railroads, electricity, even white-collar office automation, the fiber optic moment. They're so analogous in terms of this growing share of nonresidential fixed investment, the increasing contribution to GDP growth, even the leakage into other markets, like we saw in late the summer, were increasingly at the same duration fundraising around AI. was increasingly causing at the margin pressure
18:05in treasury markets, which is remarkable, this idea, right? But this is the kind of leakage you see as in these moments where CapEx becomes such a circular and gigantic phenomenon. And so that's the part of it that's really, I think, interesting. So those are the parallels through history is that the technology works, but the damage with respect to the investment is calamitous. And why? And it's just a question. of who pays the price.
18:28Unidentified speaker (speaker_1): And it seems, tell me if you agree with this, it's almost a feature, not a bug. It's almost that you can't really get around it. When you have something this path breaking, you're gonna overbuild. Zuckerberg himself has said the risk of moving too slow is just greater than the risk of moving too fast. We'll burn some capital if we have to. We don't want to, but to be left behind is unacceptable.
18:51Unidentified speaker (speaker_0): And I think that's true with one proviso, and I think it's really important. So there's one of the people that Mark and many others like to point to who's written extensively on this is Carlotta Perez, her book Technology Revolutions and Financial Capital, which is about exactly what you're describing, this notion that it's inevitable that we'll have these kinds of overbuild cycles. And at the other end of the overbuild cycle, there will be some assets that many of the assets will become stranded or even potentially become useless. But what we've constructed will still inevitably drive what she called a sort of a future golden age, which is a bit too much religiosity for me. But nevertheless, the idea that they'll be valuable.
19:26Unidentified speaker (UNKNOWN): And
19:26Unidentified speaker (speaker_0): I think that's largely true. But I think it there's a really important distinction to make that's different about this episode was that in prior episodes, we didn't tell ourselves that it's all going to work out because people like Carlotta Perez says it will. And so this notion of justifying overexpenditure by saying it always works out has been growing over the last two or three episodes. But in this one has become structurally fundamental to it that everyone alibies their own activity by saying that it always works out when we do this. And of course, that's not what happened in history. People weren't telling themselves when they built the railroads that overbuilding railroads always works out. right? So you create an extra layer of a feedback mechanism that accelerates both the speed at which it happens and the inevitable failures and fragility, because you created this self-justification based on history, but you're using it as a justification ex post, wherein before it was a description, right? And so this is, I think, really important about this moment. So he's not wrong, but using it as justification changes. the dynamic of the system dramatically.
20:29Unidentified speaker (speaker_1): In markets and maybe in poker as well, there's this Kelly criterion. It's your bet size, right? And clearly situational awareness didn't read that manual. They were just way too big relative to the vol of the assets and the leverage they were deploying. And I've had this theory around the system at large is just wrong sized from a Kelly perspective. while no one's individually making that decision. And I think maybe this is what you're saying is that the system is so big, its demands are so insatiable that it's just going to be accident prone. Is that a good rendering of kind of some of your thinking?
21:06Unidentified speaker (UNKNOWN): Yeah,
21:06Unidentified speaker (speaker_0): I think that's part of it. And so I'll give you two aspects of exactly what you describe. So one of them is that it's so large that it's fundamentally going. to trip over itself for exactly that this reason you describe. And part of it's because the nature of the underlying commodity is so unusual. So I like to describe tokens, which are sort of the atomic unit of AI, if you will. Think of them as the first hyperdeflationary industrial commodity, that it's structurally a deflationary commodity that has It been deflating and will continue to, by some measures, about 13-fold 13 fold year over year. We can do the math on that. the consequences of that. That implies that roughly every 95 days, the volume of tokens must double just to stand still. So just to stand still, I need to double the volume of tokens every 95 days. Albert Bartlett has this great line Well, one of the greatest failings of the human species is it's an inability to understand the exponential. This is that that inside of this system, when you start seeing these kinds of exponentials playing out, we struggle to understand. how to extrapolate from that to the needs of the system. And
22:06as a result, the implications for overbuild become even more dramatic, which makes the system more fragile. Because all we see everywhere we look are growth exponentials and price collapse exponentials. So if you want to play the game of speculating that collapsing prices will induce more demand, there is your evidence. You can play it right here. Look at this 13-fold year-over-year decline. If you want to play the game that we're never going to sell enough, we're never going to catch up. We have an exponential over here to describe exactly that phenomenon. And so this is the problem that leads to this greater fragility is it's a series of embedded and interlocking exponentials in terms of demand, in terms of collapsing price. And underneath it all is this kind of naive assumption that people's marginal propensity to consume, the elasticity, is infinite. That it doesn't matter how much token prices collapse, we'll always find more uses. And we know that's a bounded relationship. It's not true through all things. I do not want a hundred cars if the price of cars drops to zero. There's a regime in which this applies, and there's the industry is busy extrapolating out wildly based on the last three or four years of what's happened to say that that marginal propensity to consume tokens is infinite.
23:14Unidentified speaker (UNKNOWN): And
23:14Unidentified speaker (speaker_0): as a result, we should build accordingly because look at the underlying geometric sequences, which in turn creates this fragility. And the second source of fragility in that context is along the way, there's a lot of unintended consequences. So models are becoming much better
23:29Unidentified speaker (UNKNOWN): in
23:29Unidentified speaker (speaker_0): some narrow domains than we can keep up. And the canonical example is what we've seen over the last week with respect to what I've been calling kind of the hacking, this idea that. the surface area of systems worldwide built around information technology that can be attacked, the surface area that can be attacked is expanding much faster than we can defend it. So as a result, we're in this precarious balance where it feels like at any given moment it could be airlines, it could be banks, it could be natural gas supply, it could be something else. This isn't entirely predictable because the ability of these systems, as they grow more competent. expands rapidly in terms of being able to attack other things. And just to give you some data on that, I mean, and it's just staggering. Like, for example, about 8% of all critical software vulnerabilities ever found in the history of the industry back to the 1960s were found in the first three weeks of October of this year. 8% were found in the first three weeks. 50% of all the critical software vulnerabilities. in installed software were found in the first since January of this year. So half of the vulnerabilities ever found. And so we're playing this incredible catch-up game against a nemesis, against a foe that can move much faster than we can. And of course, when you're defending against these things, it's no different than defending against in guerrilla warfare. They only have to be lucky once. You have to be lucky every time. And this is the other source of
24:49fragility. We will get to the other source. side of this, but in the interim, the system becomes vastly more fragile.
24:55Unidentified speaker (speaker_1): Let's talk about how the various major players along the value chain interact. So I'll give you one of my little frameworks. So I talk about market risks in terms of valuation, duration, inflation, concentration, correlation. So those are my five. And then I had to add two. Polarization, which is really unfortunate in the US, but it definitely speaks to data center resistance and so forth, especially post the big democratic blue tsunami. And then I'll just throw one more post-the NVIDIA conference call, fornication. That's a risk factor,
25:31Unidentified speaker (speaker_2): too.
25:31Unidentified speaker (speaker_3): That's a funny way.
25:33Unidentified speaker (speaker_1): I had to throw that in there.
25:34Unidentified speaker (UNKNOWN): I
25:34Unidentified speaker (speaker_0): did not have that on my bingo card.
25:36Unidentified speaker (UNKNOWN): Yeah,
25:36Unidentified speaker (speaker_2): exactly.
25:37Unidentified speaker (speaker_1): So concentration, though, and correlation. I think, are really worthy of focus. And it's definitely been something you've talked a lot about. There's always going to be in the let's say financial services industry, there are trading counterparties to one another, right? There's no way around it. So if you're in the same business, you're gonna be suppliers, customers, you're gonna have To financing arrangements. But this definitely does feel pretty overlapping in a lot of ways.
26:03Unidentified speaker (UNKNOWN): Yeah,
26:03Unidentified speaker (speaker_0): the concentration is really dramatic.
26:05Unidentified speaker (speaker_1): So one of the things, let's start with the concentration part of it, because you've talked about just the way in which the buyers of AI type services are vastly funded by VC, much more so than any other type of firm. Maybe just give us a little bit of data and some of your framework and then just how you think about it from a risk standpoint, that kind of concentration.
26:27Unidentified speaker (speaker_0): Yeah. So I'll start with the data and then go upwards in terms of why it's like this. So there's data out there and there's lots of quibbling about how credible it is or not. But one of the providers of data in terms of usage is a company called Ramp. Ramp's data, which tracks organizational spending on AI on a per company basis anonymized, but with some characteristics, shows us that the largest spenders on a per employee basis by far are venture back companies. So venture back companies, which are obviously funded by venture capitalists, are the largest spenders, like to the point of something like 99 to 1%. Like there's no comparison. There's vastly more spending on a per person basis. Now, this is not aggregates, it's just per person basis. So it's showing you where the spending is coming from, it's largely coming from these organizations who are increasing, who are spending for the purposes of who are venture backed and unprofitable. So that's one slice of this. The other slice is that the industry, writ large, that's the biggest spender on this. And we saw this in the early leaks from the Anthropic S1, are the software industry, where at one point Meta, for example, was something like 18% of Anthropic's revenues, which It's changed, but that was only a year ago. It's still a very material fraction. So there's huge concentration in two different ways. Huge dependency on fragile young companies, huge dependence on a small number of technology companies and a small
27:47group of employees, relatively speaking, within those companies. So just one group, software engineers in particular. And so that kind of concentration is really unusual to see, both in terms of the sectoral dependence and the employee group dependence, as well as the dependence on a single sector. And the reason why that's important isn't just because it creates fragility. It's also because, and this is an old tenet of venture capital, is that never extrapolate to your eventual market from your earliest buyers, which is to say that your early adopters almost certainly look nothing like the long term users, what they want, the risks they're willing to take, and so on.
28:24Unidentified speaker (UNKNOWN): So
28:24Unidentified speaker (speaker_0): as you transition from early adopters to the mass market. do not make the mistake of extrapolating the needs, interests, and propensity to spend of the early buyers. That's in a sense what has happened right here, because the early adopters of large language models, the foundational technology being used by current generation AI and frontier AI companies, was initially most appealing to software engineers for a host of different reasons, which we don't need to get into. But why that's really important is that there could hardly have been a field more amenable to the use of these tools than software. And the reason is because it's highly rules-based. There's a sense in which I can work within a language-based model to process the rules and then produce tokens that represent software. But it's even more fundamental than that. There's a notion in machine learning of this idea of what's called gradient descent, which is to say that as I do things, and I take a signal back, as my model does something and it takes a signal back from the outcome, I'm trying to minimize the variance. I'm trying to find the strongest signal by minimizing the variance in what happens that's descending a gradient of outcomes. Software provides a very strong signal because there's no notion of aesthetics or of whether or not this is prettier than that. It's
29:35the software works or it doesn't. It runs faster or it doesn't. So you can descend a very tight gradient, which is one of the reasons why software is a profligate user of tokens, creating software, because I can let it run semi-autonomously, because there's a ground truth. There's a signal that it ingests and says, okay, this is working, this isn't. Now turn this to the average white-collar worker as you think about the future of what the next group of customers look like. Can I let a model loose on a memo and have it get better over the next 12 hours? No, but I'll it'll get a uniform. uniformly worse. It doesn't even necessarily take a few hours. It could happen in two iterations. Same thing with a PowerPoint, same thing with writing a memo or an email. So the world of work outside of software lacks that same gradient descent that exists inside of software, which is why extrapolating from those people who were monolithic and very aggressive early adopters doesn't tell us particularly much about what the next cohorts of buyers will look like. because of this extrapolation problem, which goes back to this idea of these monolithic early users and how concentrated they were, not just sectorally, but also just occupationally, and why that's such a source of risk and how unusual that is.
30:41Unidentified speaker (speaker_1): So you're saying that perhaps there's just a overestimation of the TAM. That's the
30:47Unidentified speaker (speaker_0): Yeah. So what you end up with is exactly that. There is a huge overestimation of the TAM. if for many reasons, but that is, I think, one of the most important ones is this naive extrapolation from an unusual cohort of early users who had some unique needs, but also whose domain was uniquely amenable to the things that large language models are best at and that turn out to be not particularly representative of most of the rest of white colour work. Doesn't mean it can't work in those areas, but it does mean the way they work with it and this autonomous usage of these models that generate huge amounts of tokens can't happen. Definitionally can't happen because the models have lack of ground truth. I don't know whether my boss will like this memo more this way or that way until my boss tells me.
31:29Unidentified speaker (speaker_1): There's all kinds of overlap. There's Oracle with OpenAI. And AI. That's a strong tie-up. Of course, NVIDIA's got investments in Core Coreweave. Weave. I think what Broadcom has done, something with Meta, I think, or a bunch of companies invested in Meta's equity round. Which of those or others matter to you most? Like where, from a fragility standpoint, is the market vulnerable to some linkage getting disrupted and that kind of propagating? We You used the word contagion before. I don't think the banks are going down. but just in terms of it propagating outward, where are the things that you're looking at that could be the starting point? So
32:09Unidentified speaker (speaker_0): I look at it in terms of the sources of fragility and the destinations of fragility. So I'll start at the destination because that's the most maybe probably the most obvious one. But you look at the companies that are in the most exposed position and trying to protect themselves by escaping from their exposed position by embracing this as quickly as possible. And the canonical example. is obvious there are a couple of classes of companies, but one of the most obvious ones is Oracle. Oracle's in this very precarious position as it tries to transition away from this its monolithic database business. And it sees this as a path to escape from its dependency on what is really a mature and even declining database business. And so it has this problem of trying to balance the transition and the classic problem of trying to move from one sector to another while not having revenues decline so quickly that you get punished for it. And so they're in that transition, but it's heavily debt assisted. They're using prodigious amounts of debt to the point that there's petering on the edge of yet another ratings downgrade and their credit default swaps trade accordingly as do the spreads
33:04Unidentified speaker (UNKNOWN): on
33:04Unidentified speaker (speaker_0): much of their debt. So they are the destination of fragility, right? So there's a place where fragility is being sent because you recognize that it as a company, and there are many of these companies. Some of the neo clouds are, I would put in this camp as well, who are recipients of fragility, if you will, because they're trying to solve a problem. The neo cloud problem they were trying to solve, which is less less well understood than it should be, was that many of the neo clouds were in crypto. And so the problem they had was that mining crypto was no longer a very good business. So they began hunting around for other things that I can use massive amounts of compute for. So the problem that they're solving is getting away from a dependence on crypto. to something else that can use your large GPU farms in the same way. Well, welcome to AI, where inference and training are prodigious consumers of compute. And so this looked like a great place to be, but it required vast amounts of debt to build it out such that you could be useful for doing it. And so again, a source of fragility because they're trying to solve a problem, but they have to use debt to do it. So that's at one end of the continuum in terms of thinking about this. And then you have the dispenser. dispensers of fragility. And so, you know, I often call NVIDIA the central bank of AI, that they're running a quasi fractional reserve system where they're vastly expanding the scope and consequence of their balance sheet in a host of different ways, including unsecured loan, unsecured lending, including repayment schedules that stretch out well into the future and require no cash exchange. It's
34:24creating currency much like a central bank or a central bank might in a fraction does in a fractional reserve system. And that currency becomes its own kind of shadow banking system, very ways, and analogous to what happened as risk accelerated during the GFC. And so that is the sort of a source of fragility in the system. And we're seeing some of that just in terms of the pressure that NVIDIA was under to continue to grow revenues, which led it to do things that Jensen Huang thought was in the company's best interest. interests, And he argued. were in the nation's best interests, that involved really stretching the rules with respect to export policy and transshipment of GPUs that were servicing China, whether through cloud services in Southeast Asia or through the black market and chips that ended up in Asia. But nevertheless, these are all are best thought of as consequences of what's happening in terms of the pressures being building up and being felt in the system rather than being a one-off. or anything else. So that's the way I look at it is these two poles in the system in terms of these sort of sources of fragility and the consequences of fragility with hyperscalers and others sitting in the middle and in a perverse way, almost the frontier AI companies in the middle of this, where they're not really the dispenser or the source of the they're not really the dispenser or the recipient of the fragility. They're just in the middle. And I don't actually hold a lot of hope for the frontier A AI kind of companies over time. I think increasingly they're in this awkward position of neither being true cloud service providers nor able to stand on
35:43their own without cloud service providers. So it puts them in an awkward spot over time.
35:47Unidentified speaker (speaker_1): You talked about the anthropic. So it's we're recording this the day after NVIDIA, but the day before Anthropic.
35:54Unidentified speaker (UNKNOWN): So
35:54Unidentified speaker (speaker_2): it's
35:55Unidentified speaker (speaker_1): the day after the day before. You called Anthropic a bank with a side hustle in AI. They'll have 200 billion in the bank, which is a fair amount.
36:03Unidentified speaker (speaker_0): Which is remarkable for a tech company.
36:06Unidentified speaker (speaker_1): Yeah, exactly. When I look at anthropic, and then I look at OpenAI's commitments, seem almost just unmeetable. And that seems to me to be some risk for the system. But I want to go back to the central bank of AI. It's just such a great moniker. I was covering long-term capital in the mid-90s. I was at Lehman Brothers. They called LCCM the central bank a vol.
36:28Unidentified speaker (speaker_0): Oh, nice. I like it.
36:29Unidentified speaker (UNKNOWN): Yeah,
36:30Unidentified speaker (speaker_1): Morgan Stanley came up with that. And the idea was that long-term was there to govern spreads, that if a spread got out of line, it was there to supply risk-bearing capital, whether it was in vol, swap spreads, mortgages, etc. And it created a little bit of a circularity. It was really, it greased the wheels of the system and it kept it going. That just really uh is a good framework, I think, to think about this central bank of AI. Would you say that this central bank of AI has truly moved to tightening mode? Are the midterms and Jensen Wang being moved out of his role? Are those really significant events with implications going forward, or we just don't know yet? How do you think know? about
37:14Unidentified speaker (UNKNOWN): it? They're
37:14Unidentified speaker (speaker_0): very significant. And it was remarkable to me that in the wake of the midterms, that the initial market reaction was much more positive than it was. This was a tightening moment. There was no question about it. One of the ways I characterize it is that if you think about this global AI data center build out, when we separate it into sovereigns in terms of the construction, one of the ways I characterize it is that the largest sovereign in the world in terms of AI data center build out is the US, not counting Texas. The second largest in the world is China. The third largest in the world is Texas. So if Texas were a sovereign, which I always joke, would probably I could just end that sentence right there, it would be the third largest sovereign in the world. And yet we have a governor elect, and this is not making any political position one way or the other. It's just recognizing that the third largest sovereign in the world just announced a go slow policy with respect to what's happening in data centers. Now, there are things happening this week and new. next likely that will lock in some of the construction there. But nevertheless, it's still important in terms of, as you describe it, this sort of implicit de facto tightening cycle that's underway with respect to what's happening. And it matters because of the size of Texas as a de facto sovereign with respect to the spending on data centers. And the market initially got that wrong and then rethought it the next day all of a sudden they all got sold off. The Bulls made strange arguments And one of the arguments was that building more data centers is actually good for us because it means that you can better protect margins in data
38:34centers because there will be fewer of them built. And of course, that implies something that isn't true, which is that all the data centers will be built in Texas or in the United States at all. And that's simply not true. I mean, it's obviously this a massive construction project underway with hundreds of data centers in various stages of construction in China alone. And so the notion that somehow in an inference based world where I'm simply generating the next token, not just betraining new models, that I can constrain it to geographic proximity is really naive. It's a bizarre argument. And as I said, it was if both having more tokens available and having less tokens available for you is your argument, then you no longer have an empirical argument. You have a theological argument. This is a faith faith-based based argument, right? It's no longer based on anything. other than what you want to have happen. So, yes, in answer to your question, it was a tightening cycle. And then NVIDIA on top of that creates another tightening cycle because with the transition to at least an interim CEO, if not a new full-time CEO yet. And you heard this in the earnings call this week, there was a question that came up of what their views were with respect to the prior administration's policy with respect to fairly aggressive lending to neo clouds. And she said we're hugely and we're huge supporters of neo clouds. And in the very next breath, the response was something to the effect that, but we're going to be very careful in future to make sure that they're the right recipients of what we plan to do. So ignore the first part of that statement. The reality is that the tightening cycle has begun with respect
39:53to the extension of credit to neo clouds, who represent among the largest consumers, something on the order of 60% of token sales, which in turn turn. turns into compute sales in future, were tied to NeoClouds, and NeoClouds just found out that they're they just entered a tightening cycle.
40:09Unidentified speaker (speaker_1): And the cost of that credit is definitely high. You certainly see it in Core Weave CDS. There was an announcement over the summer, which is really a central bank of AI from NVIDIA, which was Jensen Wang saying, we'll backstop GPU prices. I think there was a $500 billion kitty with black. Blackstone and Goldman. I think it was Blackstone and Goldman, might have been KKR, but you just gotta wonder if that's still in the offing. I mean, I don't even know if it was if it's in the price of anything at all. It was conjured up and didn't really materialize.
40:41Unidentified speaker (UNKNOWN): No, it
40:41Unidentified speaker (speaker_0): didn't, but it was a good example of an attempted going back again to the central bank of AI phenomenon. This was an attempt at a currency peg. One of the ways to think about it was an attempted currency peg. that the market realized fairly quickly wasn't sufficiently backed. There was inevitably going to be a run on that. We're going to see a moment where you break the peg, but it was an attempt to create a peg given the role of GPUs as currency in this marketplace. Again, this is why it helps to think about in systems terms and think about who's the central bank, what are they trying to do? What's the currency? What are the underlying commodities? And as soon as you start thinking in those terms, all of this begins to feel very familiar and you see. the points of stress and the interlocking pieces. And so you know, I think that's gone away, but the intent was what he was trying to do was really instructive because you realize immediately that GPUs are a commodity in this marketplace. People were unhappy with a floating rate exchange. He felt he needed to peg the currency, and here's how I'm going to do it. And so once you see it in those terms, you realize what the dynamic is and where the inevitable breakage is.
41:37Unidentified speaker (UNKNOWN): So
41:37Unidentified speaker (speaker_1): there seems like there's a couple of different vulnerabilities. One is you've outlined it's just declining token prices, which are just a function of the incredible technology itself, right? It's almost the technology turning on itself. And then I'm just wondering if the other side is also a risk, which would just be some declining prices of GPUs, GPU rental rates. And that being, you know. This is
42:02Unidentified speaker (speaker_0): one of my bugs. bugbears, by the way. And I'll let you continue your point, but I'm just preparing you to hear a rant from me. So
42:07Unidentified speaker (speaker_1): Go ahead.
42:08Unidentified speaker (speaker_0): Just one of the things people tried out, and again, if once you start thinking about these things, what are the key data series and proxies for the outcomes that you're most concerned about? You know, if it was the GFC, you're looking at things like CDS and what have you. One of the things people like to look at in this and act as if it's predictive is GPU rental prices. There are various indices out there. Bloomberg has some. Silicon Labs has others. There's a host of them out there. And I think these are, I won't say next to completely useless, but you can see completely useless from there. So they're adjacent to that problem. And the reason, of course, is that there's at least two things going on. There's incredible variance. These are not true indices in the sense you're averaging across vastly different providers of tokens and vastly different providers of these GPUs themselves. So there's incredible dispersion. And so until you understand the dispersion, you run the risk of averaging across two clusters, right? So this is this Simpsons paradox in statistical terms. So you actually think you think the relationship goes this way, but there's actually two clusters. And if you did it properly, it actually looks like this, right? This is the idea that there's really so what we have in the GPU marketplace is you have people who show up who are essentially buying commodity or buying this commodity called. tokens at the margin. And when you buy it at the margin, what you're buying is compute at the margin. And so when you buy compute at the margin, it's no different than buying anything at the margin, whether it's a Las Vegas hotel room during a big conference or any commodity under
43:27supply constraints with temporarily tight demand. You pay a very high price. But it's not reflective of what's happening in the broader market in terms of what the hyperscalers and others are paying in terms of their long-term commitments. The prices are a reflection of this sort of aggregated index across two functionally different aspects of the market. One, this kind of spot price, if you will, almost in terms of what's happening with the provision of tokens and compute to people who are sort of latecomers and across all of these services, and then also the sort of the frontier models themselves. And so the idea, Jensen has said this, Jensen Hawing Hawaiian has said this recently and others that oh, look, everything. must be good because GPU rental prices continue to flatline and even trend upward, is just a profound misunderstanding of the nature of what's happening and what these things, insofar as they measure anything, what they actually measure. And I'll add a third twist to this is that one of the things that NVIDIA has been very good at, people forget that NVIDIA back in the day was just a provider of graphics chips back in the 70s and 80s. I was an analyst, I covered one of their peers, a company called ATI Technology. Technologies back then. And they were just providers of these boards you put in your computer and you could play a Duke Nukem faster or something. But what changed was they lost that market because it went to gaming consoles, it went to PlayStations and other things. And so they had to sell more upmarket chips into a new market, and that became premium gaming. And to his credit, he realized that the best way to
44:46protect his margins was to position yourself as a premium provider with engineered. scarcity. And so you watch, if you watch their gross margins over time, they began ticking up in that period when they went into this premium gaming mode and there were the same calls, scarcity and everything else. And then as that began to fall away a little bit, along came crypto and the exact same phenomenon. He played the game again. Engineered scarcity, Jack the margins got up to like 55 and 60%. And then as crypto fell off in 2022, along came Jen. Gen AI. And we could play the scarcity game one more time. So this is not this thing that we're measuring in terms of the cost to compute is a reflection of not just disparate markets, but consciously engineered scarcity.
45:27Unidentified speaker (speaker_1): If we go back to the GFC, and this is where I wanted to finish, just try to solicit some of your thoughts on tail risk, the price of insurance, how to think through trading this. So my first question is just. if we look at the circularity that was the pre-GFC period, it was really to me, mortgage credit was getting supplied endlessly. You were able to reach very subprime borrowers, put them into teaser rate loans, that propelled housing prices higher and that collateral recycled throughout the system many times through CDOs and those types of structures. And once housing prices stopped going up and started falling, the whole thing started to reverse viciously. Is there a price in this system, the AI ecosystem, around which that could happen? In other words, if X price goes down or up, it's going gonna to set in motion a reversal of a lot of what's happened in the last. couple of years?
46:28Unidentified speaker (speaker_0): The naive answer has always been that these GPU rental prices. And I think that's wrong. I don't think they are for all the reasons I've just described. I think more likely the place, and this is one of my more out there tail risk hypotheses, is that the providers of these things will explicitly tell you that the target, the metric they look at, this sounds ridiculous, and I admit it straight up, is the TAM for global white-collar white-colored labor. So the TAM, the total available market for white-color labor. label. And they justify the AI capex on the basis of getting a material percentage of the TAM for global white-color labor. Now let's put aside the second for a second whether that's possible or whether that's ridiculous. Take them at their word that that's the objective. So that's the metric that they're optimizing for. And so they're saying to themselves, I need to get to $5 trillion in 2031, 2032 revenues to justify with appropriate hurdle and margin. the current AI capex spending. So if that's where they're headed and you're building on that basis, what we know about commodity markets is if you reprice a corner of the market by, say, for example, they're successful at getting some slice of that, you reprice the whole market. You don't reprice the corner of a commodity, you replace the entire thing. So when you reprice a slice of global white-collar white collar labor, you end up repricing much of white-collar white collar labor. You don't have 18% of a $30 trillion market anymore. You have 18% or you have a larger percent of a much smaller market because you've repriced the market. So one of the tail risks that I think about a
47:47lot is what happens in terms of this repricing at scale of white collar labor and what it means in aging societies with high dependency ratios, and how the entity that could end up being having to be bailed out is no longer banks or institutions or even frontier. AI companies. It's you and I. It's that the system itself has lost the ability to fund itself because of this repricing of a fundamental component of the system, which is to say labor itself. This is my out there idea in terms of what actually gets repriced. Labor itself gets repriced if you take them at their word. And that has consequences for the, how shall I say, the creditworthiness and functioning of the entire system. But in a way we don't predict. and don't expect because we're so busy focused on credit default swaps and yield spread and token rental prices and not taking the industry at its word that what it's trying to do is reprice white colour work.
48:36Unidentified speaker (speaker_1): If we look last little topic, we just go again, going back to the GSC, that period was one of incredibly lean vol in the 2005, six, 6, even early two 2007, thousand seven. You had CDS on the banks trading at twenty 20, 25 basis points. You could buy it for a song. It never moved. This cycle is very different. I mean, there's a lot of volatility, at least on the single stocks. They're very uncorrelated, but there is a lot of vol in equity markets. And I'm curious, as you talked about your business in venture of buying long data call options, do you have a thought process or framework in this context of how to think about? the price of those options? Where to find options that may be cheaper than the amount of uncertainty that could be coming, which I think you think is quite a bit.
49:25Unidentified speaker (UNKNOWN): So
49:25Unidentified speaker (speaker_0): two lenses we bring to bear on this. And one is just the typical venture capital rhetoric. But the only way to be successful in venture capital is to regularly ask yourself when you look at something, what if this goes unexpectedly right? And it's because it's too easy to say what if this goes wrong? There's always reasons it could go wrong. But what if this goes unexpectedly right? What does it look like? And we often ask CEOs and entrepreneurs this question. We'll say, just between you and I, whenever you wake up in the middle of the night and you say to yourself, you know, this could be a billion dollar company, and you sort of lie to yourself in the middle of the night about it, what is it you tell yourself that might go right? Rise. So you have to always be, we look for companies and places that have the kind of leverage where people can hypothesize. ways in which everything could go right. And not just in a naive sense, like I could get five percent of the Chinese market for, you know, avocados or something, but explosively right. And if you can't find a mechanism for something could go explosively right and get it cheap enough, we're out. And this is our answer is we need to see both aspects of it, inexpensive long dated call options on things that could go explosively right.
50:27Unidentified speaker (speaker_1): That's a great way to end. Paul, it's been a real pleasure. I really appreciate you taking the call to come on. And as I said, you know, the listeners of the Alpha Exchange are folks in markets wrestling with this stuff every day, and your perspective is valuable. So thanks very much.
50:42Unidentified speaker (speaker_0): Thanks for having me on your anniversary episode.
50:44Unidentified speaker (speaker_1): You've been listening to the Alpha Exchange. If you've enjoyed the show, please do tell a friend. And before we leave, I wanted to invite you to drop us some feedback. As we aim to utilize these conversations to contribute to the investment community's understanding of risk, your input is valuable and provides direction on where we should focus. Please email us at feedback alpha exchange podcast.com. Thanks again and catch you next
51:10Unidentified speaker (UNKNOWN): time.