The Most Dangerous Investment Cycle In History | Paul Kedrosky
Watch on YouTube →
Overview
Paul Kedrosky argues that the AI buildout is an unusually dangerous investment cycle because it combines a fast, labor-substituting general-purpose technology with loose credit, data-center real estate speculation, and government-backed US-China competition. He connects the scale of capital spending to labor repricing, concentrated AI revenues, GDP and S&P 500 exposure, and the possibility that debt and excess capacity could weaken the usual policy response after a downturn.
Key takeaways
- Kedrosky's central historical distinction is that AI can substitute for white-collar work at a pace faster than earlier general-purpose technologies, so past claims that technology creates more jobs than it destroys should not be treated as guarantees.
- If data-center spending approaches $10 trillion, Kedrosky's return-on-capital calculation implies $4 trillion to $6 trillion in revenue is needed by the early 2030s, putting a large share of white-collar work under pressure to be repriced.
- AI investment is unusually exposed to four crisis drivers at once: loose credit, real-estate speculation through data centers, a powerful technology narrative, and government-backed US-China competition.
- Kedrosky says AI capex supplied 25% to 60% of quarterly US GDP growth in recent quarters, while seven major stocks produced nearly 45% of five-year S&P 500 gains, making headline growth and index diversification potentially misleading.
- A market where Chinese providers offer tokens near marginal cost and ordinary users cannot distinguish frontier models could face severe pricing pressure, even as US labs continue training models to preserve their growth narratives.
- If AI-driven borrowing leaves companies highly leveraged and capacity underused, Kedrosky expects low interest rates to be less effective because businesses may prioritize paying down debt over taking on new loans.
Chapters
- Kedrosky combines experience as an equity research analyst, a venture fund manager for nearly 15 years, and a researcher in complexity and path dependency.
- He says his academic work on how systems become fragile informs his analysis of financial cycles.
- He has published independent research for more than 15 years and works with a small number of hedge funds and institutional investors.
- Kedrosky compares AI with railroads, electrification, highways, and fiber as a general-purpose technology that supports activity across the economy.
- Unlike many earlier technologies, current AI can substitute for symbolic and white-collar tasks as well as complement workers.
- He cautions that the claim that technology always creates more jobs than it destroys is based on a small set of historical examples, most of them more complementary than substitutive.
- Kedrosky expects AI's disruptive effects to arrive faster than railroad deployment, the industrial revolution, or the roughly five-to-six-year fiber buildout.
- He identifies four recurring crisis drivers: loose credit, real estate speculation, technology narratives, and government policy; he argues the AI cycle brings all four together.
- Data centers link the boom to commercial real estate, while US-China competition and claims of an existential technological race reinforce spending.
- Kedrosky says projected data-center spending could approach $10 trillion over four or five years, if current investment plans continue.
- A return-on-capital calculation implies roughly $4 trillion to $6 trillion in annual revenue may be needed by the early 2030s.
- He estimates that scale would require repricing about 40% of US white-collar work or roughly 23% to 25% of the global white-collar market, with potential layoffs in the low millions.
- Kedrosky describes inference as a scale business where the marginal cost is largely power, heat, and operating the data center.
- Frontier labs launch new models partly to sustain premium pricing above token-production costs, but cheaper Chinese models increase competitive pressure.
- He says Chinese providers are producing tokens near marginal cost with models roughly eight months behind leading US systems, intensifying price competition.
- Kedrosky says software engineers accounted for an estimated 70% to 80% of early frontier-model usage, making them an unrepresentative early-adopter group.
- Coding can expand a short prompt into large amounts of code, while many ordinary white-collar tasks compress long documents into summaries or decisions.
- He argues that most users may not notice small differences between frontier models, limiting providers' ability to charge a premium for every performance gain.
- Kedrosky says composite benchmark improvements have slowed from about 18% to 19% year over year to roughly 5% to 6%.
- He argues that stopping model training could lower a frontier lab's costs, but the need to demonstrate progress to investors keeps the spending race going.
- He cites reports that more than 70% of Anthropic's revenue comes from eight companies, with Meta alone possibly accounting for 17% to 18%.
- Kedrosky estimates AI-related capital spending contributed around 25% to 60% of quarterly US GDP growth over the previous seven or eight quarters.
- He warns that this spending can obscure weakness elsewhere in the economy and distort interpretations of policy effects and inflation.
- He says the Magnificent Seven generated almost 45% of S&P 500 gains over five years, while technology made up about 55% of the index, leaving index investors heavily exposed to AI.
- Kedrosky says AI-related borrowing competes with US Treasury issuance as more than half of Treasury buyers are now yield-sensitive investors rather than yield-insensitive foreign governments.
- He compares a possible post-boom deleveraging to Japan after its 1980s and 1990s real-estate crisis, where low rates did not persuade indebted companies to borrow and invest more.
- He argues that high public debt and corporate leverage could make both fiscal stimulus and lower interest rates less effective after an AI investment downturn.
- Kedrosky disputes the expectation that modern policymakers will reliably prevent depression through stimulus, arguing political resistance and debt-market pressure could delay action.
- The host points to the Federal Reserve's post-2008 response and COVID-era spending as evidence that governments can deploy quantitative easing and fiscal aid.
- Kedrosky says a deflationary crisis could still produce austerity and policy mistakes, drawing a comparison with resistance to deficits in the late 1920s and early 1930s.
- Kedrosky's framework asks whether removing a product's complexity unlocks demand or reveals that users do not need the product at all.
- He uses Photoshop as a case where image prompting might replace the need to learn complex software, and Cloudflare as a case where AI connectors can make useful services easier to access.
- Roblox could benefit if AI reduces the difficulty of building on its platform, while a surge in app-store submissions alongside falling ratings per app suggests many new apps may not find users.
- Kedrosky says AI has commoditized basic statistical modeling and routine analysis, reducing the investment edge available from work that is easy to replicate.
- He expects more value to come from synthesis across disciplines, heuristics, and experience rather than relying on creativity alone.
- He describes his own approach as iterative, building research around client questions and testing what produces an edge before it stops working.
- Kedrosky sees models as strong at producing the center of a distribution but weak at identifying tails and inflection points, where he is exploring tipping-point detection in complex systems.
Transcript excerpt
0:01 Paul Kedrosky, welcome to Hidden Forces. >> Hey, thanks for having me. >> I'm excited to have you on. This is your first appearance on the podcast. And uh >> It's the first appearance on this podcast. Unfortunately, it's not the first appearance on any podcast, but yeah. >> No, you are a prolific podcast guest, Paul. How long have you been How long have you been independently So, actually, let's just start with the background for people that don't know you. Just give give me your story here. Like, how did you get involved in this space? Did you come in from the finance side? Did you come from the technology side? Like, what's your origin story? And how long have you been independently producing research and putting it out for the public? >> Uh so, I come at it from both the finance and technology side, I suppose. I mean, I'm My day job, one of my day jobs is I I run a a series of venture
0:48 funds with a partner of mine that we've been doing for almost 15 years. So, you know, so I come at it from a sort of financial technology side. But way back, way back, I was actually an equity research analyst at one of the major brokerage firms. And so, where I was coming at this largely from a financial standpoint, this by which I mean, you know, com- financial systems and how they function and how companies do what they do. And so, I come at it from sort of a sense a synthesis of those two perspectives, I guess. And way, way, way back, um in grad school, uh my my PhD work my grad school work was in sort of complexity and path dependency about how systems get into fragile states and what
This is the start of the transcript. Save this video free for the whole transcript, searchable and time-coded.
Related notes
Summary, takeaways, and chapters were generated by AI from the video's transcript and may contain errors. The video belongs to its creator, Hidden Forces.