The Most Dangerous Investment Cycle In History | Paul Kedrosky

Anthropic

Across the shows: 44 speakers in other episodes, 24 bullish and 37 bearish calls, past 30 days

• Paul Kedrosky · SK VenturesBearish

  Anthropic's revenue is unprecedentedly concentrated in a few customers, creating dependence risk.

  Per leaks, over 70% of Anthropic's revenue comes from eight companies, and Meta alone may be 17-18%. For a company reportedly on a $50-70 billion run rate, that concentration is unprecedented. Paul compares it with analyst experience: a concentrated customer base can cause an earnings air pocket when one buyer holds off. Defense contractors have similar concentration but with long-term government contracts, while a commercial company dependent on a few buyers gives those customers leverage to dictate terms.

  “we're told that something in excess of 70% of their revenues are eight companies. One alone, Meta, might be 17 to 18% of [Anthropic's] revenues.”at 30:52

Cloudflare

• Paul Kedrosky · SK VenturesBullish

  Falling product complexity, aided by AI connectors, is driving a tailwind for Cloudflare's usage.

  Kedrosky says complexity used to limit how many people could use Cloudflare's many products. AI tools, MCPs and connectors now let people use these services without dealing with the complexity themselves. That has opened up a lot of its products, which he calls an incredible tailwind. He notes the stock has tripled over about eighteen months, and says most analysts tell simplistic AI stories. He presents this as a way of thinking rather than a bull call.

  “just to pick on them, that's created an incredible tailwind in their business. And so I think the stock has tripled or something over the last eighteen months.”at 52:53

Roblox

• Paul Kedrosky · SK VenturesBullish

  Easier app-building on Roblox from AI tools helps Roblox, since its platform becomes more useful.

  Building on Roblox has long been annoying, and Kedrosky hears this from people. As AI tools reduce the complexity of producing apps for the platform, he argues Roblox's large gaming ecosystem becomes more accessible to developers. He says this is good for both builders and Roblox, which already drives a large share of YouTube gaming content. He relies on second-hand accounts for the complexity point.

  “it's not all of a sudden that you realize that Roblox is irrelevant. You realize that, wait a minute. There's this incredible platform sitting there that I could launch applications and run on top of, which is good for me and good for Roblox.”at 54:24

AI & jobs

• Paul Kedrosky · SK Ventures

  AI substitutes for white-collar work, faster and more than past general-purpose technologies.

  Unlike railroads, electricity or highways, which complemented human work, AI is mostly aimed at symbolic white-collar tasks, so it substitutes for tasks and occupations rather than just adding productivity. Early evidence includes call-center data from Avelo and Anthropic reporting less entry-level hiring. Because the substitution is also happening far faster than past technology diffusion (railroads took 25 to 30 years, fiber about five or six), Paul expects significant layoffs and social disruption. He agrees with Andrew Yang's alarm, calling it 'baked in'.

  “So the two things that are different, it's much more substitutive and it's happening much faster. So the two together make it much more consequential in terms of the likely disruptive force that it has in the economy.”at 10:42

• Paul Kedrosky · SK Ventures

  AI-driven capex requires repricing about a quarter of white-collar work, not just layoffs.

  Taking the data center buildout as rational for argument's sake, a return-on-capital calculation (similar to analyses from Columbia and Bain) implies roughly $4-6 trillion in AI revenue by the early 2030s. Since the spending mostly replaces existing spending, that revenue must come from white-collar work. $4-5 trillion is about 40% of the US white-collar workforce and about a quarter of the global market for white-collar work. Paul argues this means at least a quarter of white-collar work gets repriced, which may show up as fractional work or lower pay rather than mass layoffs. Once one segment reprices, the rest of white-collar pricing also shifts.

  “So four to five trillion dollars is about 40% of The US white collar workforce. ... It doesn't necessarily mean layoffs. Right? So you have to be careful about making the distinction. But what AI is doing is repricing the value of cognitive work.”at 20:06

• Paul Kedrosky · SK Ventures

  AI has commoditised basic analysis, so analytical edge from standard modelling is disappearing.

  Kedrosky says AI makes basic modelling easy and largely a commodity, so he now needs less work to produce models. He argues that anyone claiming to do analysis faces the same commoditisation, so edge from straightforward analysis is falling away. He thinks any remaining edge must come from synthesis across domains.

  “And I think that's now largely a commodity ... the notion that you can have an edge from straightforward analysis is largely falling away.”at 56:15

US growth & recession risk

• Paul Kedrosky · SK Ventures

  AI capex accounts for 25-60% of quarterly US GDP growth, distorting the economy.

  Paul was drawn to this topic by US GDP growth that seemed unexplained by orthodox statistics. Digging in, he found that over the last seven or eight quarters, AI capex accounted for somewhere between 25% and 60% of quarterly GDP growth, varying by quarter. He calls this wildly unusual and incredibly distorting, which means headline growth depends heavily on a single spending category.

  “for at least the last seven, eight quarters, something like... And it varied from quarter to quarter, 25% to 60% of GDP growth on a quarterly basis has been tied to AI CapEx, which is, again, wildly unusual, incredibly distorting”at 33:29

• Paul Kedrosky · SK Ventures

  AI capex masks weakness in the rest of the US economy, so headline GDP overstates underlying growth.

  Because AI capex is such a large share of GDP growth, it hides the fact that most other parts of the economy are flat or shrinking. Kedrosky argues this leads to a wrong causal model of growth. He gives the tariff example: people who think tariffs work see strong GDP and misread what drove it, when the real driver was AI spending.

  “because when you're that such a high percentage of GDP growth, it masks the weakness and everything else.”at 34:05

• Paul Kedrosky · SK Ventures

  Highly indebted firms will deleverage over the next decade, so lower rates won't revive borrowing.

  Kedrosky says highly leveraged companies don't respond to lower rates by borrowing more. They prioritise reducing debt, as Japanese firms did after their real estate bust. He applies this to today's large companies with rising debt and long-term lease commitments. Rates at zero won't make them borrow, so monetary and fiscal policy will be less effective than usual. He says they'll spend the next decade reducing indebtedness.

  “You lower rates to zero, are they gonna take on more debt? Absolutely not.”at 43:50

AI infrastructure & data centers

• Paul Kedrosky · SK Ventures

  Data center capex is irrational; the AI buildout has all four ingredients of past crises.

  Paul says data center spending is not rational, although he assumes it is for argument's sake. He projects roughly $10 trillion of data center spending over four or five years. Most major Western financial crises had one or two of four drivers: loose credit, real estate speculation, a technology story, or government backstops. This episode has all four: a strong technology narrative, data centers as implicit commercial real estate, very loose credit (though tightening after the CoreWeave debt scare), and a US-China 'existential' race that justifies spending at any cost. AI capex now exceeds every prior episode as a share of GDP in 150 years, larger than the telecom build-out, PCs, rural electrification, railroads and canals, and behind only WWII rearmament and the interstate highways.

  “As a percentage of GDP, as a CapEx moment, we're at... We've now exceeded all prior episodes in the last hundred and fifty years. It's larger than the telecom build out...”at 17:01

• Paul Kedrosky · SK Ventures

  AI capex is adding roughly a quarter to over half of US GDP growth, a railroad-scale driver.

  Kedrosky estimates AI capex accounts for something between a quarter and a little over half of GDP growth depending on the quarter. He says that makes it comparable in materiality to the railroads, and it is the dominant driver of recent growth. He argues the concentration is so large that it distorts the GDP picture.

  “So something on the order of between a quarter and a little over half depending on the quarter, and that's up there with prior episodes. It's probably on the order of the railroads, something on that order in terms of its materiality with respect to GDP growth.”at 34:05

Frontier AI labs & model race

• Paul Kedrosky · SK Ventures

  Labs can't slow spending; the US-China prisoner's dilemma locks in the race.

  Each firm believes that if it stops spending, competitors will win a market likely dominated by a few players in inference and token supply. That makes continued spending mandatory. Anthropic's call to 'pace the frontier' is a prisoner's dilemma: even if labs agree to slow, China may not, and then China wins. A West Coast culture that treats technological progress as inexorable and anti-slowdown reinforces the dynamic. Capital markets, collecting fees, add more money to keep it going.

  “What if we all agree but China doesn't? Well, then we're into this problem where China wins again, and we're back to this existential crisis. And so the race dynamic is baked into the very nature of what we're doing.”at 17:01

• Paul Kedrosky · SK Ventures

  Frontier model gains are slowing: composite benchmark improvement fell from ~18% to ~5-6%.

  Paul argues that model progress is becoming asymptotic. On composite benchmarks, year-over-year improvement fell from about 18-19% three years ago to about 5-6%. Many benchmarks have also been ingested by the models, so headline results overstate progress. Labs keep launching new models and citing benchmark wins to maintain the story that justifies funding. If training returns had not declined, the first lab to stop training would win on costs, but the race makes that move impossible.

  “while models are improving, it's becoming somewhat asymptotic... We went from, like, say, around 18, 19% performance year over year on composite benchmarks three years ago to more like five or 6%.”at 28:46

US rates & the Fed

• Paul Kedrosky · SK Ventures

  AI capex is masking inflation data, pushing the Fed to raise rates while the economy is weak.

  Kedrosky argues AI capex distorts inflation readings as well as GDP. He says that distortion is causing the Fed to raise rates even though most of the economy is fairly weak. He treats the mistake as a consequence of a misread causal model.

  “AI CapEx is badly masking what's happening in inflation, which is causing us to raise rates when most of the economy is actually fairly weak.”at 35:37

• Paul Kedrosky · SK Ventures

  Fed independence is under pressure, making policy a series of rolling mistakes.

  Kedrosky argues that the crisis-era response produced politicisation of the Fed, with its mission increasingly under pressure. He says this is part of why he expects policy to deliver a series of rolling mistakes rather than a technocratic fix. He sees the trend as continuing.

  “It gave rise to the politicization of the Fed, where the entire Fed mission is increasingly under pressure.”at 49:11

US Treasuries & yields

• Paul Kedrosky · SK Ventures

  Rising long-term yields reflect US debt worries plus AI debt competing for yield-sensitive buyers.

  Kedrosky says the rise in long yields reflects both US debt concerns and competition from AI-related debt issuance. Sovereign buyers used to be yield-insensitive because they had trade surpluses to park. Now more than half of Treasury buyers are yield-sensitive hedge funds and individuals. A triple-A hyperscaler bond can compete for the same money, so each AI issuance pushes up yields at the margin.

  “More than half of the buyers of treasuries ... are yield sensitive. It's not sovereigns anymore. It's hedge funds. It's individuals.”at 41:44

• Paul Kedrosky · SK Ventures

  Investors' rising yield demands are already signalling nervousness about US debt before any crisis.

  Kedrosky says debt markets are already requiring yields not seen in decades, which reflects nervousness about US finances. He argues the market would keep demanding higher yields if the US printed money to fund a fiscal rescue. He disputes that the US can spend with impunity because it has balance-sheet room, saying the market is already signalling otherwise.

  “what the market is telling you even in advance of something like that happening is that there's real nervousness in debt markets such that they're requiring considerably higher yields, yields we haven't seen in decades.”at 45:28

Credit markets

• Paul Kedrosky · SK Ventures

  Capital providers earn fees on AI debt structures without holding the risk, feeding overspending.

  The financiers backing the race are paid for issuance, not for holding the debt. They earn fees by creating special purpose vehicles, off-balance-sheet financing and syndication for AI infrastructure. That incentive makes capital endlessly available and keeps the spending race going. Paul adds that loose credit is starting to tighten after a debt-market scare over CoreWeave and others in the prior few weeks.

  “They're making money from an agency business of helping to create these special purpose vehicles, off balance sheet financing, syndication... So they make money from issuance, not from holding.”at 17:01

AI pricing & unit economics

• Paul Kedrosky · SK Ventures

  Frontier token prices won't fall to marginal cost; new models are a halo-pricing defence.

  If AI is a scale business that consolidates, providers compete down to marginal cost, which is mostly power and cooling for inference. Frontier labs resist this by launching constant new models to justify a premium, a halo price above production cost. Pricing pressure is worse than it looks, because most mainstream users (not the software engineers who dominated the first four years) don't care about frontier performance and can't tell models apart in blind tests. China already sells tokens near marginal cost, roughly eight months behind US models, so the premium is hard to defend.

  “And way to think about that is it's a defensive tactic to try and make sure that they can charge what the marketers call, you know, have a halo price above the cost of actually producing tokens.”at 23:19

US–China AI & chip race

• Paul Kedrosky · SK Ventures

  China makes tokens near marginal cost, roughly eight months behind US models, far cheaper.

  China is trying to become an industrial-scale token manufacturer, producing tokens near marginal cost. Benchmarks from groups like Epoch AI suggest its models trail US models by about eight months. Its tokens are sold at a much lower price. Paul uses this to argue that US frontier labs cannot charge a premium for model quality indefinitely, which squeezes US pricing and unit economics.

  “China is trying to essentially become like almost an industrial manufacturer of tokens. So producing tokens at very close to the marginal cost of production that are, say, eight months behind US models in terms of their performance... And they're producing these tokens at a vastly lower price.”at 24:55

Tariffs & trade policy

• Paul Kedrosky · SK Ventures

  Strong US GDP under tariffs reflects AI capex, so tariff believers misread the growth data.

  Kedrosky says that if you think tariffs work, you look at strong GDP growth and conclude tariffs had no impact. He calls this a misunderstanding of what drove GDP, which was AI capex. He uses it as an example of how a wrong causal model leads to bad decisions about growth.

  “If you believe that tariffs work, you look at GDP growth and say, oh, look. It had no impact. As a matter of fact, GDP was growing in a really sprightly way. And that's just a misunderstanding”at 35:01

AI bubble & tech valuations

• Paul Kedrosky · SK Ventures

  The S&P 500 is now an AI bet, with the Mag Seven driving about half of index growth.

  The Magnificent Seven, largely AI-related, delivered almost 45% of S&P 500 growth over the last five years. Kedrosky says tech is now about 55% of the index on a proper reading. He concludes that a 'diversified' index fund is really a concentrated AI bet, and that both concentration and valuation levels are back at levels last seen in the railroad and rural electrification episodes.

  “They were almost 45% of S and P five hundred growth for the last five years.”at 36:08

US consumer

• Paul Kedrosky · SK Ventures

  Household stock exposure is at near-record levels, with more wealth tied to the market than ever.

  About 52% of Americans now own stocks, and a slightly higher fraction of their total financial wealth is exposed to the stock market. Kedrosky says both measures are near record levels. He presents this as a concern because households are heavily exposed to a market he sees as concentrated in AI.

  “So it's one thing to say I own a couple of stocks, but it's another thing to say a significant fraction of my wealth is tied up in the stock market. Both of those are at near record levels.”at 38:08

US deficit & government debt

• Paul Kedrosky · SK Ventures

  Stretched US debt limits the government's capacity to bail out the economy in the next crisis.

  Kedrosky argues the state's capacity to keep bailing out the system is already stretched, because the US is still working off the last two or three bailouts. He says this is visible in the size of the debt and deficit and in the market punishing US sovereign debt. He also argues that 1930s-style austerity pressure makes fiscal rescue less likely, so a sharper fall would come before policy shifts.

  “the state capacity to continue to do these bailouts is very stretched and you can see that because of the size of The US debt and deficit”at 39:43

Big Tech platforms & digital ads

• Paul Kedrosky · SK Ventures

  App store launches are surging, but per-app ratings have collapsed, pointing to a saturated app market.

  Kedrosky points to a surge in App Store submissions. He says the less-visible side is the collapse in ratings per app, which he links to a National Bureau of Economic Research report. He concludes that people's ability and interest in adopting new apps is near saturation, so launching more apps isn't producing more attention.

  “we're launching lots of things, but people's ability or even interest or awareness of these things has really fallen off a cliff.”at 55:17

AI: overall outlook

• Paul Kedrosky · SK Ventures

  Language models excel at the centre of distributions but are weak at tails and inflection points.

  Kedrosky describes next-token prediction as pulling outputs toward the centre of a distribution. That is often right, but it misses every inflection. He says he is running machine-learning experiments on tipping-point detection in complex systems, and he focuses on tails and inflections because he thinks that is where value lies.

  “I think models are outstanding at generating the center of the distribution and genuinely terrible at the tails.”at 1:00:45