Paul Kedrosky — AI is the First Bubble With Every Ingredient at Once

0:00 The frame — the first bubble with every ingredient at once

• "This moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in US history." Two supporting lines set the whole episode up: lenders don't care what happens inside the data centers, and "tokens are the first hyper-deflationary commodity in the history of modern economies."
• What draws him to it is scale — "big things that move in ways that are both unexpected and more consequential than people realize," the same lens he brings to the GFC, the dot-com meltdown, and back to canals and railroads.

1:42 Albert Bartlett and the exponential function

• "The greatest failing of the human species is an inability to understand the exponential function" — another way of saying things that scale quickly confuse humans. The algae-in-a-pond image: three periods before it covers the pond, it's only an eighth covered.
• AI is "the canonical current example of prodigious scale and exponentials that are completely misunderstood by people, both by bulls and by bears."

2:30 Two exponentials, not one — adoption up, price down 70–80% a year

• "There's an exponential adoption curve but there's also an exponential deflation curve. It's the fastest deflating commodity in the history of quasi-industrial commodities, falling something like 70 or 80% a year." Almost everyone models the first and ignores the second.
• 2:52 Grid interconnection can't keep up, so "bring your own power" — natural gas turbines behind the meter — is now applauded. The irony: the people promising AI will solve climate are, at the margin, among the biggest emissions contributors. "It's going to take a lot of AI to solve the problems created by AI in climate alone."

3:41 What every great bubble shares — and why this one has all four

• Two hundred years of "paroxysm moments" have a common kit: loose credit, a great technology story, sometimes a real-estate component, sometimes a policy angle (even the South Sea bubble had one). This is the first that sits at the intersection of all of them.
• That's why nobody can size it — "they approach it through the lens of credit or… real estate or the geeks on X approach it through the lens of five prompts that'll change your life." Precedents: railroads, canals, 1920s rural electrification — "massive amounts of capital allocated towards something hugely consequential and then created a lot of breakage along the way."

4:54 Owen Lamont, the bumper sticker, and Ritholtz's point about behavior

• Lamont: "Stocks are unreasonably high when prices rise to a valuation level that cannot be justified by rational forecasts of subsequent cash flows — then they double."
• The Bay Area dot-com bumper sticker: "Please God, give me just one more bubble, this time I'll know what to do." They didn't. Borrowing from Barry Ritholtz: behavior around capital, manias and fads doesn't change — "they panic at all the wrong times. They chase things at the wrong times." Not stupidity, just limited information against phenomena that exceed the ability to think concretely about them.

8:27 The original sin — extrapolating the future from coders

• The first domain where LLMs worked brilliantly was software, because software has a strict grammar and a tight gradient descent (you learn a lot from small errors, and being wrong has a real consequence — repaint a subroutine pink and it stops working). An essay about Tolstoy has no such feedback.
• "The first domain where AI was applied could hardly be less representative of AI's future if you tried… and yet that's the domain from which we're extrapolating our futures."

9:36 Compressive vs expansive — why token projections break

• White-collar AI is compressive: take a 40-page sell-side document, "give me five bullets." Coding is expansive: one prompt out, a million lines of code back. Token demand curves built on the expansive case don't transfer.
• 11:00 "The projections that were made in the first four years of large language model adoption are largely useless" — the venture-capital rule that early adopters aren't like anybody else, "crazy people who try half-baked products." Build for them and you never find the valuable customers.
• 12:01 That one sampling decision cascades into everything downstream: how much data-center capacity to build, internal vs external financing, hurdle rates, whether these are commercial real-estate projects, whether to syndicate the debt.

12:19 The Minsky moment — when the money divorces the application

• Financial episodes reach "escape velocity" where it no longer matters whether the applications match the original indication. The '07–'08 flywheel was pushing debt off the balance sheet so the liabilities were somebody else's problem — asset-backed securities and syndication.
• "Whenever these things become financialized and it becomes divorced from the underlying application, you need to pay a lot of attention. And that's the moment we're in right now."

12:58 The crossover — more than half of data-center financing is now external

• "In 2026, the first half of '26, we crossed over from more than half of the financing for data centers… no longer comes from internal cash flows." The bull argument was always it's their own cash flow, who are you to object — "and of course, what's happened was the predictable — it has now flipped."
• 13:36 The new money is "ABS, private credit, sovereigns and everything else, and all of these SPVs and new structures." The same people who said cash-flow funding was safe now say outside investors are smart money — "this is a narrative problem… motivated reasoning."
• 14:32 The flow is big enough to distort sovereign markets — "one of the reasons behind the bond freakout we just saw because it's pushing up longer-term rates." Lending to a data center with a hyperscaler behind it beats "this flawed credit known as the United States or the UK." The textbook crowding-out has reversed: data centers are crowding out sovereigns.
• 15:21 Lenders look through the building: "a 12-year lease renewable and they're good for it… much better than holding a 10-year." Hence more construction, more GPUs, more SK Hynix HBM — demand created by financing structure, not by usage.

17:08 35–40% GPU utilization against a scarcity narrative

• A study of one GPU warehouse ("think of it like an AWS of GPUs") puts usage at 35–40% — "eye-popping," given the story about scarcity pushing used A100 prices higher.
• The reconciliation is hoarding: "colossal amounts of hoarding… double and triple ordering because everyone's terrified" of being caught short at some parabolic inflection. Idle even at peak loads — "a kind of potential flux of product into the market at some point in the future."

18:29 Tokens — the first hyper-deflationary commodity, and the 400% stand-still

• Deflating 70–80% a year under both a technology curve and a capital flood: "as tokens brush up against other parts of the economy, they do huge damage to it… you see huge deflationary waves entering that sector." Great for consumers, "toxic to the frontier companies."
• 19:41 The arithmetic: at −80% price, holding revenue flat needs 400% unit growth minimum — before pleasing Wall Street, and before servicing debt that is now external. "This is among the most difficult things in capitalism. Just to stand still I need to grow 400% year-over-year."
• 20:37 Read the frontier numbers through that lens: OpenAI's 18% quarter-on-quarter "was deemed a disappointment."

21:09 15–18% of investment grade is now data-center related

• Per a JP Morgan report: data-center-related paper is now 15–18% of the IG market — "if it were a sector, which it isn't, it's now larger than the financial services sector."
• 21:36 The reversal of tech's defining attraction: "one of the prime attractions of technology after growth was that it had no debt. The balance sheet was pristine. These were cash flow monsters." Plus maintenance capex — Michael Burry's point — "continuing capex obligations into eternity to keep these data centers at the cutting edge."
• 21:55 That turns orthodox tech companies into utilities "with similar obligations, but wildly overvalued compared to an orthodox utility." So: "who gets re-rated first? Do utilities all of a sudden see a huge spike in valuation, or do technology companies… get re-rated to look more like utilities? And I obviously think the latter."

22:56 The IPO-supply chart — $4–5.5 trillion, more than all post-WWII listings

• The one bubble ingredient absent until now was supply. If SpaceX, Anthropic and a couple of others list, they're "bigger than all the IPOs from the '90s combined" — "not just that, it's all post-World War II combined." His $4T estimate is now "more like 5.5 trillion."
• 23:24 Faber's corollary: even in listed names the direction has flipped — "from companies like Oracle, like buybacks versus share issuance and dilution… shifting pretty hard in the other direction."

24:39 Where the money comes from — you sell the liquid winners

• Retail imagines big funds have "a printing press of cash in the basement." They don't: cash is a performance drag, so allocations to a mega-IPO are funded by selling other stocks.
• What gets sold has three signatures: most liquid (minimal price impact), overlapping (avoid doubling factor exposure), and perversely the best performers ("I want to lock in my gains… and tell everybody what a good boy or girl I am").
• 25:29 He modelled this in March/April: the selling has to start "weeks and even months in advance… in a way that the market's not paying attention," which is why the pressure on winners showed up in April–June.
• 26:11 He believes that dynamic "played into the implosion at Situational Awareness" — long the most liquid, best-performing names that were destined to become the funding source. "Anticipation of the upcoming flood of new issues became a pin." Structurally predictable, not idiosyncratic.

28:36 "At high valuations, failure is overdetermined"

• Nike is down 75% and every post-hoc explanation names product, competitors, politics or macro — none names the fact that "the P/E ratio at the peak was like 70 and now it's 20."
• 29:22 The math: twenty independent ways to fail at ≤5% each compounds to "greater than a 60% chance of failure in the period that the 5% applies." "So what looks not particularly risky, or… unpredictable, is actually highly predictable." The mechanism is being "pecked to death by ducks."
• 29:52 Same for AI — sovereigns, rogue AIs, or the one he's watching: unprecedented cash inflows into Taiwanese and Chinese chip makers implying "a tsunami of supply in early 2028." In a boom-bust industry "once you lock in supply, my friend, prices are going to zero" — fixed costs force product out the door.
• 30:39 Semis are "probably the most capital intensive boom-bust industry on earth," and the current this-time-is-different story is the super-cycle — RAM never coming back, GPUs a permanent duopoly. In ten years it'll be "remember when Micron was at such and such a price."

31:40 Vibe chipping — AI-designed silicon dissolves the semiconductor moat

• A Wall Street Journal story on the inference-ASIC startup Edged looked like the usual brash-founders piece; the real news was the process — "42 days in the initial design verification stage versus what should have probably been six or seven months," and "their first design worked," which normally never happens.
• Chip design code sits "exactly in the sweet spot of what large language models do really well," so expect "vibe chipping": new designs arriving at fabs at unprecedented rates, low-power ASICs running on-chip LLMs inside security cameras, and "the notion that some kind of tribal knowledge protects you as a chip manufacturer is going away."
• The recursion: "the process of producing chips itself is being transformed by AI, which in a perverse sort of way will help prick the bubble itself."

34:54 Models have flatlined — and harnesses are hiding it

• "Large language model performance really hit its maximum inflection in terms of year-over-year change almost four years ago, 2022, 2023." Ignore the promoted benchmarks — models ingest them, "no more than if you had seen the SAT before you'd done it."
• On composite indices, year-over-year gains "essentially flatlined over the last 6 months": 10–12% has become "one or two% at most." We're at "the iPhone 7 moment" for LLMs.
• 35:38 The mask is harnesses — Claude Code, the codexes — wrapping the model "and giving the appearance of things changing much faster than they are." His Sound of Music analogy: models are the bratty kids, harnesses are Julie Andrews making them sing.

36:33 Convergence — Anthropic, Qwen, DeepSeek behind a harness

• Cross-model variance "has collapsed": "very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek… in practical composite terms." His blind Pepsi-Coke tests — "inevitably no one can tell the difference."
• 37:24 Therefore "the game is almost over in terms of pretending that you can justify multi-billion dollar training runs," and "the most successful frontier AI company will be the first one to stop pretending they can train new AI models — because it's basically a gift."
• 38:27 Why the outputs converge: "the median data nugget inside of a large language model is a 37-year-old male on Reddit." Homogeneous training data, homogeneous answers — so models will amplify herding, not break it (contra the Economist piece on the TikTok-ification of tourism).

38:49 75% don't want a data center in their county — the agency problem

• Higher opposition than nuclear. His read: lost agency. "It's a great hulking presence that I didn't ask for" — a physical and digital manifestation of losing control over work, relationships and place, the same reflex that drives American attitudes to vaccines.
• 40:18 Anomalous, since the US is normally the earliest technology adopter ("we've never seen a digital toaster we didn't like"). Yet the US is more negative on AI than other G7/OECD countries, and OECD countries are more negative than sub-Saharan Africa, where it reads as a leg up. Partly self-inflicted — "I blame Dario and others… for all of the 'some large percentage of jobs are going to disappear.'"
• 41:32 The uniquely American amplifier: "any threat to employment is a threat to health care, and a threat to health care is a threat to… your personal solvency." Elsewhere job loss doesn't mean bankruptcy from dental work. "We've largely promoted AI on the basis of job losses. So big surprise."

42:43 The chart that started it — AI as a majority of US GDP growth

• He first published the AI-share-of-GDP-growth statistic half expecting to be corrected. Instead "that share grew and has persisted over the last six quarters since I first began writing about it." For "probably the sixth time in Western history, we have some force that's non-governmental that's actually so large as to shift the tides of GDP growth."
• 43:36 The dog-and-mailman causality trap: the dog barks, the mailman leaves, the dog takes credit. "If you thought your policies, tariffs or whatever else, were causing US GDP to grow… you have a messed up model of causality." With AI >50% of US growth, "your tariffs did not cause the US GDP to grow in that period. What caused it was this remarkable thing called data centers and hyperscalers" — and mis-attributing it invites "actions that are actually really consequentially negative."