Intro
Paul Kedrosky, a partner at SK Ventures, argues that the AI build-out is the first US bubble to sit at the intersection of every force that produced the previous ones, and walks through what changed when more than half of data-center financing stopped coming from cash flow. Meb Faber brings his own charts on IPO supply and Nike's multiple into it.
Guest: Paul Kedrosky, partner at SK Ventures and a fellow at the MIT Initiative on the Digital Economy, formerly a sell-side analyst
Host: Meb Faber, co-founder and chief investment officer at Cambria Investment Management
Published: 28 August 2026 on The Meb Faber Show
Show notes | 45 min
Key Takeaways
This is the first bubble that has every ingredient of the big ones at the same time
Loose credit, a real technology story, a real estate component and a policy angle, all present 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"
More than half of data-center financing now comes from outside the hyperscalers
The crossover happened in the first half of 2026, into ABS, private credit, sovereigns and SPVs
Tokens are deflating faster than any commodity in modern economic history
"Tokens are the first hyper deflationary commodity in the history of modern economies."
At 80% annual price declines a frontier company needs 400% unit growth just to stand still
Lenders are underwriting the credit behind the data center, not what happens inside it
"there could be hide and go seek competitions going on inside the data centers"
GPUs are being hoarded, and utilization is far below the scarcity story
"The GPU usage is only sitting at around 35, 40%"
Technology companies are being turned into utilities with debt and permanent capex
Kedrosky thinks tech gets rerated toward utilities rather than utilities rerated up
The coming IPO supply has to be funded by selling today's most liquid winners
He puts the new issuance from a narrow group of private companies above $5 trillion
At high valuations, failure is overdetermined
Twenty independent 5% failure modes compound into better-than-even odds of failure
Frontier model gains have flatlined and the models have converged on each other
Year-over-year composite gains have gone from 10 or 12% to one or 2%
"There is very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek or somewhere else in practical composite terms."
AI is about to collapse the chip industry's barriers rather than protect them
One startup ran initial design verification in 42 days instead of six or seven months
American hostility to AI is about lost agency and health insurance, not technophobia
AI capex, not policy, is what has been driving US GDP growth
Why This Moment Has Every Ingredient the Big Bubbles Had
Kedrosky's interest is scale rather than any particular mania — the global financial crisis, the dot-com meltdown, canals and railroads are all the same category of thing: big movements that are more consequential than people realize
He leans on the physicist Albert Bartlett's line that "the greatest failing of the human species is an inability to understand the exponential function", and on the algae-covering-a-pond illustration where the pond is half covered one period before it is entirely covered
AI is misunderstood by bulls and bears alike because there are two exponentials running at once, not one — an exponential adoption curve and 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"
The grid cannot keep up with the scale, so operators are being applauded for bringing their own power — natural gas turbines behind the meter — with the emissions consequences that implies
On the joke behind the episode's premise: Dario Amodei, Sam Altman and others promise AI will solve drugs, longevity and climate, while one of the biggest marginal contributors to emissions right now is the gas being burned to run the data centers
Looking back two hundred years, the largest financial paroxysms share a set of features — loose credit, a great technology story, sometimes real estate, sometimes a policy angle, traceable back to the South Sea Bubble
"this moment is the first one that sits at the intersection of all of the forces that created the largest bubbles in US history" — which is why people talking about it through only credit, or only real estate, or only prompts keep missing the size of it
Nobody Learns, Including the People Who Lived Through the Last One
Faber put to him a line from a previous guest, Owen Lamont: "stocks are unreasonably high when prices rise to a valuation level cannot be justified by rational forecasts of subsequent cash flows, then they double"
Kedrosky's answer was a Bay Area bumper sticker from the end of the dot-com crisis — "please God, give me just one more bubble, that this time I'll know what to do" — and the observation that it turns out they don't know what to do
The behavioral pattern does not change, a point he credits to Barry Ritholtz's playbook: people panic at the wrong times and chase at the wrong times, every time
He is careful not to call it stupidity — people are doing the best they can with limited information against phenomena at a scale they cannot think about concretely
"All the same phenomena that we saw in the GFC in the lead up are all in place again now. They just have different names."
Coders Were the Least Representative First Customer Imaginable
Faber's framing was that these numbers only land when they are made relatable, and offered one of his own: "these stocks are now bigger than the entire energy sector combined"
Kedrosky's structural objection is that token-usage projections are drawn from a badly skewed sample
Software is unusual on three counts at once — a strict grammar, a tight gradient descent where small errors teach a lot, and a real consequence to being wrong
Repainting a subroutine to be prettier gives you software that does nothing; rewriting a sentence about Tolstoy might just be an A as well
The bigger difference is direction. Most white-collar AI use is compressive and software is expansive
"I take a huge sell side document, 40 pages long, and say, give me five bullets" — that is what the rest of the economy wants from AI
A prompt for a fitness app comes back as a million lines of code, which is the opposite motion
"the first domain where AI was applied could hardly be less representative of AI's future if you tried" — expansive, tight grammar, rapid gradient descent, and almost nothing else in economic life shares those properties
"the projections that were made in the first four years of large language model adoption are largely useless" — not that the models are useless, but that the extrapolation is
The venture capital version of the same lesson: early adopters are unlike everybody else, and building for them is how a company never finds the customers who matter. Kedrosky says when a founder tells him a product is perfect for him, he answers that he is completely weird and nothing perfect for him is good for five other people on Earth
The Financing Crossed Over, and Almost Nobody Changed Their Argument
The pattern he watches for is financialization — when the money becomes divorced from the underlying application, which is the moment sometimes called a Minsky moment
The flywheel in 2007-08 was pushing debt off the balance sheet so the liabilities never came home; the AI flywheel is the same shape
The crossover has already happened: "crossed over for more than half of the financing for data centers specifically, but let's call it hyperscalers writ large. It no longer comes from internal cash flows"
"Now most of the financing as of mid twenty twenty six is external financing coming from a constellation of things like ABS, private credit, sovereigns, and everything else", plus SPVs and new structures
The people who used to dismiss him have reversed themselves without noticing. The old rebuttal was that this is smart companies spending their own cash flow; the new one is that it is coming from outside investors who are smart — the opposite claim, offered as the same reassurance
He calls that motivated reasoning: a story told because it feels good, not because it makes the world work better
The scale is now large enough to distort sovereign funding. "It's one of the reasons behind the bond freak out we just saw because it's pushing up longer term rates."
Lending to a data center with a hyperscaler's prime credit implicitly or explicitly behind it beats lending to what he calls a flawed credit like the US or the UK
The textbook worry was sovereigns crowding out private borrowers; right now it is running the other way, with sovereign fundraising challenged by data-center fundraising
Lenders Are Not Looking Inside the Building
"there could be hide and go seek competitions going on inside the data centers" — Kedrosky's joke about what the lenders he talks to actually care about, which is the credit on the other side
They look through the building to a twelve-year renewable lease and a prime credit, and treat it as a secure cash flow they prefer to a ten-year
That look-through drives more construction, more GPUs, and more high-bandwidth memory and NAND from suppliers like SK Hynix, none of it tied to what is happening inside
The technology working is a required ingredient, not a counterargument. Large language models are wildly useful, and if this were Beanie Baby-grade technology the conversation would not be happening
The analogy he reaches for is electrification and the build-outs of the 1920s: a genuinely great story, one of the most consequential technologies of the century, and "the capital allocation that happened in over sixty years in electrification happening in four years"
A study he cites puts GPU utilization far below the scarcity narrative: "The GPU usage is only sitting at around 35, 40%" in a large rental fleet he describes as an AWS of GPUs
That sits awkwardly beside the story about used A100 and H100 prices being bid up by scarcity
The reconciliation is behavior, not economics — "there's hoarding going on, colossal amounts of hoarding going on, double and triple ordering going on" because nobody wants to be caught short if some more parabolic moment arrives
Idle capacity at peak load is also latent supply: a potential flux of product into the market later on
Tokens Are Hyper-Deflationary, and That Is Toxic to the Frontier
"Tokens are the first hyper deflationary commodity in the history of modern economies." They are under pressure from the technology curve and from the capital flooding in at the same time
Kedrosky expects it to arrive as waves — as tokens brush against a sector, that sector gets a deflationary shock from a structurally deflationary force running at 70 to 80% a year
He sees no obvious mechanism, structural, technological or financial, by which that force ebbs
Consumers probably do very well out of it. The frontier model companies do not.
"This is among the most difficult things in capitalism is just to stand still, I need to grow 400% year over year." If price falls 80% a year, 400% unit growth is the minimum to hold revenue flat
That is before pleasing Wall Street, and before servicing fixed obligations that now sit on the balance sheet
His image for it is Wile E. Coyote over thin air, legs spinning, hoping not to look down
"OpenAI apparently in this last quarter did 18% quarter after quarter and that was deemed a disappointment."
Technology Companies Are Turning Into Utilities
A report Kedrosky had recently seen put "something like 15 to 18% of the investment grade marketplace is now data center related", a share that on his reading now exceeds financial services
The reversal from what technology used to be is total: "I'm old enough to remember that one of the prime attractions of technology after growth was that it had no debt" — pristine balance sheets, cash flow monsters
Maintenance capex makes it permanent rather than one-off, a point he credits to Michael Burry: the obligation to keep the data centers current runs into eternity
"turning orthodox technology companies into a kind of utility company, right, with similar obligations but wildly overvalued compared to an orthodox utility"
The open question is which way the rerating runs — utilities up, or the capex-heavy technology names down toward utility multiples. Kedrosky says the latter
The IPO Wave Has to Be Paid For by Selling the Winners
Faber raised a Kedrosky chart he called one of the most jaw-dropping he had seen this year: the late-nineties bubble had a supply-and-IPO component that has been missing for a long time, and a handful of pending listings — SpaceX and Anthropic among them — would outweigh the entire nineties IPO cohort
Kedrosky's correction was that it is bigger than that: "it's all post World War II combined"
Faber's own reading is that the direction of travel has flipped across the board, from buybacks toward share issuance and dilution, naming Oracle as an example
The number has already moved since he published it: "I said $4 trillion in new issuance from that narrow subset of companies alone. It's gonna be more than that now. It's gonna be more like $5.5 trillion"
On what retail investors get wrong about how large funds buy: "they tend to act as if these large funds are just sitting on a printing press of cash in the basement"
Cash is a drag, particularly for long-only funds, so an allocation to a new listing is funded by selling something else
What gets sold is predictable, and it is the good stuff. "So I tend to sell things that are liquid, that look like the thing I'm buying, and perversely, some of the things that are performing the best."
Liquid, so the sale does not move the price; overlapping, so exposure is not duplicated; and often the winners, so gains can be locked in before the quarterly letter
He modeled this in March or April and concluded the selling would start well ahead of the listings — weeks and months early, quietly, so as not to look big-footed — which put pressure on winners through April, May and June
He thinks that mechanism helped detonate a fund: "I think that played into the implosion of Situational Awareness", which was long the most liquid, best-performing names precisely as they became the funding source for the coming issuance
He credits a conversation with a derivatives desk at JPMorgan for the read, and allows that the fund's book and the pressure on it were separately visible
Nike, and Why Failure at High Valuations Is Overdetermined
Faber brought a Derek Thompson post: a chart of Nike down 75% from its peak, hundreds of comments blaming product, competitors, politics and macro, and "the PE ratio at the peak was like 70. And now it's 20." — which nobody mentioned
"At high valuations, failure is overdetermined in a statistical sense." There are so many low-probability ways to fail that the combination is not low-probability at all
The arithmetic is the argument: twenty independent failure modes at 5% each, and "there's greater than a 60% chance of failure in the period that the 5% applies"
So what looks unpredictable is actually highly predictable in aggregate, even though the specific trigger is not
The AI version has an unusually long list of candidate triggers — sovereigns, rogue AIs, and the one he has been pointing at lately, unprecedented cash inflows into Taiwanese and Chinese chip manufacturers pointing to a supply tsunami in early 2028
"this is a boom bust industry. So once you lock in supply, my friend, prices are going to zero." Fixed costs have to be covered, so the product goes out the door rather than sitting in inventory
How the unwind actually feels, in his words: "Kinda got pecked to death by ducks, man." Different things take it lower and lower until it finds a bottom, and ten years later people invent ad hoc explanations for a Micron price
The current this-time-is-different story he distrusts is the semiconductor super-cycle: that RAM never comes back, that GPUs are a duopoly or effectively a monopoly, and that this is the most consequential computing industry on earth
Vibe Chipping, and the End of Tribal Knowledge in Semis
A Wall Street Journal story about a new inference ASIC company — Etched — got the usual brash-young-founders treatment, which Kedrosky found funny because "the semiconductor industry historically as a venture capitalist has been a friggin' elephant's graveyard"
The interesting part was not the founders, it was the clock. Chip production runs from design specification through verification and test cycles to tape-out and a fab
"It was forty two days in the initial design verification stage versus what should have probably been six or seven months." And it worked on the first design pass, which he compares to marrying the first person you met walking down the street
The real story is that this is one of the first companies structurally using AI to design chips, and chip code sits squarely in what large language models are good at, for the same tight-grammar reasons he laid out earlier
On what comes next: instead of vibe coding, "we're gonna have kind of vibe chipping" — new designs arriving at fabs at rates nobody has seen
"the notion that some kind of tribal knowledge protects you as a chip manufacturer is going away"
He expects a flood of low-power ASICs doing on-chip large language models at the edge, inside things like security cameras, coming out of nowhere — and thinks that, perversely, is part of what pricks the bubble
The Models Have Flatlined and Converged
Asked whether the field ends with fifty indistinguishable winners or one Google emerging above the Ask Jeeveses, Kedrosky said the convergence has already happened
"we're kind of in the iPhone 7 moment when it comes to large language models" — the geeks still line up, and everyone else cannot tell this one from the two before it
He dates the maximum inflection in year-over-year model improvement to almost four years ago, 2022 and 2023
He discounts published benchmarks because the models ingest them, which he likens to having seen the SAT before sitting it, and looks at composite indices instead
"the year over year gains in the models themselves has essentially flatlined over the last six months" — down from 10 or 12% year-over-year changes in the composite indices to one or 2% at most
Harnesses are what disguise it. Things like Claude Code and Codex wrap the model and make it look like the underlying rate of change is faster than it is
His analogy is Julie Andrews in The Sound of Music: the models are the bratty kids, the harness is what gets them to sing
The second effect is convergence. The variance from best model to worst, once immense, has collapsed
"There is very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek or somewhere else in practical composite terms."
He runs occasional blind Pepsi-Coke tests behind a harness, and says nobody can tell them apart even though everybody believes they can
"the most successful frontier AI company will be the first one to stop pretending they can train new AI models" — because not spending that money is a gift
On a piece in The Economist he thinks was wrong: it argued chat models would break the TikTokification of tourism by making people's inputs more divergent
His rebuttal is the training data — "the median data nugget inside of a large language model is a 37-year-old male on Reddit", so homogeneous output should not surprise anyone
Why Americans Are Anomalously Sour on AI
Faber cited a stat he had seen that week: "75% of people said they wouldn't want a data center in their county, which was higher than nuclear"
Kedrosky's first read is lost agency. A data center is a hulking physical presence nobody asked for, and a stand-in for a broader sense of lost control over work, relationships and daily life
He links it to the same impulse behind American attitudes to vaccines: the objection is to being told, not necessarily to the thing itself
The comparison that interests him is international. The US is normally among the earliest and most aggressive adopters of technology, and yet "how the US is anomalously negative about AI compared to other G7 and OECD countries"
Broaden the data and the OECD as a whole is more negative than Sub-Saharan Africa and other developing regions, where AI reads as a leg up rather than a threat — a way to feed a family or get educated without going to college
He blames the industry's own messaging in part, naming Dario Amodei: tell people a large share of jobs will disappear and they will take you at your word
The uniquely American piece is health insurance: "any threat to employment is a threat to healthcare, and a threat to healthcare is a threat to your personal solvency", which is not the case in almost any other Western country
Tying health care that tightly to employment makes job loss existential, and AI has been sold largely on job losses
AI's Share of GDP Growth, and the Dog Who Thinks He Chased the Mailman
Asked which of his charts stand out, Kedrosky picked the one that got him thinking about AI as a system problem in the first place: AI as a large and growing share of GDP growth
He published it partly hoping to be corrected, and "that share grew and has persisted over the last six quarters since I first began writing about it"
The charts he values rotate your cognitive axes rather than showing that x is not as y as you thought — this one, because for perhaps the sixth time in Western history a nongovernmental force is large enough to move the tides of GDP growth
His illustration is his dog, who barks at the mailman, watches the mailman leave, and concludes he did it — the mailman would have left anyway
The policy version of that error is the point: if you believed tariffs were driving US growth "because it was more than 50% of US growth", your model of causality is broken, and you may take badly damaging actions on the strength of it
What actually drove the growth, on his account, was data centers and hyperscalers
Kedrosky's bottom line is that the AI story has stopped being a technology story and become a financing story — the models genuinely work, which is precisely the ingredient that lets capital detach from what is happening inside the buildings and grow to a scale that bends bond markets, GDP prints and the semiconductor cycle around it.
Products, Companies & Tools Mentioned
OpenAI (Its 18% quarter-on-quarter growth being read as a disappointment is Kedrosky's example of what the deflation treadmill does to expectations)
Anthropic, Qwen and DeepSeek (The frontier models he says are now practically indistinguishable in composite terms)
Claude Code and Codex (The "harnesses" wrapped around models that make the underlying rate of improvement look faster than it is)
Nvidia A100s and H100s (The used GPUs whose prices are supposedly bid up by scarcity, against utilization he puts at "around 35, 40%")
SK Hynix (High-bandwidth memory and NAND — the downstream purchases the data-center flywheel keeps pulling in)
SpaceX and Anthropic (The pending listings behind the issuance wave Kedrosky says beats every post-war IPO cohort combined)
Etched (The inference ASIC startup that ran initial design verification in 42 days, and his evidence that AI is about to collapse the chip industry's barriers)
Micron (His stand-in for the semiconductor boom-bust that people will explain after the fact with ad hoc stories)
Nike (Faber's example of a stock down 75% where the commentary blamed everything except a multiple that went from about 70 to 20)
Oracle (Faber's example of the shift from buybacks toward issuance and dilution)
JPMorgan (A derivatives desk there informed his read on what drained liquidity out of the winners ahead of the IPO wave)
Situational Awareness (The fund whose implosion he thinks was partly caused by being long exactly the names that had to be sold to fund new issuance)
Cambria Investment Management (Faber's firm — he is its co-founder and chief investment officer)
SK Ventures (Kedrosky's firm, and the source of the charts Faber subscribes to)
Books & Resources Mentioned
PaulKedrosky.com (Where Kedrosky told listeners to find his writing and charts)
Albert Bartlett on the exponential function (The physicist's line that Kedrosky uses as the spine of his whole argument about scale)
The Economist on chat models and the TikTokification of tourism (A piece he calls completely wrong but still interesting — his counter is the homogeneity of the training data)
The Wall Street Journal on a new inference ASIC company (The brash-young-founders story that led him to the 42-day design cycle)
Michael Burry on maintenance capex (Cited for the point that data-center capex is a permanent obligation, not a one-time build)
Barry Ritholtz's playbook on behavioral investing (Credited for the framing that people panic and chase at exactly the wrong moments, bubble after bubble)
Owen Lamont's line on valuations (Quoted by Faber from an earlier episode: prices that cannot be justified by rational forecasts then double)
Get the latest market chatter and takes as they happen:
X | Threads | Instagram | YouTube | TikTok | Facebook