One Man's Uprising: Why Things at the Top Don't Need a Reason to Fall

One man raised a revolt, and the seven ancestral temples fell; the emperor died at another’s hands, a laughingstock to the world. Why?

—— Jia Yi, On the Faults of Qin (Western Han, c. 2nd century BCE; author’s translation)

Qin built the Great Wall, confiscated the weapons of the realm, and believed it had founded a dynasty to last ten thousand generations. It was brought down by a conscript. Jia Yi’s point was never that the conscript was formidable. His point was how small the hand can be, once a thing has been stacked that high. Episode 648 of The Meb Faber Show (28 August 2026), with guest Paul Kedrosky, spends an hour on AI infrastructure and ends up in the same place: asking “what will pop it” is the wrong question, because at that altitude there are too many things that could.

What the episode is about

Paul Kedrosky is a venture investor and a fellow at MIT’s digital economy institute, and was a sell-side analyst in an earlier life. His position in this episode is clear but unusual: he does not think large language models are useless. His words are that they are “wildly useful.” He thinks the problem is that the money now moving has almost nothing to do with that usefulness.

The organizing claim is this. Go back through the largest financial paroxysms of the last two hundred years and they share a handful of ingredients: loose credit, a genuine and compelling technology story, a real estate component, and a policy angle. Canals had some of them. Railroads had some. Rural electrification in the 1920s had some. This is the first moment that sits at the intersection of all of them.

And because all the ingredients arrived at once, everyone is touching one side of the animal. Credit people approach it through credit, real estate people through real estate, the enthusiasts online through five prompts that will change your life. Each hand is reporting honestly. None of them is describing the elephant.

The main points

1. The greatest failing of our species is an inability to understand the exponential function. He borrows the line from physicist Albert Bartlett, and the old pond metaphor that goes with it: algae doubling daily covers the whole pond on the last day, which means the day before it was half covered and the day before that a quarter. Bulls and bears misread exponentials equally. And AI does not carry one exponential curve; it carries two pulling against each other — adoption rising, price falling.

2. We are extrapolating the future from the least representative possible starting point. The first place large language models really worked was software, and software has three rare properties: a strict grammar (making a subroutine more aesthetically pleasing does not give you prettier software, it gives you broken software), fast error feedback, and — crucially — it is expansive. One prompt goes in, a million lines come out. Almost every white-collar use runs the opposite direction: it is compressive. Forty pages of research go in, five bullets come out, and what you actually want to know is how those five differ from yesterday’s. Hardly any other domain in economic life has all three properties, “and yet that’s the domain from which we’re extrapolating our futures.” He closes with the venture cliché that fits perfectly: be very careful who your first customers are, because early adopters are like nobody else.

3. The nature of the financing flipped in the first half of 2026 — and the counterargument flipped 180 degrees with it. For years the standard rebuttal was: this is their own cash flow, these are smart companies, who are you to tell them how to spend it? By mid-2026, more than half of data center financing comes from outside — asset-backed securities, private credit, sovereigns, a constellation of special purpose vehicles. And the same people now say: the money is coming from outside investors, they’re really smart, why do you care? He points at that reversal as the finding itself. Two opposite facts marshalled to support one conclusion is motivated reasoning: you are telling the story because it feels good, not because it makes the world work better.

4. The lenders do not care what happens inside the data center. His running joke is that from the standpoint of most lenders he talks to, there could be hide-and-seek competitions going on inside and they wouldn’t care. They look straight through the building to a prime credit on the other side, a twelve-year renewable lease, a secured cash flow — better than holding a ten-year Treasury. That is the Minsky condition: when something becomes financialized and divorced from the underlying application, pay attention. The flow is large enough to have inverted a textbook relationship. Sovereign issuance is supposed to crowd out private borrowers; right now data center fundraising is crowding out sovereigns, which is part of why long rates got restless. He cites a JPMorgan figure: roughly 15% to 18% of the investment-grade market is now data-center related. If it were a sector, it would be larger than financial services within that market.

5. Tokens are the first hyper-deflationary commodity in the history of modern economies, falling something like 70 to 80% a year. He does the arithmetic out loud. If price falls 80% a year, then simply to stand still in revenue you need roughly 400% unit growth — before pleasing Wall Street, and before servicing the debt that now funds more than half of the buildout. He describes the frontier labs as Wile E. Coyote out past the cliff edge, legs still spinning, thinking: please don’t let me fall, I’m growing this fast just to stay level. One number in the same stretch is worth holding onto: utilization in a large GPU pool sits around 35 to 40%, at the same time as stories circulate about used accelerator prices rising on scarcity. Both can only be true under one explanation — enormous hoarding and double and triple ordering. What is hoarded today is supply that lands on the market later.

6. Models have converged, and the harnesses around them are covering for it. Year-over-year gains on composite measures have gone from 10 to 12% down to one or two points. The promoted benchmarks mean progressively less, because models ingest the benchmarks — like seeing the exam before you sit it. More importantly, variance across models has collapsed. He occasionally runs a Coke-Pepsi blind test, hiding models behind one harness and asking people to tell them apart; nobody can, and everybody thinks they can. For the coding harnesses wrapped around the models, his image is The Sound of Music: the models are the bratty kids, the harness is the governess, and she gets them singing beautifully. It looks like the children improved. Someone organized them. He also offers a sharp prediction: the most successful frontier company will be the first one to stop pretending it needs to train new models. And a line you can use tomorrow: the median data nugget inside a large language model is a 37-year-old man on Reddit. Work out what he’d say, and you have roughly predicted the model.

7. The hostility toward data centers has an American-only ingredient. The host cites a survey where about 75% of people don’t want one in their county — higher than for nuclear. Kedrosky reads it in two layers. The first is lost agency: a great hulking presence nobody asked for, standing in for a broader sense of having less and less say over work, relationships, and daily life. The second is specific to the United States: a threat to employment is a threat to health coverage, and a threat to health coverage is a threat to personal solvency. And the industry has spent two years marketing AI on the basis of job losses. Tell people that and they take you at your word. The revealing contrast is that in sub-Saharan Africa and other developing regions, attitudes are markedly more positive — there it reads as a ladder up, not as something being taken away.

Going further

1. “This position is up a lot. Should I get out?”

The question is hard because you want an answer about whether something bad will happen, and markets don’t produce those. This episode offers a conversion I think is the single most portable idea in it.

At high valuations, Kedrosky says, failure is overdetermined. Suppose there are twenty paths down, each individually unlikely — say 5%. Multiply out the chance that none of them happens and you land above a 60% chance that one does. So a state that looks unpredictable and not particularly dangerous is in fact highly predictable. What’s predictable is the sum of the probabilities, not any one path.

The host supplies a lovely piece of supporting evidence. Derek Thompson posted a chart of Nike down 75% from its peak and asked what happened. Hundreds of replies: product, competitors, the political controversy, macro. Not one mentioned that the P/E at the peak was about 70 and is now about 20.

There is a practical reading order buried in that. When many competing stories can each explain the same decline, look at the multiple first. If the change in multiple accounts for most of the move, the stories were mostly written afterward. The reverse also holds: if the multiple barely moved and earnings did the falling, then it really is about the business.

The episode has no view on whether you should sell, and shouldn’t. But it rewrites the question into one you can actually compute: how many paths am I currently exposed to, and what do they sum to?

2. “Demand for AI is obviously real, so buying the related names should work.”

The most counterintuitive move in the episode is holding two things at once: the technology genuinely works, and the money has already decoupled from whether it works.

To hold both, split one question into two. The first is does the thing work — his answer is yes, emphatically, and he notes that this is a required condition for moments like this. If it were nonsense collectibles, none of this capital would have shown up. The second is who is paying, and on what terms — and that answer changed this year. Moving from internal cash flow to external debt is not merely a change of funding source; it changes the character of the thing. Cash flow can slow down. Debt cannot.

Split that way, the contradictory headlines stop contradicting. Utilization at 35% alongside rising used-card prices isn’t anyone lying; it’s hoarding. Data centers approaching a fifth of the investment-grade market doesn’t mean the bubble bursts tomorrow; it means the sector’s balance sheet has gone from debt-free cash monster to something closer to a utility with perpetual maintenance capex — while still being valued as technology. Which sets up his sharpest fork: do utilities get rerated upward toward tech, or do the tech names most tied to the capex wave get rerated downward toward utilities? He takes the latter.

There is a subtler layer too. He describes an inference-chip startup that compressed design verification from six or seven months to forty-two days, precisely because chip design code sits in the sweet spot of what these models do well. He jokes that after vibe coding we get vibe chipping. And the irony writes itself: if tribal-knowledge moats really are dissolving, then the thing dissolving the belief that chips are a scarce oligopoly is AI itself.

So between “demand is real” and “the layer I bought captures it” sit three questions: will this layer be drowned by its own success, who is funding it, and is its moat being dissolved by the same wave that created the demand.

3. “I didn’t do anything wrong. Why did my holdings fall first?”

The episode contains a structural argument that ordinary investors rarely encounter and are frequently hurt by.

The setup is new issuance. He calculates that if just a handful of the largest private companies go public, the issuance exceeds all IPOs since World War II combined, inflation-adjusted. Four trillion dollars, and he now thinks it looks more like five and a half.

The inference comes next. Retail investors tend to act as if large funds keep a printing press in the basement — as if buying an allocation means calling down the hallway for money. It doesn’t work that way. Cash is a drag for a long-only fund, so buying something new means selling something old. And what gets sold follows a pattern: the most liquid (to avoid price impact), the things that overlap with what’s being bought (to avoid doubling an exposure), and — least intuitively — some of the best performers (locking in gains reads well in the quarterly letter).

Timing matters most. You cannot free that money up the night before, like borrowing from your dad to buy skis. It has to happen weeks or months ahead, and quietly. So the pressure from a flood of new issuance arrives on today’s winners long before the flood does. He says he modeled this early in the year, and that some of the most liquid, best-performing names coming under pressure in the spring was connected to it.

Two lessons for the rest of us, pointing in opposite directions, and you need both. The consoling one: your position may be falling because somebody has to sell it to pay for something else, which says nothing about whether you were right. The warning: that explanation is so comfortable it can rationalize any decline. So it needs a falsifiable check. Forced selling is broad, simultaneous, and indifferent to fundamentals. If comparable, similarly liquid names are all soft over the same window, that’s a funding rotation. If yours is the only one that’s soft while its peers hold up, nobody is raising cash. Something is wrong with your thing.

Sources worth your time

• The Meb Faber Show, episode 648, with Paul Kedrosky (28 August 2026), hosted by Cambria co-founder Meb Faber.
• Paul Kedrosky’s own site, paulkedrosky.com, which is where he points listeners at the end. Most of the statistics in the episode come from the chart series he publishes there.
• Albert Bartlett’s public lectures on the exponential function. The upstream idea for the whole episode is his: the greatest failing of the human species is an inability to understand exponential growth. Understand that before anything else here.
• The JPMorgan work on data-center-related share of the investment-grade market, cited in the episode at 15 to 18%. The value isn’t the precision. It’s that the number forces you to drop the old assumption that technology companies carry no debt.

The One Thing to Take Away

At the top, falling doesn’t require a reason. It only requires probabilities adding up.

We assess risk almost entirely by imagining one specific way things go wrong and then judging how plausible that story is. If we can’t picture anything frightening enough, we conclude it’s probably fine. That method breaks at altitude — because what characterizes altitude isn’t that any one path is especially dangerous, it’s that there are so many paths. Twenty at 5% each is above 60%, and afterward you will only ever see the one that happened, and someone will give it a very persuasive name.

Something to do today, no stocks required: pick the arrangement you currently depend on most, the one whose failure would genuinely hurt. A source of income. A machine you use daily that will eventually break. A person every decision has to pass through. The last remaining working channel of communication in a relationship. Write down ten ways it could break, and give each one a probability you consider low — 2%, or 5%. Then multiply out the chance that none of them happens: 0.95 to the tenth power is 0.60. Ten paths you each dismissed as unlikely add up to a 40% chance something breaks.

Then don’t stop at the sigh. Pick the one that is cheapest to fix and fix it this week — one more backup, one more person who knows the password, one more phone number obtained. It won’t make you safe. But it will be the first time you know how high you’re standing.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.