What Counts as a Bubble

The usefulness was never the point

Kedrosky starts with a definition and holds it the whole hour:

"A bubble is whenever asset prices detach from any semblance of economic reality."Kedrosky · 2:02

His evidence is history. The West has built out a general-purpose technology eight or nine times: canals, railroads, electrification, the interstates, fiber. Every one overbuilt. Railroads were enormously useful, and still, he says, something like half the track miles ever laid turned out to be useless. 3:19

The mechanism is a race. Each company fears that if it slows down, a competitor or its own shareholders will punish it, which is why calls to "pace the frontier" go nowhere. 6:26

"As long as you have a supply of capital and a race dynamic, you're guaranteed overcapacity."Kedrosky · 7:12

What makes this one dangerous is where the bet sits. Half of household net worth is in equities, most of the index is tech, most of tech is AI. If you hold an S&P 500 fund, "you are very, very long AI." He calls it a one-factor world. 4:02

Two tests, and today both pass

Azhar asks two questions of any technology boom. Are customers finding it useful, and paying more for it over time? And is the financing sound? Bubbles that burst almost always break in the financing, when short-term debt can't be refinanced and the capital goes away.

"There is not an overbuild yet. There is vast, vast undersupply of the compute needed for AI."Azhar · 8:32

He goes a layer below the tokens, to the asset actually being bought: data centers, about 65% of whose cost is the chip. Those chips are earning back their price in 18 to 24 months, at what he calls Uber surge pricing. 11:34

The obvious worry is circularity: about half that rent comes from OpenAI and Anthropic. Held steady, that would look like someone "paying themselves around the back door." But the share is falling as real customers arrive. His verdict: a productive boom, "with the hints of maybe getting, forgive my French, a little bit stinky." 12:31

Volume is not value

Kedrosky answers with an old joke: a child digging through a pile of manure under the Christmas tree, sure there must be a pony in there somewhere. Talk of enormous token counts, he says, confuses scale with value.

"What matters is tokens times price."Kedrosky · 13:36

And the price is falling, roughly 13-fold a year, performance-adjusted. The labs reply that their newest models hold their price. Kedrosky's answer is that most people don't need the newest models.

The deeper error, in his venture-capital adage, is to "never confuse your early customers for what the real market will look like." 14:31 The early customers were coders, by his count around 80% of the big labs' usage. Coders are unusual twice over: they want the frontier, and their use is expansive, seven words in, a million lines of code out. Most people's use is one-to-one or compressive: twenty-five emails in, one sentence out. 15:29

The labs and the chips are two questions

Azhar says Kedrosky's answer had "some sleight of hand": it was about the frontier labs, while his own was about the data centers. If enterprises move to cheaper open-weight models running on the same hardware, that is just the demand moving.

"It may hurt Anthropic's margins, but it still monetizes the underlying capital expenditure."Azhar · 19:16

Ezra Klein presses both on a definition. Azhar gives a classic one: investment in the asset falls by half, and the relevant market crashes by half or more. Anything short of that is pain, not a pop.

"I think we may end up in a world where there are grindingly low and embarrassing returns for credit funds and for hyperscalers … But that degree of pain is not a bubble bursting."Azhar · 21:03

Kedrosky: "I disagree on almost every point." The definitions are arbitrary; the fragility is in plain sight. He points to three IPOs, OpenAI, Anthropic and SpaceX, together larger than all post-war IPOs combined, and a cyclically adjusted price-earnings ratio at the level it reached before 1929. 21:48

"So I like to say the failure here is overdetermined."Kedrosky · 23:04

The IPOs themselves are part of his case. They used to raise money for growth. These, he says, are insiders cashing out: "So these are exit events." 26:32 He adds customer concentration: perhaps eight companies making up 80% of a lab's revenue, mostly writing code. 30:30

Azhar turns that into something testable. Eight customers at 80% is 10% each, and a customer that size has to be disclosed in the filing, so the S-1 will settle it, "a very, very falsifiable prediction." 31:18

From cash flow to debt

Apple declined to build a frontier model and has outperformed the companies that did. Azhar reads it as culture: Apple "cares about detail and they care about control," and large language models are not about control. 33:27 Kedrosky reads luck: Apple "mostly screwed it up and got lucky," a free rider on everyone else's spending. 34:00 Azhar's name for that is complementary business, the normal way a general-purpose technology spreads. 35:36

Klein then asks the reassuring question: these companies can afford to lose hundreds of billions, so aren't the rest of us protected? Kedrosky: "No, we're not." A year ago the answer was that spending came from cash flow. Now it is moving to debt. 37:35

"What Wall Street is very, very good at is making it less pristine."Kedrosky · 38:15

Some of it sits off the balance sheet, in special-purpose vehicles. And an industry once prized for growth without debt has become "the largest issuer of investment-grade bonds in the United States." 39:16

Azhar shares part of the worry. In 2025 he set a threshold for external financing of about 30%, a level the industry is now nearing. And the crowding out of homes and factories is real; he compares it to the Netherlands' "Dutch disease," and doubts market forces alone will fix it. 41:19

If it pays off, what does work look like?

Azhar reports what he hears from the dozens of executives he talks to each year:

"We're making progress with AI. It's harder than we thought."Azhar · 45:15

Most of the gain, for now, arrives as slow attrition: more revenue per employee, headcount falling where growth lags. There are almost no AI-native companies on Main Street yet, just as most of America electrified fifteen or twenty years after Ford. His horizon is five, seven, ten years. 48:26

Kedrosky thinks that horizon is the problem:

"This is the fastest adopting technology in history, and its fundamental function is to reprice cognitive labor."Kedrosky · 49:04

The promise that technology always creates more jobs than it destroys has become, he says, a mantra. "I think the disruption will vastly outpace our ability to create new positions." 50:12

The risk no balance sheet shows

Klein raises a different way for it to deflate: people may simply refuse it, in classrooms, in hospitals, in their towns. Kedrosky says the companies have been "conducting unlicensed drug trials on the public," and that regulation will come after some terrible event. 52:02

His argument for inevitability is arithmetic. No single disaster is likely, but there are many of them:

Kedrosky's figures drift low and then high: six risks give about 26%, not "like 20%", and twelve give 46%, not "more than 50%"; it takes fourteen to cross half. The point survives the arithmetic. It also assumes the risks are independent, and risks that share a cause compound less. 53:12

Azhar agrees self-regulation won't work, and reads the protests differently:

"That is a democratic expression, right?"Azhar · 54:56

Kedrosky ends somewhere neither balance sheet reaches. Klein's agents phoning restaurants for reservations, a surgeon's ten-thousand-word emails from patients, the dollar a month Kedrosky pays to screen out robocalls: costs pushed into the system, and, he says, counted as GDP.

"In a weird way, societies require friction … almost a denial of service attack on modernity."Kedrosky · 1:00:32

Each names his own test

Klein closes by asking each what would move him. The answers return to what each was counting from the start.

Neither test is about the technology. One is about whether people pay for it, the other about whether the economy shows it. Short of those numbers, Kedrosky says, this looks "more like the 1930s" than the 2000s. 1:03:55