AI Is an Industrial Bubble, Not a Financial One
The two-bubble distinction
The framework comes from Jeff Bezos via Tim Ferriss, relayed in conversation with Bill Gurley:
| Financial bubble | Industrial bubble | |
|---|---|---|
| Archetype | 2008 Global Financial Crisis | Late-1990s dot-com boom |
| What’s being traded | Derivative claims on other claims (CDOs, synthetic exposures) | Equity in companies building new technology |
| Durable value created | None; the system was betting on a fiction | Significant: internet infrastructure, new business models, durable companies |
| What survives | Very little; the premises were empty | The underlying technology, the operating lessons, a handful of winners |
| Aftermath | Recession, regulatory tightening | Survivors become the next decade’s dominant firms |
Both patterns share the same speculative psychology and the same narrative overheating. They differ in whether there is something real underneath. That difference determines what the aftermath looks like.
Why this framework matters for AI positioning
The common objection to “AI is a bubble” is that the technology is clearly transformational. The common objection to “AI is real” is that valuations have decoupled from any plausible near-term revenue. Both objections are correct. The two-bubble distinction resolves them: a technology wave is real precisely because it invites the kind of speculative capital that produces bubble behavior. Carlota Perez (in Technological Revolutions and Financial Capital, 2002) shows that every major technology revolution in the past 250 years followed this pattern. Canal mania, railway mania, electricity, automobiles, dot-com. The speculation and the real value creation have been systematically correlated, not opposed.
For investors: bet on surviving the correction with exposure to the companies that will still be standing afterward. Those companies look expensive during the bubble phase (everyone bidding them up) and cheap during the correction (bubble capital fleeing indiscriminately). The discipline is holding through the correction with a differentiated view of which companies are real.
The dot-com analogy
Most “AI is a bubble” arguments reach for the dot-com comparison. Right comparison, wrong reason.
People who use the comparison to predict an imminent crash miss what actually happened. The dot-com crash destroyed speculative valuations but left the underlying technology, the companies that dominated the next two decades, and the investment theses that led to them intact. Amazon, Google, Netflix, PayPal, Cisco, Adobe all survived or benefited from the dot-com era. The crash made them cheaper, not wrong.
The correct version of the analogy: AI in 2025-2026 probably faces a dot-com-style correction. The correction will be painful and will destroy most pure-speculation vehicles. The companies building real things will emerge with dominant positions. The question investors should ask is “which companies survive the correction.”
What makes this one different
Two differences from the dot-com era, both cutting against a clean pattern match:
1. Capital intensity is much higher. Dot-com companies could fail cheaply on generic servers and software. AI companies require hundreds of millions to billions in compute commitments to train frontier models. When the correction comes, bankruptcy costs will be larger and the secondary effects on compute supply, power contracts, and chip demand will be structural.
2. Circular financing at the top of the stack. Microsoft invests in OpenAI, OpenAI commits to buying services from Microsoft. Nvidia invests in AI companies, those companies commit to buying Nvidia compute. The dot-com era had nothing comparable, and the pattern makes AI supply-chain revenue figures less reliable as indicators of end-user demand. The correction, when it comes, may unwind more of the revenue base than observers currently assume.
Neither breaks the industrial-bubble framing; they change the shape of the correction. More circular financing and higher capital intensity probably mean a longer, more painful correction, but one that still leaves durable winners standing.
Practical posture
- Accept that some current valuations are bubble valuations. Trying to justify them on DCF is a mistake.
- Accept that the underlying technology is genuinely transformational. Dismissing it as hype is also a mistake.
- Look for businesses where the fundamental demand does not depend on the circular financing. Real customer revenue, measurable unit economics, clear paths to cash flow.
- Expect a correction. Time it based on the financing structure unwinding (circular deals collapsing, SPVs defaulting, retail flows reversing) rather than on specific valuations.
- Hold through the correction if your thesis is right. The companies that survive an industrial-bubble correction typically end up dominant for the next decade.
2026-08 Update — Which asset actually survives
The framework says an industrial bubble leaves durable value behind. It does not say which asset. Ben Thompson names one (Ben Thompson - What Happens When the AI Boom Runs Out of Money, August 2026): the residue this time is most likely new power infrastructure, not the compute.
The reasoning is depreciation schedules. GPUs are roughly 70% of data-centre capex and are largely worthless after three years, while buildings, substations, transmission, and generation last decades. So the durable layer is the one nobody is excited about.
That has a specific historical shape. Railroads ran out of money in the 1870s and kept operating, opening the American West and contributing to GDP for a century. Fibre from the dot-com build was lit years later by companies that did not pay for it. In each case the surviving asset was the long-lived physical layer, and the investors who funded it were not the ones who captured the return.
Thompson notes the echo directly: Berkshire’s railroad profits are part of what funds its Google stake today.
The investing consequence, if this holds: exposure to the durable layer is not the same as exposure to the bubble. A financial blowup that leaves power infrastructure behind is very bad for whoever financed the compute and quite good for whoever ends up owning generation and transmission. That is a different position from “AI is real, so own AI.”
The tell to watch: whether power or compute is the binding constraint as capacity lands. Thompson notes the US has brought online more electricity than he expected, which weakens the power-scarcity case and extends the runway for competing chips — see AI is power and compute constrained.
Related Notes
- Circular Financing as a Bubble Signal in AI
- AI and Investing Thesis
- AI Stack Value Accrual - Chip, Infra, Intelligence, App
- Grow 10 or Earn 40 - Two Paths for Mature Software Companies
- Investing in the AI Era - Bill Gurley (source)
- Risk Migration Down the Capital Stack — the financing-structure question an industrial bubble leaves open: durable assets survive, but the sequence of who funded them decides who eats the loss