The AI Supercycle: is it a bubble? Why railways and dot-com already answered

The question everyone asks about AI is "is it a bubble, yes or no?" That is the wrong question, and the previous nine chapters explain why. Railways were a real, world-changing technology and a savage bubble. The internet was a real, world-changing technology and a savage bubble. $AMZN fell about 95% in the dot-com crash and went on to become one of the largest companies on earth. "Real" and "bubble" are not opposites. They are the default combination.

So the useful question is not whether the AI supercycle is a bubble. It is which part is the durable build-out and which part is the 1720-style financing scheme wearing it as a costume. That is a much harder question, and it is the one this whole series has been training for.

The capex is real, and that is not the reassuring part

The four largest US hyperscalers are spending on the order of several hundred billion dollars a year, combined, building AI data centers, and the figure has been revised up almost every quarter. Power, not chips, is increasingly the binding constraint. This is the railway parallel made literal: enormous, sunk, physical capital being laid down on the belief that demand will arrive to fill it.

That capex is the strongest argument for the build-out being real. It is also, read the other way, the largest single bet in corporate history that the return on this infrastructure will justify its cost, and nobody can yet point to the revenue that pays it back at this scale. Railways got built. Most of the first railway investors still lost their money. Fiber got laid in 1999. Most of the companies that laid it went bankrupt before the broadband era they enabled. Real infrastructure and ruined investors are not a contradiction in this series. They are the rule.

The circular financing is the 1720 tell

Here is the part that should make a reader of the South Sea Bubble sit up. The dominant supplier of AI compute, $NVDA, has become an investor in, and a financing backstop for, several of its own largest customers. Chip vendors take stakes in the cloud providers who buy the chips. Model labs sign compute commitments with cloud providers who are funded, in part, by the chip vendor. Capital flows out to a customer, comes back as a purchase order, and is booked as demand.

In 1720 the South Sea Company lent people money to buy South Sea shares, and the rising share price was treated as proof the scheme worked. The mechanic is identical: when the seller finances the buyer, revenue and the stock price start referencing each other instead of an outside source of cash. That is reflexivity, the third recurring mechanic, and it is the single hardest thing to see from inside the story, because every step looks like a normal commercial deal.

Why "is it a bubble" is the wrong question

Run the dot-com tape forward and you get the answer the binary framing misses. In 2000, "the internet is a fad" and "the internet will eat the economy" were both being said, and both were wrong in the way that matters. The internet ate the economy and the Nasdaq still fell about 78%. The people who were right about the technology and wrong about the price lost just as much money as the people who were wrong about everything.

AI can be the most important technology of the decade and the AI trade can still be a bubble, at the same time, for the same reason railways and the internet were. The build-out can be real, the survivors can be enormous, and the median investor can still be carried out, because the median investor cannot tell the survivor from the casualty in advance. Picking the AMZN out of a field that also contains Pets.com is the entire game, and "I believe in AI" does nothing to help you do it. Conviction in the technology is not a position in the equity.

The four mechanics, applied to 2026

Strip away the specifics and the AI supercycle reruns the same four-part template every chapter in this series has mapped:

1. A genuinely new and scarce thing. Not a virus-streaked tulip but AI compute itself: leading-edge GPUs, the foundries that make them, and the power and land to run them. Genuinely scarce, genuinely new, genuinely valuable.
2. A new instrument that adds leverage or removes friction. Vendor financing, special-purpose vehicles raising debt against data centers, circular supplier-to-customer capital, and the public-market concentration that lets a handful of names carry an index. The leverage is in the financing structure, not yet in retail margin accounts.
3. A reflexive circle where the price is the story. "AI is the future, so own the picks and shovels, which makes the picks and shovels rise, which proves AI is the future." The circular financing hard-wires this loop into the actual revenue line.
4. A top that needs no catalyst. As in 1637 and every chapter since, the unwind does not require a scandal or a failed harvest. It requires only that the marginal buyer, the next hyperscaler willing to raise the capex number again, simply does not show up one quarter.

You will notice none of those four tells you it is a bubble. They tell you it has the shape of one, which is a different and more useful claim. Tulips had the shape. So did railways, which were also real. Shape is what you can observe before the fact. Verdict is only available after.

What history actually tells you to do

The mistake every chapter of this series documents is the same: people tried to answer "is this a bubble" with an opinion about the technology, and the technology was almost never the question. The flowers were real. The railways were real. The internet was real. The thing that broke was always the price and the financing, and the only people who navigated it watched the capital flow rather than argued about the narrative.

That is the entire reason the QuantAbundancia bubble map exists, and the reason this site is organised the way it is. We do not try to tell you whether AI is a bubble, because that question has no answer until afterward and no edge even then. We map the AI supercycle by capital flow, group the names into bubbles by how they actually move together, and validate the groupings against 252-day correlations, so the unit of analysis is "where is the money going and is it still going there", not "do I believe the story". Watching when a cluster breaks correlation, when the marginal buyer leaves, is the closest thing to an edge this 400-year history offers. Bubble-level shifts and rule-based alerts when a cluster breaks are part of /pro.

This is the final chapter of A History of Market Bubbles. It ends where the site begins: with the live map of the one that is happening right now. You have read what the last nine looked like from the inside. The point of all of it is to make the tenth one legible while you are still standing in it.

The live version of this story: the QuantAbundancia bubble map tracks the AI supercycle by capital flow, validated against 252-day correlations. Track the dominant name live at /stocks/nvda.

Start the series from the beginning: A History of Market Bubbles.

QuantAbundancia is educational research. Nothing here is investment advice. See /disclosures.