There is a 250-year-old framework that turns "is AI a bubble?" into a better question. It is genuinely useful. It might also not be real. Both of those things matter, and almost nobody holds them at the same time.
Sometime in the late 1990s, telecom companies buried an astonishing amount of fiber-optic cable under the world's streets and oceans, convinced the internet would need it any minute. Then the dot-com bubble burst, and the bill came due on an embarrassing fact: somewhere between 85% and 95% of that fiber was never lit. It sat there in the dark, "dark fibre," the industry called it, a monument to mania, glass laid for traffic that didn't exist. Investors were wiped out. Companies like Global Crossing went bankrupt with billions in cable in the ground.
And then, quietly, over the next decade, that "wasted" glass became the physical body of the modern internet. The broadband, the streaming, the cloud, the video call you took this morning: they run, in large part, on capacity that was overbuilt by doomed companies during a bubble everyone agreed was insane. The people who paid for the fiber lost everything. The fiber won.
Hold that story, because right now five American companies are preparing to spend somewhere north of $7 trillion building AI data centers this decade, and the question dominating every earnings call and dinner party, "is AI a bubble?", turns out to be the wrong question. There is a 250-year-old framework that asks a better one. It is genuinely useful. It might also not be real. Both of those things matter, and almost nobody holds them at the same time.
The framework belongs to Carlota Perez, whose 2002 book Technological Revolutions and Financial Capital reads two centuries of economic history as a sequence of five "great surges." Each runs roughly fifty to sixty years. Each is organized around a new techno-economic paradigm and a defining cheap input that gets lavished on everything. And each, Perez argues, has the same internal shape.
The five, by her dating: the Industrial Revolution (from 1771, cotton, canals); Steam and Railways (1829); Steel, Electricity and Heavy Engineering (1875); Oil, Automobiles and Mass Production (1908); and Information and Telecommunications (1971, the year Intel shipped the 4004 microprocessor). Cheap iron, then cheap steam, then cheap steel and power, then cheap oil, then cheap computation: each paradigm has its near-free input that reorganizes what's worth doing.
The shape inside each surge is the part worth tattooing on your forearm. It opens with installation: financial capital, impatient, greedy, gloriously irrational, pours into the new technology, builds its infrastructure far faster than any sober plan would, and inflates a bubble. The bubble bursts at a turning point, a crash that ruins the speculators and forces institutions to recompose. And then comes deployment: production capital takes over, the technology spreads into every corner of ordinary life, and the gains broaden into what Perez calls a golden age.
Here is Perez's genuinely subversive claim, the one that separates her from the gloom of Kondratiev and Schumpeter before her: the bubble is not a flaw in the system. It is the system's funding mechanism. Rational capital would never overbuild infrastructure that far ahead of demand. Only a mania will do that, and a mania is exactly what's required to lay more rail, or fiber, or compute, than the present can justify, so that the future inherits it cheap. The crash transfers the infrastructure from the people who financed it to the people who will use it. Investors lose; society keeps the rails.
This isn't a theory floating free of evidence; it has two clean precedents, and their numbers are the concrete heart of the whole argument.
The first is Britain's Railway Mania of the 1840s. The economic historian Andrew Odlyzko, who has studied it closely, calls it "a giant, wildly speculative, and successful investment mania," successful not for the investors, most of whom were ruined, but for the country. The startling figure: the rail lines laid down during those frenzied bubble years came to represent something like 90% of the total length of Britain's eventual railway network. A generation of speculators was wiped out building track that then carried Britain for the next 150 years. The mania's victims and the mania's beneficiaries were almost entirely different people, separated by the crash.
The second is the dark fiber we started with: the 1990s telecom build-out, 85 to 95% unlit after the bust, that became the substrate of the 2000s internet economy. Same structure, one full wave later. And the mechanism is identical: a bubble is a machine for funding large-scale, parallel, reckless experimentation in physical infrastructure. Most of the bets fail. The infrastructure remains.
You can feel the AI parallel writing itself, and that is precisely the moment to slow down. Because the seductive version of this essay, the one written several hundred times already in 2024 and 2025, by Stratechery and KKR and half the substacks on the internet, stops right here, says "the AI data-center build-out is the new dark fiber, the crash will fund the deployment, relax," and goes home. That take is not wrong, exactly. It's just doing the one thing this framework most tempts you to do: mistaking a story for a law.
Drop AI onto Perez's map and you get two honest readings, not one.
Reading A: AI is the irruption of a brand-new sixth wave. On this view the surge began around 2020 to 2022 with GPT-3 and ChatGPT; the paradigm is foundation models and agentic systems; and the defining cheap input is cognition itself. The way Moore's Law made computation almost free, AI is driving the cost of a unit of cognitive work toward zero. The enormous data-center capex is then a textbook installation-phase over-investment, financial capital racing to build the carrier infrastructure of a new age.
Reading B: AI is the late fifth wave. The dot-com crash of 2000 was that wave's turning point; we've spent the two decades since in ICT deployment (broadband, smartphones, cloud, e-commerce), and AI is simply the newest deployment layer, or at most a second frenzy inside a wave that's already old. On this reading AI is sustaining, not foundational: a powerful new feature of the internet era, not the dawn of a successor.
Most essays pick one and write with confidence. The honest move is to notice that Carlota Perez herself isn't sure, and this is the nuance almost everyone omits. In her more recent thinking she argues the ICT wave's golden age is still delayed, that we have not yet had the broad deployment her own model predicts, that the 2008 Great Recession should be folded into a still-unfinished turning point, and that there may be another crash before the golden age finally arrives. Sit with the implication: the woman who built the five-wave machine is uncertain whether we have even exited the fifth wave's turning point, which makes the confident "AI is the sixth wave!" proclamations a little awkward, since they're racing ahead of the very theorist they cite.
What the data tells you cleanly is only that you are in some installation frenzy. The five biggest US cloud and AI firms have guided toward a combined $635 to 690 billion of capital expenditure in 2026, more than double their 2024 level, with roughly 75% of hyperscaler capex now flowing into AI infrastructure. Goldman Sachs and McKinsey model on the order of $765 billion of annual AI capex in 2026 rising toward $1.6 trillion by 2031, about $7.6 trillion cumulatively across the build-out. That is unmistakably what a Perez installation phase looks like from the inside. It tells you the financial capital is racing to lay the infrastructure. It does not tell you which wave's infrastructure, or whether the deployment that's supposed to follow will be a golden age or a long, disappointing plateau.
Now the inconvenient turn, the one that separates an honest essay from a flattering one. Long-wave theory is not mainstream economics. It is closer to a heterodox tradition that the discipline has mostly declined to accept.
The waves trace back to Nikolai Kondratiev in the 1920s, and the standard criticism of his work is unsparing: he derived an extraordinarily broad historical generalization from very limited data, made selective use of the statistics that supported his thesis, and produced no sufficient evidence to establish that any genuine regularity exists. The blunt verdict that recurs in the literature is apophenia, the human tendency to perceive patterns that aren't there. And the damning thing about a fifty-year wave is that you can always draw one through the data after the fact. Give me any two centuries of booms and busts and a free hand, and I will find you five waves, or six, or four, each beginning exactly where my argument needs it to.
So the five-wave model is a narrative lens, not a validated predictive law, and here is the insight that makes it worth keeping anyway: its greatest strength and its greatest danger are the same property. Its strength is that it turns a century of economic chaos into a reassuring story with a moral: the crash is normal, the golden age comes after, hold the line. Its danger is that this exact story can make any boom feel like destiny. It can launder reckless capex as "installation phase." It can reframe a genuine plateau as a mere turning point. It is most comforting precisely when we are most lost, which is the property of a good myth, not a good model.
And the honest counterweights pile up once you're looking for them. "The crash funds deployment" is true in aggregate, over decades, with brutal survivorship bias: the rails and the fiber look inevitable in hindsight only because we don't catalog the bubbles that left nothing behind but bankruptcies and wrecked lives. Capex is evidence of a frenzy, not proof of a durable paradigm. And the dark-fiber analogy has a quiet flaw that should make you suspicious of how good it feels: fiber didn't rot. A strand of dark glass laid in 1999 was just as useful in 2009. A GPU loses most of its value in three to five years. If AI's capabilities plateau before the deployment phase arrives, a meaningful chunk of that $7.6 trillion won't wait patiently in the dark to become the substrate of a golden age. It will simply be obsolete, stranded with no second life. Be suspicious of any historical analogy that tells the people spending the money exactly what they want to hear.
So you're a builder or a leader, the framework is real-but-contested, and the placement is genuinely unknown even to its author. What do you do with that? You use the lens for its insights and refuse it as a timetable.
The durable insight is this: the lasting winners of a technological revolution are usually not the darlings of its installation phase. They're the deployment-era companies that buy the stranded infrastructure cheap and build the boring, profitable, world-changing applications on top of it. Amazon and Google were not dot-com bubble stocks; they were the deployment phase, quietly inheriting the cheap fiber and cheap talent the bust left behind. The famous names of the frenzy are mostly gone; the names you use every day showed up after the crash. So if you're placing strategic bets, don't index them to the incumbents of the current frenzy. Ask instead: who is positioned to use nearly-free cognition once the build-out is paid for, and probably written off, by someone else?
The second insight is to separate the infrastructure from the equity. The railway investors lost; the railways won. The cleanest thing to believe in here is not any company's valuation but the capability itself, the cost of a unit of cognitive work falling toward zero, because that is the part that survives the crash the way the rails and the fiber did. Bet on the cheap input outliving the bubble, not on the bubble.
And the reflexive discipline, the one this whole essay has been circling: when a framework makes your uncertainty feel resolved, that is the precise moment to trust it least. Use "five waves" to ask sharper questions (Is this installation or deployment? Who keeps the infrastructure after the crash? Is the cheap input real and durable, or a story?) and never to extract a comforting answer about when the golden age arrives, because that answer is exactly what a pattern-that-might-not-exist most wants to sell you.
The honest location of the AI bubble, then, is this: somewhere inside a pattern that is genuinely illuminating and might not be real, at a spot its own inventor cannot confidently mark, in a wave we may not have finished or may have just begun. The five waves may ultimately be a story we tell because the alternative, that it's all contingent and no one actually knows, is harder to live with. Take the story's wisdom, which is considerable: the crash builds the future, and the winners come afterward. Just refuse it the one thing it most wants to give you, and the one thing it cannot honestly provide: a date.
When a framework resolves your uncertainty, that is the moment to check its basis, not trust it.
The essay's discipline is to keep using a story for its questions while refusing the comforting answer it wants to hand you. The same goes for any confident conclusion an agent reports. Chain of Consciousness is the tamper-evident record of what an agent actually did to reach a result: the evidence it used, the inference it made, the step it took. It lets a reviewer check the conclusion against its basis instead of accepting a clean narrative on faith, which is exactly the move this essay asks for, applied to your own systems.
See Hosted Chain of Consciousness · See a verified action chain
pip install chain-of-consciousness · npm install chain-of-consciousness