Every technological mania arrives wrapped in a new narrative, yet all of them reveal the same human truth:
how we react when the future seems to accelerate faster than our ability to understand it.
AI triggers hope, fear, ambition, and vertigo in equal measure. And from the inside, everything feels inevitable, logical, even obvious… until it no longer is.
In recent weeks, I read four Q3 2025 letters written by investors I respect —Wedgewood Partners, Praetorian Capital, Greenlight Capital, and Hayden Capital— each of them observing the same fever from a different angle.
Rather than trying to decide who was “right,” I was interested in understanding what we could learn by comparing their perspectives.
These four letters didn’t just teach me about the current state of AI.
They taught me about the way investors interpret a mania.
Lesson #1 — (Wedgewood Partners): The Scale of the Boom
“Manias feel inevitable when the numbers get so big they stop having meaning.”
Wedgewood Partners was the first to remind me of something uncomfortable:
Behind every boom, there is a point where enthusiasm stops acknowledging arithmetic.
Their letter includes an observation that perfectly captures the current moment:
“Artificial Intelligence (AI) stocks long continue to be darlings of Wall Street… yet, to try to grasp the ever-growing forecasts of trillions in the necessary infrastructure to support the current insatiable demand does give an investor pause — particularly given how crowded the ‘AI trade’ has become.”
What’s interesting is the idea beneath that sentence:
The notion that AI is not only absorbing capital, but also the collective imagination.
To articulate that feeling, the Wedgewood team turns to an unexpected cinematic metaphor: The Blob.
Just as that creature devoured everything in its path, today’s investment in AI is creating a growing mass of CAPEX, credit, and hype—within which economic logic begins to dissolve.
The numbers explain why:
The tech sector—traditionally “asset-light”—is mutating into a capex-heavy, debt-dependent model.
For two decades, data centers were funded with internal cash flow; yet in 2025 alone, the tech giants raised US$150 billion in debt to sustain the AI fever.
According to Bain & Co., the US$500 billion in required CAPEX will demand US$2 trillion in revenue to be justified.
Figures that, in Wedgewood’s words, defy analytical comprehension… and yet, the market is cheering.
But the most unsettling part isn’t the magnitude of the spending—it’s the speed.
The GPUs powering the revolution depreciate in 3–5 years, while NVIDIA is already accelerating toward annual upgrade cycles.
This means massive amounts of capital could become obsolete long before recouping their investment.
Wedgewood calls this a potential wave of depreciation-driven earnings compression.
And then there’s another constraint people discuss far less: energy.
Some data centers consume as much electricity as an entire U.S. city.
Wedgewood warns that the real bottleneck won’t be technological, but political.
Without a massive expansion of the power grid—and possibly water access—AI will run straight into the physical limits of the system.
Lesson #2 — (Harris Kupperman of Praetorian Capital): When the Math Doesn’t Add Up
How much revenue is needed to justify the current level of capex spend and give AI investors a return on their capital??
Harris “Kuppy” Kupperman of Praetorian Capital describes something unsettling.
After speaking with insiders in the data-center ecosystem, he discovered that many of them—the very people building the AI infrastructure—are confused.
They don’t understand how these projects are supposed to generate economic returns.
And the most concerning part isn’t that they don’t understand it. It’s that investors don’t either.
Kuppy agrees with Wedgewood on a critical point: GPUs become obsolete in 3 to 5 years, precisely as NVIDIA accelerates its upgrade cadence.
This means that to reach a reasonable breakeven, the industry would need roughly US$1 trillion in revenue—an amount that turns any financial model into an act of faith.
Then comes the uncomfortable question he raises: hyperscalers have spent years subsidizing money-losing businesses to pull users into their ecosystem.
Kuppy wonders whether AI is simply another one of those businesses that, strategically, can lose money indefinitely.
Part of the market believes these costs will be offset once AI replaces highly skilled jobs.
Harris isn’t buying it: that disruption—if it arrives—is years away and will require technology far more advanced than what exists today.
His diagnosis is broader: without U.S. government intervention, and without the investment boost created by AI, the United States would already be in a recession. In fact, he argues that the AI bubble may represent more than 100% of recent economic growth.
To illustrate the phenomenon, he turns to the 19th-century railway bubble: a transformative industry, enormous promise, insiders convinced that investment had to continue endlessly…
Yet the cycle still blew up multiple times, triggering financial panics and sending many companies into bankruptcy.
In his sharpest line, he compares the two eras:
Raising capital for a money-losing business is hard. Back then, you got a free train trip. Now you get the free use of an LLM.
This investor is also skeptical of the circular investment patterns among Big Tech: companies investing in one another to justify growth—alarmingly similar to what happened with fiber-optic networks during the dot-com bubble.
Finally, he warns that this bubble won’t burst because of a technical signal, an obvious metric, or a sudden narrative shift.
His conclusion is human, almost psychological:
Like all bubbles, it implodes when people get tired of funding it.
Lesson #3 — (David Einhorn of Greenlight Capital): The AI Spending Delusion
There is a good chance that 25 years from now, AI will turn out to be even more important than we currently imagine. But, the path from here to there is likely to be very bumpy for investors.
David Einhorn of Greenlight Capital arrives at conclusions eerily similar to Kupperman’s: He too sees a deep disconnect between the enormous sums being allocated to AI and the basic arithmetic required to justify them.
The most revealing part is that not even the protagonists themselves know how they’re supposed to earn back what they’re spending: Tim Cook (Apple), Mark Zuckerberg (Meta), and Sam Altman (OpenAI) have all failed to articulate a clear path to economic returns in the face of this massive wave of CAPEX.
Their answer is always the same: “This will eventually pay off.”
But the key word is “eventually,” and Einhorn wonders how long that eventually can last.
In his letter, he cites a McKinsey study projecting something monumental: by 2030, global AI-related CAPEX could reach US$6.7 trillion—with US$5.2 trillion going to data centers alone.
The question he raises is simple and brutal:
Where will those nearly US$7 trillion come from?
His analysis is straightforward: the math doesn’t add up.
Not even the “Mag Seven” —the most profitable companies on the planet— generate enough free cash flow to finance this vision of the future.
Collectively, they produce around US$500 billion in FCF per year.
Half of that is already committed to stock buybacks and dividends.
Their total book equity is roughly US$1 trillion — a fraction of the spending required.
Even if Wall Street, private equity, and venture capital emptied their coffers, there would still be a multi-trillion-dollar gap that could only be filled with massive new debt.
To justify all this spending, AI would need to generate about US$2 trillion in annual revenue by 2030—equivalent to the combined size of today’s global market for advertising and software subscriptions.
Einhorn’s question is obvious: Is that even remotely plausible?
He then dives into something even more troubling: the illusion of revenue.
He describes how much of the “AI revenue boom” doesn’t come from real customers, but from AI companies buying products from one another.
His line is precise:
Much of the revenue in AI simply comes from AI companies buying products and services from each other.
He even walks through a concrete example:
A single dollar spent on a ChatGPT subscription can turn into eight dollars of “reported revenue” along the chain—OpenAI to Microsoft, to CoreWeave, to NVIDIA—without a single dollar of net economic value being created.
It is, quite literally, the same dollar inflating multiple income statements.
This dynamic, he warns, creates a false illusion of growth while masking value destruction, overcapacity, and a fragile structure dependent on cheap capital.
Einhorn’s other critical point is the almost religious faith in AGI (Artificial General Intelligence).
He challenges the belief that we are close to machines that “think” or improve themselves.
Current models, he argues, are statistical systems that mimic patterns; they do not understand, reason, or make real inferences.
To illustrate this excess of optimism, he points to the case of self-driving cars: for a decade, they were always said to be “one year away.” And yet, that year never arrived.
The expectation of imminent AGI, he argues, is not technology.
It is wishful thinking, amplified by financial incentives.
Lesson #4 — (Fred Liu of Hayden Capital): Bubbles as Engines of Progress
Not all bubbles destroy. Some prepare the ground.
In contrast to the skeptical tone of the other managers, Fred Liu of Hayden Capital takes a different stance:
Stop looking at the trees and step back to see the forest.
He doesn’t focus on whether we are or aren’t in a bubble.
He proposes something more interesting: to ask what kind of world is being built by this wave of AI investment.
His thesis runs counter to market intuition:
Bubbles are a feature, not a bug, of technological progress.
Instead of seeing them solely as financial tragedies, Liu sees bubbles as mechanisms of civilizational advancement: periods of excess that finance critical infrastructure, massive experimentation, and new mental models.
To explain this, he turns to three illuminating historical examples.
1. “Canal Mania”: The Infrastructure That Survived the Speculation
In the late 18th century, the British financed thousands of miles of canals in a speculative frenzy. Many companies went bankrupt. Many investors lost everything.
But the country was left with a network of 4,000 miles of canals that drastically reduced transportation costs, enabled the movement of fragile and heavy goods, and became the logistical backbone of the Industrial Revolution.
Liu highlights something fascinating:
Canal transport became cheaper after the bubble burst because the collapse erased the original capital costs and forced competition among overbuilt routes. New owners could price at marginal cost instead of the much higher, original construction cost, turning canals into near-free public goods.
The capital was lost.
The infrastructure remained.
And with it, a new economy became possible.
2. The Railroad Boom: The Excess That Unified a Country
In the late 19th century, the United States experienced another wave of euphoria: railroads.
In 40 years, the nation went from 30,000 to 200,000 miles of track—the largest railway system in the world.
It was a period of extreme overinvestment: fares fell 80%, competition was brutal, and a quarter of all rail lines went bankrupt.
But the country emerged more unified, more industrialized, and with new industries that would have been impossible without that excess:
national catalogues (the original Amazons),
industrial-scale agriculture,
unified markets,
mobility and economic expansion.
Financial chaos gave way to an unprecedented productive leap.
3. The Dot-Com Bubble: Overbuilding That Created the Modern Internet
This one is closest to our memory.
The dot-com bubble produced:
excess fiber-optic cable,
unused data centers,
massive bankruptcies,
delusional expectations,
“dark fiber” sold for ridiculous prices.
But that excess became the backbone of the modern Internet:
cheap broadband,
ecommerce,
streaming,
SaaS,
cloud computing,
smartphones.
Liu describes it as a process of parallel experimentation:
New browsers, search engines, ecommerce models, and most importantly:
An entire population coming online, creating the cognitive foundation necessary for the digital economy that would dominate the following decades.
Conclusion
What changes is the technology.
What doesn’t change is how we react to it.
AI can be a bubble, an opportunity, or a revolution. It’s likely a bit of all three.
What is certain is that the challenge remains the same as always:
To think calmly in a world that keeps accelerating
I’m curious: which of the four perspectives resonated most with you?
Let me know in the comments.
If you’d like to keep reading pieces like this one, I’d be glad to have you as a subscriber.
Ezra’s Notes is free, and it helps me write more essays worth your time.
and that’s why behavioral economics will always be my favorite!
I definitely do think that AI is a bubble but I do think that after everything AI will stick around in our lives after the bubble pops. Just like the Dotcom bubble, internet stayed with us after the bubble popped.