When Bubbles Build: How Investors Might Lose and Society Wins
In 1845, Britain had roughly 2,000 miles of railway track. Over the next two years, Parliament authorized another 8,000 miles. The intervening period, known as Railway Mania, saw hundreds of new railway companies launch. Shares traded at enormous premiums. Lawyers, clergy, and factory workers all speculated. By 1850, about a third of these companies had collapsed or were never built. Share prices fell 66%. Many investors lost everything.
Britain kept the railways.
This pattern repeats with eerie consistency across technological history. Investors fund transformative infrastructure at a loss. The infrastructure outlasts the investors. Society wins; shareholders don't.
The Pattern
The 1990s fiber optic boom laid 80 million miles of cable across the ocean floor. When the dot-com bubble burst, companies like WorldCom (once valued at $180 billion) and Global Crossing went bankrupt. Roughly $5 trillion in market value evaporated between 2000 and 2002. The capital was gone. The fiber remained. That "wasted" infrastructure became the backbone for YouTube, Netflix, cloud computing, and everything we now consider the modern internet.
The US railroad expansion of the 1870s-1890s followed the same script. By 1894, a quarter of American railroad mileage was in receivership. Cornelius Vanderbilt's rivals went bust. Jay Gould's empire crumbled. But the transcontinental railroad had already made it possible to ship goods coast-to-coast in days instead of months, enabling an industrial transformation that dwarfed any stock market losses.
The mechanism is always the same: competitive pressure drives overinvestment, capital floods in faster than revenue can justify, a correction wipes out the weakest players, and the surviving infrastructure generates value for decades.
The $600 Billion Question
In June 2024, Sequoia Capital's David Cahn published an analysis that landed like a cold shower on AI optimism. His math was blunt: NVIDIA was on track for $150 billion in data center GPU revenue. But GPUs are only part of the bill — data centers, energy, networking, cooling, and maintenance multiply the total cost. Cahn estimated the full AI infrastructure bill at roughly $600 billion per year.
Against that, AI revenue (excluding NVIDIA's own sales) totaled roughly $100 billion. The gap: $500 billion in spending that has to be justified by revenue that doesn't yet exist.
Big Tech Total Capex ($B, AI-driven surge from '23)
┌─────────────────────────────────────────────────────────────────┐
│ ++│ 400
│ ++++ │
│ +++ │
│ ++++ │
│ ++++ │
│ ++++ │
│ +++ │
│ ++ │ 200
│ +++ │
│ ++ │
│ +++++++++ │
│+++++++++++++++++++++++++ │
└─────────────────────────────────────────────────────────────────┘
'20 '22 '24 '26
The numbers have only grown since Cahn's analysis. Microsoft committed $80 billion in AI infrastructure for fiscal year 2025 alone. Combined, the four hyperscalers (Microsoft, Google, Amazon, Meta) are projected to spend over $320 billion in 2025, up from $230 billion in 2024. For 2026, estimates approach $400 billion.
This is not venture capital chasing concepts. These are profitable companies investing their own cash flows. Microsoft's AI-related revenue hit $13 billion annualized by early 2025, growing rapidly but still a fraction of its capex commitment.
Not Your Father's Bubble (But Maybe Your Great-Grandfather's)
The instinct to compare AI to the dot-com bubble is understandable but imprecise. In 2000, Pets.com went public on a concept. Today's AI companies have real products used by hundreds of millions of people — many of them free or heavily subsidized, which is precisely how massive adoption coexists with a massive revenue gap. ChatGPT reached 300 million weekly active users. GitHub Copilot is embedded in professional workflows worldwide. Google's AI-enhanced search handles billions of queries daily.
The better comparison is the railroad or fiber optic parallel: companies with real technology and real demand, but whose investment pace assumes a future that may take longer to arrive than investors expect.
Howard Marks of Oaktree Capital framed it precisely in a January 2026 memo: what makes bubbles dangerous isn't that the underlying technology is bad. It's that "the price is too high." NVIDIA trading at 45x earnings, or data centers being built on demand projections extrapolated from 18 months of ChatGPT growth, could represent sound investments. They could also represent 1999 Cisco (which took 25 years to recover its dot-com peak price, finally doing so in December 2025).
Even AI's champions acknowledge the risk. Sam Altman said some AI companies "are going to get washed out" and that "a lot of the companies that have been built on top of AI... are going to struggle." Cahn's analysis carried a similar warning: the picks-and-shovels strategy only works if the gold rush materializes at scale.
The Paradox That History Resolves
Here's the observation that makes the AI bubble conversation more nuanced than it first appears: the bubble thesis and the transformative technology thesis are not contradictions. They're historically codependent.
Every transformative infrastructure in modern history was overbuilt by investors who lost money, then utilized by society at a fraction of the construction cost.
Carlota Perez, the Venezuelan-British economist who has spent decades studying technological revolutions, argues that financial bubbles are not bugs in the system. They're features. Bubbles channel massive capital into infrastructure that rational, patient investment alone would never fund at the necessary speed or scale. The railroad investors of the 1840s wouldn't have authorized 8,000 miles of track with conservative return expectations. The telecom companies of the 1990s wouldn't have laid transoceanic fiber for reasonable internal rates of return.
Bubbles, in Perez's framework, are the mechanism through which speculative capital builds the infrastructure for the next era of productive growth. The speculative phase inevitably collapses. The deployment phase that follows uses the overbuilt infrastructure at bargain prices.
Who Pays, Who Benefits
The AI version of this pattern would look something like this: hyperscalers spend $1-2 trillion on GPU clusters and data centers over 2024-2028. Some of that spending generates direct returns through AI products. Much of it doesn't — at least not on the timeline investors expect. Stock prices correct. Some AI startups fail. The infrastructure remains.
And then the interesting part begins. Just as cheap fiber after 2002 enabled a generation of startups (Dropbox, Netflix streaming, AWS) that the fiber builders never imagined, cheap AI compute after a correction could enable applications we can't predict today.
The DeepSeek episode in January 2025 offered a preview. The Chinese lab demonstrated that comparable AI performance could be achieved at a fraction of the cost, wiping $600 billion off NVIDIA's market cap in a single day. Whether DeepSeek's specific claims hold up matters less than what they revealed: the cost curve for AI inference is dropping precipitously. The infrastructure being built today will likely become cheaper to operate than its builders are projecting. Whether today's GPU clusters prove as durable and general-purpose as fiber optic cable remains an open question — hardware becomes obsolete in ways that cable does not.
This is exactly what happened with fiber optics. The overcapacity of the late 1990s became the cheap bandwidth that made streaming video, cloud computing, and the smartphone ecosystem possible.
The Uncomfortable Conclusion
The honest assessment of AI's trajectory is that both sides of the debate are probably right, just as both sides were right about railroads and fiber optics.
AI is likely transformative. It is also likely overvalued. These are not competing claims. They are the same historical pattern, playing out on a larger scale and at a faster pace than any previous technology cycle.
For investors, the implication is uncomfortable: you can be right about the technology and still lose money. Cisco shareholders were right that the internet would transform everything. They still waited 25 years to break even.
For society, the implication is quietly optimistic. The infrastructure being built right now — the data centers, the chips, the models, the tooling — will very likely generate enormous value. Just not necessarily for the people paying for it.
The railways carried freight for a century after the speculators went bankrupt. The fiber still lights up the ocean floor. And the AI infrastructure being poured into the ground right now will process queries long after the current stock prices are forgotten.
Bubbles destroy wealth. They also build worlds.
Links: AI's $600B Question (Sequoia Capital) | The Bubble Memo (Howard Marks / Oaktree) | Technological Revolutions and Financial Capital (Carlota Perez) | British Railway Mania (Wikipedia) | DeepSeek market impact (CNBC) | Big Tech AI Capex (Reuters)