The AI Bubble is Too Big to Pop Quietly

If this market deflates, it could shake pensions, governments, and the very infrastructure of the internet.

When the dot-com bubble burst in 2000, its symbol was almost comic: a hand puppet selling dog food. Pets.com, which went bankrupt nine months after its $82.5 million IPO in February 2000, became the punchline for an era of easy money and exuberant storytelling. It was embarrassing. Costly. And, in retrospect, relatively contained.

Fast-forward 25 years to shares of AI infrastructure firm Coreweave erasing roughly $24 billion in market value in just two trading days. That single collapse is almost 60 times the size of Pets.com at its peak. However, the difference is not just a matter of numbers on a balance sheet. It reflects something more consequential: the fact that today’s AI boom is not a niche phenomenon confined to speculative investors. Rather, it is a reordering of global capital markets on a scale that touches pension and index funds, sovereign wealth portfolios, and even the plumbing of the internet itself.

This is why the question, “are we in an AI bubble?”, misses the point. Of course we are. Prices have run ahead of reality, failures are piling up, and projects are quietly dying. But to compare this to the late 1990s is to misread the moment. The dot-com mania was a fever dream, while the AI surge is a structural bet on the future of the economy. If it collapses, it won’t just be day traders licking their wounds. It will be schoolteachers in Ohio watching their pensions shrink and municipalities recalculating budgets built on inflated asset values.

The evidence of mania is easy to find. Valuations have drifted into the surreal. Oracle’s stock price spiked nearly 40 percent last month in its best day since 1992, briefly catapulting Larry Ellison into the position of the world’s richest man. That kind of move might be expected from a meme stock or a thinly traded biotech, but from a company that has been a fixture of enterprise software for decades, it’s eyebrow-raising at the very least. Meanwhile, the IPO market, dormant for years, has roared back with a vengeance. Companies are doubling on their first day of trading simply because they can attach “AI” to their narrative. 

Beneath the surface, the infrastructure binge is staggering. Tech giants are spending hundreds of billions to build out data centers, training clusters, and custom chips. The parallels to the fiber-optic overbuild of the 1990s are plentiful: too much money chasing too much capacity, long before the demand to justify it arrives. A new MIT study, in fact, found that 95% of generative AI projects are failing to deliver measurable returns. The hype has outpaced not just adoption, but also the very ability of companies to figure out what to do with these tools.

Even OpenAI CEO Sam Altman, who is perhaps the most important figure in the industry, has admitted the obvious, telling reporters, “are we in a phase where investors as a whole are overexcited about AI? My opinion is yes”, earlier this year.

But if the signs of a bubble are clear, so too are the reasons this isn’t just 1999 all over again.

First, today’s AI giants are not vaporware. Nvidia, the beating heart of the current mania, has more than tripled revenues in just two years. Microsoft and Amazon are generating tens of billions in annual cash flow from AI-linked cloud services. These companies are not Pets.com; they are some of the most profitable enterprises in human history.

Second, adoption is moving faster than any prior technology cycle. The Pew Research Center says 34% of adults in the U.S. have used ChatGPT, which is around double the share in 2023. That pace of diffusion makes the early internet look sluggish. Even if many of those use cases are trivial or experimental, the cultural and organizational uptake is undeniable.

Third, bubbles can be productive. The railroad mania of the 19th century bankrupted investors but built the backbone of continental commerce. The telecom bubble left behind a glut of fiber that made the broadband era possible. Today’s data center arms race may look reckless, but the infrastructure it produces could become the essential utility of the 21st century, enabling applications we cannot yet imagine.

The paradox that suggests bubbles can destroy wealth even as they accelerate progress is what makes the AI moment so complicated. To dismiss it as mere hype is to overlook the enduring infrastructure being laid. To celebrate it as a straight path to productivity is to ignore the staggering waste and risk of concentration.

The true risk lies not in over-optimism per se but in centralization. In 2000, the failure of a dozen internet darlings bruised portfolios but didn’t imperil the financial system. Today, the fate of the AI story rests disproportionately on a handful of firms. Nvidia’s chips are the bottleneck for nearly every frontier model. Microsoft has integrated generative AI into its operating system, productivity software, and cloud business. Amazon is retooling its core services to accommodate machine learning demand. If one of these giants stumbles, the ripple effects would extend far beyond Silicon Valley.

It is tempting, then, to argue that we should dampen the enthusiasm before it metastasizes. But that misses another uncomfortable truth: bubbles may be the only way societies finance certain kinds of progress. Democracies are bad at long-term industrial planning. We don’t build railroads or fiber or massive computing grids through five-year plans. We build them through speculative manias, in which investors overpay for the promise of a future they can’t yet see. Most lose money. Some make fortunes. But the infrastructure remains available for the next generation to use more wisely.

This is not to say we should cheer the bubble on. If AI stocks deflate the way dot-coms did, the Nasdaq could lose half its value, and the effect would be far more widely felt than in 2000. As University of Michigan professor Erik Gordon put it, “more investors will suffer than suffered in the dot-com crash, and their suffering will be more painful.”

The lesson, therefore, is not to deny the existence of a bubble but to prepare for its aftermath. Policymakers must ensure that the infrastructure binge leaves something useful behind. Corporations must stop treating AI as a marketing gimmick and start integrating it into workflows that actually generate value. Investors must recognize that the belief that AI will change everything can be true without every price being justified.

What makes this moment equal parts intoxicating and dangerous is that both sides of the argument are right. AI is overhyped, and AI is transformative. Valuations are unsustainable, and profits are real. The future will look nothing like the present, but the present is already reshaping the future.

The dot-com crash gave us inexpensive bandwidth, hardened protocols, and a generation of entrepreneurs who knew how to build at scale. The AI crash, if and when it comes, may give us something even more enduring: an ambient layer of computational capability, embedded in the everyday life of organizations and societies.

That is why the better question is not whether this is a bubble. Rather, we should be asking what kind of bubble we are in and what it will leave behind. If history is any guide, the losses will be painful, but if the infrastructure is durable, the pain may yet be worth it.

Michelle Yu (MBA '26) is originally from Cresskill, New Jersey. She graduated from Columbia University with a degree in Film and Media Studies and worked for CNBC, NBC News, and CNN prior to HBS, along with projects for HBO, Showtime, Oxygen, and Spectrum. Outside of work, she is a 2x marathon runner, American Songwriting Awards winner, and filmmaker whose work has screened at the Tribeca Film Festival and AMC's Empire Theaters in Times Square.