The Data Center Bubble

In the late 1990s, companies like Global Crossing spent tens of billions building fiber-optic networks that would supposedly revolutionize telecommunications. The infrastructure was real, the technology transformative, but the economics never worked. By the time Global Crossing filed for bankruptcy in 2002, it had become one of the largest corporate failures in American history. Corning, the fiber optics supplier that was the NVIDIA of its day, lost 97% of its share value between 2000 and 2002.

Today, we’re watching the same script unfold with AI data centers, except the scale is far larger and the potential economic damage more severe. As Harvard economist Jason Furman recently pointed out, data center construction was the primary driver of GDP growth in the first half of 2025. Strip out data center spending and GDP growth was essentially zero.

The AI industry generates $15-20 billion in annual revenue against $40 billion in depreciation, investment manager Harris Kupperman estimates. The hardware becomes obsolete far faster than traditional infrastructure. To break even, the industry would need approximately $1 trillion in revenue across 2025-2026. Current projections suggest that’s not remotely achievable.

This isn’t just about individual company finances. When a single category of speculative infrastructure investment props up national GDP growth, you don’t have a healthy economy—you have a bubble waiting to pop. The technology is real and transformative, but that doesn’t mean current valuations and spending levels make sense. Fiber optics changed everything, just not in the timeframe or manner that justified the capital invested in the late 1990s. AI will likely follow the same path.

What’s different this time is scale. As Practice Capital notes, the investors footing today’s bills will probably regret ever making these investments. Companies are building out infrastructure on borrowed money, buying hardware that depreciates rapidly, hoping revenue will eventually materialize to justify the expense. Kupperman’s warning cuts to the core: doing things at massive scale doesn’t make them work—it just turns an industry crisis into a national one. Right now, we’re taking what should be a manageable tech sector correction and potentially engineering it into a broader recession through sheer capital misallocation.

The bulls will argue that AI revenue will ramp faster than anyone expects, that the technology will find unexpected applications that justify current spending. Maybe they’re right. But betting hundreds of billions on “maybe” has historically ended poorly. The market hasn’t fully priced in this risk because the AI narrative—like the internet narrative in 1999—is too compelling to question.

When reality reasserts itself, the correction won’t be limited to AI companies and data center operators. The pain will spread through companies making data center equipment, commercial real estate trusts holding data center properties, power utilities that invested heavily in capacity for these facilities, and banks that financed the construction.

I’m not predicting precise timing or magnitude. But you can’t sustain an industry on depreciation that’s twice your revenue, and you can’t sustain an economy on speculative infrastructure spending that never generates adequate returns. The Global Crossing bankruptcy caught investors by surprise because the narrative was too seductive to question. When today’s data center bubble pops, it will be a meaningful fraction of the American economy. Anyone who remembers what even smaller bubbles have done to broader financial stability should be terrified.