The AI Infrastructure Bubble, Explained

Maybe. But “eventually” is doing a lot of work in that sentence, and the financial structure of the current buildout doesn’t have the luxury of waiting for eventually. The debts come due on a schedule. The interest payments are real. And the gap between what’s being spent and what’s being earned is not a rounding error. It’s an 18x chasm.

The $650 Billion Question

AI revenue vs. AI infrastructure spending

JPMorgan published the number in November 2025: for the current pace of AI infrastructure investment to earn a 10% return on invested capital, the AI industry needs to generate $650 billion per year in revenue. Sustainably. Every year. That’s the breakeven for the buildout to make financial sense.

Current generative AI revenue, according to Menlo Ventures’ 2025 State of Generative AI report, is approximately $37 billion per year. Real products generating real revenue. But $37 billion against a $650 billion requirement is an 18x gap. For context, the entire global video game industry generates about $180 billion. The AI buildout needs to generate 3.6x the entire gaming industry’s revenue just to break even.

When One Bet Becomes the Entire Economy

AI capex as share of US GDP growth

The spending itself has become a macroeconomic event. Harvard economist Jason Furman calculated that AI infrastructure investment accounted for 92% of US GDP growth in the first half of 2025. Not tech sector growth. The entire economy’s growth.

When a single capital expenditure cycle becomes nearly the entire growth engine of the world’s largest economy, the stakes change. If AI infrastructure spending slows (credit tightening, demand disappointment, a shift in corporate priorities), there is almost nothing else propping up GDP growth. What happens to the other 8% when the 92% pulls back?

The Debt That Funds the Dream

AI infrastructure debt issuance

To fund the buildout, companies have turned to debt markets at a pace that has no precedent in tech history. Over $200 billion in AI-related debt was issued in 2025. Hyperscalers alone issued $121 billion in bonds for the full year, more than four times the five-year average of $28 billion per year.

Bank of America reported that $75 billion of that came in September and October alone: twice the annual decade average, compressed into two months. This debt is secured against data center assets and future revenue contracts that may or may not materialize.

The credit markets are noticing. AI infrastructure credit spreads (the premium lenders charge above risk-free rates to fund data centers and GPU clusters) have been widening since mid-2025. Credit markets have historically repriced risk before equity markets in every major cycle. In 2006, credit default swap spreads on subprime debt widened 18 months before equities peaked. We’ve traced where that repricing shows up first in the current cycle.

The Capex Hockey Stick

Hyperscaler capital expenditure, 2020–2026

The spending curve tells the story most clearly. Combined hyperscaler capital expenditure has gone from $94 billion in 2020 to an estimated $650 billion in 2026. A 7x increase in six years.

The closest historical analog is the telecom fiber buildout, which peaked at roughly $150 billion per year in inflation-adjusted dollars. AI infrastructure spending is on track to exceed that by more than 4x at the equivalent point in the cycle. The companies funding this buildout are already cutting headcount to pay for it, a pattern we document in AI Layoffs or Capex Layoffs?

The Depreciation Trick

GPU useful life extensions and earnings overstatement

Meanwhile, the companies at the center of the buildout are making accounting choices that obscure the true cost. All five major hyperscalers have extended the stated useful life of their server equipment, in some cases from 3 years to 6. Perfectly legal. And it has the effect of suppressing depreciation expenses and inflating reported earnings.

Michael Burry estimates the cumulative overstatement at $176 billion over the 2026–2028 period, based on his analysis of public 10-K filings. The GPUs don’t care about accounting. The actual product replacement cycle in AI infrastructure is 2–3 years, driven by the pace of architectural improvement. A 6-year useful life assumption implies that a GPU purchased today will still be commercially viable in 2032. Nobody in the industry actually believes that.

“Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives... Yet this is exactly what all the hyperscalers have done.”– Michael Burry, November 10, 2025

We’ve Seen This Before

Telecom fiber bubble vs. AI infrastructure bubble

The technology is real. It’s improving rapidly. That was also true of fiber optic cable in 1999. The financial structure built around AI (the debt, the spending ratios, the revenue gap, the accounting choices) follows a pattern that has appeared before. The telecom fiber buildout of 1999–2001 saw the same dynamics: massive capital deployment, debt-funded expansion, and a revenue curve that couldn’t keep pace with the investment curve. The fiber got used eventually. The companies that built it mostly didn’t survive to see that happen.

What the Stress Index Says

AI Infrastructure Stress Index composite score

Revenue gap, debt levels, depreciation extensions, GDP dependency. Each of these is a data point. Individually, they tell a story. Together, they form a pattern. The AI Infrastructure Stress Index was built to track that pattern in real time.

The index aggregates 12 signals across credit markets, operational fundamentals, and market sentiment into a single composite score from 0 to 100. Here are the five currently scoring highest.

As of February 12, 2026, the composite score sits at 34 out of 100, in the yellow zone. The index’s verdict: On Watch. Three individual signals are deep in the red (AI Adopters at 100, Options Sentiment at 86, GPU Spot Pricing at 81), even as the composite stays moderate. The score updates daily. The data is public. Decide for yourself.