The AI Bubble Has Its First Real Cracks. Here Is What the Data Shows.

AI infrastructure spending is accelerating. Oracle just projected $50 billion in capital expenditure for fiscal 2026. CoreWeave tripled its revenue in a year. And yet the stocks are breaking down. CoreWeave is down 57% from its 52-week high. That divergence between spending and price is the first real crack in the AI trade.

The Numbers

What the Debt Is Saying

The stock prices are not moving on sentiment alone. The debt markets are telling the same story. Oracle raised $18 billion in bonds in late 2025, one of the largest bond sales in tech sector history, and is targeting even larger raises in 2026. Despite maintaining an investment-grade credit rating, the yields on Oracle’s bonds have slipped into junk bond territory. Its five-year credit default swaps, essentially insurance against Oracle failing to repay its debts, have tripled in price and are now trading at levels not seen since the 2008 global financial crisis.

CoreWeave’s numbers are more acute. The company carries roughly $15 billion in debt, nearly four times its total annual revenue. Last quarter, it paid $311 million just to cover interest, an amount that is more than 20% of its total revenue and approximately six times its gross profit. Revenue is growing fast. Debt is growing faster.

The Concentration Problem

Both companies have a version of the same structural risk: nearly all of their revenue comes from a small number of customers. For Oracle, the dominant customer is OpenAI, which has committed $300 billion over five years to Oracle’s services. OpenAI remains deeply unprofitable and its annualized revenue is roughly one-fifth of what it has committed to spend with Oracle annually. For CoreWeave, the key customers include Microsoft and other hyperscalers, who are simultaneously its biggest clients and its most capable competitors. If AI demand cools even modestly, those hyperscalers have every incentive to bring workloads in-house and cut out the middleman.

What the Dot-Com Parallel Actually Shows

The comparison to 2000 is worth making carefully. In the dot-com bubble, infrastructure spending did not peak when stocks peaked. Cisco, WorldCom, and Global Crossing kept building fiber networks for 18 months after Nasdaq began its decline, because the contracts were already signed and the debt was already raised. The spending continued. The companies did not. The pattern in AI looks similar: capex commitments are locking companies into spending cycles that will run regardless of whether demand materializes at the projected scale.

One Thing To Watch

Watch CoreWeave’s next earnings report for any signal that Microsoft or other hyperscaler customers are pulling back or renegotiating contracts. That is the tripwire. CoreWeave at $81 is already pricing in significant stress. If a major customer signals reduced commitment, the stock has further to fall and the broader AI infrastructure trade reprices with it. The data is already moving. The headlines have not caught up yet.