The AI Spending Trap Just Ate Two Bellwethers

Wall Street spent the last 18 months building an unfalsifiable thesis: no matter what happens, AI capex wins. The Fed cuts rates? Capex accelerates. The Fed pauses? Capex accelerates. Inflation spikes? Capex accelerates. We get a war in the Middle East? Capex accelerates. It was a kind of holy writ—the one narrative that couldn't break, because everyone from Satya Nadella to Jensen Huang had stake-holders convinced that $500 billion a year in infrastructure spending was not just inevitable, it was profitable.

Thursday and Friday just poured acid on that story.

Oracle dropped 13.3% Thursday after revealing its fiscal 2026 capex would be $15 billion higher than the September guidance—a number so appalling that it triggered not just a stock selloff, but credit-default swap widening. The market didn't just worry Oracle had misjudged spending; it worried Oracle couldn't finance it. Broadcom then reported record fiscal year revenue (24% up year-over-year, AI revenue surged 65%) and the stock dropped 10-12% anyway, because management flagged margin compression. The earnings were pristine. The inventory was clean. The customer mix was accelerating toward AI. And none of it mattered. The tape said: we don't want more revenue if you can't keep the profit margins wide.

This is a guillotine hitting the spine of the 2025 rally.

The rotation story floating around—"money is just moving from AI to value"—is a useful fantasy. What's actually happening is a repricing of the entire buildout economics. Two different ways to lose at the AI capex table just got exposed.

Oracle's way: You're the buyer. You commit to massive spending, but your debt load creeps up, your payback period stretches beyond what finance models support, and the market decides you're taking on leverage to chase a trend, not to fund a cash-cow business. When you miss the spending guidance (underestimating demand is a sign of success, theoretically), the market hears: "This is spiraling. We don't have clear control." Credit traders price in risk. Stock drops.

Broadcom's way: You're the supplier. Customers are desperate for custom silicon, they're buying in volume, and your AI revenue is becoming the lion's share of the total. But they're also squeezing you on price, forcing you into lower-margin custom chip designs instead of higher-margin standard products. You sell more, you make less per unit. Revenue is a vanity metric when return-on-capital matters. Stock drops anyway.

The two bellwethers are on opposite sides of the AI infrastructure funnel, and both are getting punished because investors just realized: the question isn't whether AI capex is real; it's whether anyone making money off it actually profits.

This matters because the meta-narrative of the whole 2025 bull case was built on a cascade of certainties: capex drives demand, demand justifies pricing power, pricing power funds dividends, dividends justify valuations. Each step of the chain looked airtight when you said it fast enough in a earnings call. But earnings are brutal on assumptions. Oracle just proved you can have increasing demand and deteriorating finances simultaneously—which breaks the axiom that "AI boom = broad-based earnings lift." Broadcom proved you can have growth and margin collapse together—which breaks the other axiom, that suppliers are the safe way to play AI.

What's left is a world where:

• Hyperscalers (the mega-cap incumbents—Google, Microsoft, Amazon, Meta) are still spending because they can absorb the capex hit over time and will extract rent from AI services. They're fine. Their stocks are fine. They're trading on forward earnings and dominance, not on "capex unlocks value."

• Infrastructure vendors and suppliers are now being valued on whether this year's revenue turns into this year's profits, not on "capex sustains for a decade." That's a harder test. Broadcom failed it on optics. Nvidia hasn't reported yet.

• Secondary beneficiaries—the cooling systems, the real estate plays, the software companies licensing content to OpenAI for $1 billion (hi, Disney)—are still kicking because they're not the ones bleeding capex in both directions.

Meanwhile, jobless claims spiked to 236,000 last week, Oracle is delaying data-center openings into 2027, and the Fed just signaled a pause in rate cuts starting January. The housing of assumptions is getting darker.

The market isn't declaring AI dead. It's declaring "AI at any cost" bankrupt. There's a difference. Investors still believe in the buildout; they've just stopped believing it's free, risk-free, or profitable in the 24-month window they care about.

That's a bloodier repricing than a simple narrative flip. That's a reckoning with geometry itself.