Wall Street is panicking about AI spending — and getting the story completely backwards.

Last week, the “Hyperscale Five” — Amazon, Alphabet, Meta, Microsoft, and Oracle — revealed their combined 2026 AI infrastructure budget: over $700 billion. That’s nearly $2 billion per day being poured into chips, data centers, and power. Predictably, traders looked at the price tag, screamed “peak capex,” and started dumping AI supply chain stocks like they were going out of style.

Here’s the problem with that thesis: it fundamentally misunderstands where AI spending is headed.

For the past two years, the AI bull run was powered by training — the one-time, capital-intensive process of building models. Bears assume that once models are trained, the spending stops. But February 2026 earnings data tells a different story: inference compute volume has now officially overtaken training compute. And that changes everything.

Training is a one-time capital expenditure. You build the model and you’re done for a while. Inference, on the other hand, is a utility — it scales linearly with every single user, every query, every interaction. It never shuts off. As advanced “reasoning” models become the standard, they use something called test-time scaling, which deliberately runs more compute per query to deliver better answers. That transforms AI from a bursty workload into a 24/7 industrial process.

Translation: the $700 billion isn’t a peak. It’s a floor.

Meanwhile, the “where’s the ROI?” crowd is conveniently ignoring Google’s most important number from last quarter: a $240 billion cloud backlog, up 55% year-over-year. Google isn’t spending because it “hopes” customers show up — it’s spending because it already has $240 billion in signed contracts it physically cannot fulfill without more chips. Microsoft’s cloud backlog has ballooned to roughly $625 billion. These companies are supply-constrained, not demand-constrained.

There’s another wrinkle the bears keep missing: hardware upgrade cycles have collapsed from five years to roughly twelve months. Nvidia’s roadmap — from Hopper to Blackwell to the upcoming Vera Rubin architecture — has forced hyperscalers into a perpetual upgrade treadmill. The Rubin GPU, shipping late 2026, promises a 10x reduction in token cost. If Google moves to Rubin and slashes its AI operating costs by 90%, Microsoft and Amazon have no choice but to follow or risk being structurally uncompetitive.

So while traders are panic-selling AI supply chain stocks on “peak capex” fears, the actual data — $240B in locked-in backlog, inference demand accelerating, 12-month hardware cycles — points to sustained spending for years. The market is pricing in a cliff that the fundamentals say doesn’t exist.

When markets misprice a structural shift this badly, the opportunity tends to show up in the companies closest to the spending. Right now, those stocks are on sale.