The AI bubble already burst. It burst in the stocks nobody quotes.

Oracle closed Friday at $137.10. Its highest close was $328.33, on September 10, 2025 — which means one of the largest companies in the world has lost 58.2% of its value in a little over a year. Super Micro is 63.6% below its high. The biggest nuclear-power utility in America is down 34.8%, and the company that makes the cooling systems data centres run on is down 32.7%. Three days before we pulled these numbers, the Nasdaq 100 closed at a record.

That is the whole puzzle of this market in one paragraph. If you have been reading that the AI trade is a bubble waiting to pop, the data says something more specific and much more useful: a large part of it already popped, roughly sixteen months ago, and the index hid it from you. What is left standing is a handful of companies that sell chips for cash. What broke is everything that had to raise money to build something.

We will show you the ladder, then the one number that explains it, then the government series that says the thing everyone expects to happen next is not happening.

Start here: what “down from its high” actually measures

Almost every number in this post is a drawdown. It is the simplest measure in investing and the one most often quoted wrongly, so here it is in full:

• Take the highest price a stock has ever closed at in the period you care about. Closing, not the highest price it touched for four seconds during the day — an intraday tick is not a price anybody actually got.
• Compare today's close to it. If a stock's best close was $100 and it is $58 now, it is 58% off its high. Nothing more complicated than that.
• The convention is that −20% or worse is a “bear market” in that thing. It is an arbitrary round number, but it is the one everyone uses, so we will use it too.

Two things make a drawdown honest, and both matter here. First, each stock is measured against its own high on its own date — not against a shared calendar date, which would quietly flatter whichever stock happened to peak latest. Second, a drawdown says nothing about whether something is cheap. A stock can be 60% off its high and still expensive; that is exactly what a deflating bubble looks like on the way down.

The ladder

Fifteen stocks and one sector fund, everything anybody would call part of the AI trade, each measured against its own best close. Sorted worst first.

Now look at the dates in the appendix, because they are the part nobody reports. Nvidia's best close was May 14, 2026. Broadcom's was June 2, Arm's June 18, Micron's June 25, TSMC's June 30. Oracle's was September 2025 and Super Micro's was March 2024. Of these fifteen stocks, fourteen had already set their high before the day the Nasdaq 100 set its. The index made a record on September 22 without a single one of Microsoft, Nvidia, Alphabet, Amazon or Broadcom being anywhere near its own.

The split is not random. It is a split between builders and sellers

Group the same fifteen by what they need from the outside world, and the picture sharpens into something you could have predicted from first principles.

• Build it. Companies whose money goes into a physical asset — a data centre, a power plant, a cooling plant, a server fleet — that starts earning years later. Oracle, Super Micro, Constellation Energy, Vertiv, Arm, and the utilities that sign decade-long contracts to supply the power.
• Sell the chips. Companies that ship a component and get paid for it now. Nvidia, AMD, TSMC, Micron, Broadcom.
• Hyperscalers. Microsoft, Alphabet, Amazon, Meta — doing both at once, out of cash flow they already have.

That gap is the finding. The market has not stopped believing in AI; it has stopped paying up for anything whose payoff is a decade away. Which brings us to the number that does that.

The one number: a real interest rate at a three-year high

Three yields matter here, and it is worth ninety seconds to separate them, because the whole argument turns on the difference.

• The nominal yield is what a 10-year Treasury pays: 5.18%. It is the headline number.
• The breakeven is how much inflation the bond market expects over those ten years: 2.34%. It is derived from the gap between ordinary and inflation-protected bonds.
• The real yield is what is left — what a lender earns after inflation. It is quoted directly by inflation-protected Treasuries: 2.85%. This is the actual price of money.

Here is why that single line wrecks the build-it group specifically. A data centre that earns nothing for four years and then earns for twenty is a long stream of distant dollars. Discount that stream at 1.78% and it is worth one number; discount it at 2.85% and it is worth meaningfully less — and the further out the money is, the harder the cut. A chip sold in this quarter for cash barely notices. Exactly the same rise in the discount rate does very different damage depending on how far away the money is.

You can see the same force everywhere else it should appear. Long Treasuries (TLT) closed Friday at a four-year low. Gold, which competes directly with a real yield, is 20.7% below its January high. High-yield bond funds hit a one-year low the same day. None of those are AI stories. They are all the same story.

We wrote about the level of yields on its own in the stock market looks calm, the bond market says it shouldn't be. This post is the sequel: where in the stock market that yield actually landed.

The control: this is not fear

The obvious objection is that we have described a crash and called it a discount rate. So here is the check. If markets were frightened about AI debt going bad, the price of credit would say so — and it is the single easiest thing to look up.

So no: the bond market is not braced for a wave of defaults, and the options market is not braced for a crash. And that resolves something that looks alarming and isn't. High-yield bond funds hit a one-year low on Friday, but not because credit got riskier — because Treasury yields rose underneath them. A bond fund can fall purely because the discount rate moved, with default risk completely unchanged. The spread is the part that measures fear, and the spread is near the bottom of its range.

Which makes this an unusually clean natural experiment. No panic, no credit event, no inflation scare — just a full extra point of real interest, and a market that re-sorted itself by how distant each company's cash flows are.

What the index was hiding

One more angle on the same thing, because it tells you how little the index is telling you. Four baskets, all indexed to November 2022 — the month ChatGPT was released, which is the only sensible start date for measuring “the AI trade.”

Note where each basket sits relative to its own best day, on the same Friday: the Nasdaq 100 is 0.4% below its high, the S&P 500 0.8% below, and the equal-weight S&P 500 is 5.2% below its. The more you weight toward the biggest companies, the better the market looks — which is another way of saying the market is a few companies now, and an index is the last place you would notice.

The part that should change your mind: the spending is accelerating

If this were a genuine AI bust, the buying would be stopping. The national accounts say the opposite, emphatically. Here is what American businesses spend on computers and peripheral equipment — the actual hardware that fills a data centre — at an annual rate.

The wider measures agree. Investment in information-processing equipment and software is $1.61 trillion annualised, up 19.2% in a year. All non-residential equipment investment is up 14.8%. American semiconductor production is up 12.4% and within a whisker of its record. Whatever is going on in the share prices, the machines are still being bought.

One series does break, and it is worth naming because it cuts the other way: construction spending on manufacturing plants is $169.8 billion annualised, down 21.2% in a year and lower in six straight months. The factory-building boom is over. The data-centre-filling boom is not.

What this means for your money

If you own an index fund, you own the narrow version of this market. The gap between SPY and its equal-weight twin is 3.67 times of return over four years. That has been wonderful and it is also a concentration risk you did not choose. The honest read is not “sell”; it is that your index fund is a more aggressive bet on a handful of companies than it was in 2022, and you should know that before the next repricing rather than after.

“Picks and shovels” was the advice, and it was the worst place to be. For three years the standard line was that you did not have to pick the AI winner — just own the infrastructure everybody needs. That basket is down 39.3% on average. The companies that looked most speculative in 2023, the ones just selling chips, are the ones nearly unscathed. It is worth sitting with how completely that inverted.

A 2.85% real yield is a competing offer, and it is the highest in three years. That is the mechanism behind every chart here, and it applies to your own decisions too: a government bond now pays you 2.85% above inflation for ten years with no business risk at all. Every risky asset has to beat that, and a year ago it only had to beat 1.78%.

Your power bill is the part of this that reaches your kitchen. Electricity prices are up 3.8% over the past year against 3.4% for prices generally. That is a small gap today, and we are not going to inflate it into a crisis. It is the one line in the consumer price index where a data-centre build-out would show up in a household budget, and it is worth watching precisely because it is still small.

If you sell options, dispersion like this is the whole environment. A market where the index is calm at a VIX of 14 while individual names swing 30% is the definition of cheap index volatility and expensive single-stock volatility. That is not a trade recommendation — it is why our screen checks implied against realized volatility per name rather than trusting a market-wide reading, and why it refuses to treat a calm index as a calm portfolio.

What would change this read

Four honest caveats. A thesis without them is a sales pitch.

• We chose the fifteen names. There is no official list of “AI stocks,” so any basket is a judgment call and a different one would give a different average. We have published every constituent, its price, its high and the date of that high in the appendix, so you can drop the ones you disagree with and recompute. The median drawdown, 12.1%, is less sensitive to our choices than the mean of 20.7%, and both are in the chart.
• A drawdown is not a valuation. Nothing here says the fallen names are cheap or the standing ones are expensive. Oracle at 58% off its high could be fairly priced, or still too dear. We are measuring what the market did, not grading it.
• The discount-rate story is an inference, not a proof. It fits unusually well — long bonds, gold, utilities and the whole build-it group moving together while credit spreads stay tight — but we cannot rule out that each of these companies fell for its own unrelated reason in the same four months. The test is forward-looking and specific: if the 10-year real yield falls back toward 2%, the build-it group should bounce hardest. If it doesn't, we were wrong about the mechanism.
• The spending figures lag badly. Q2 2026 is the newest quarter published, so the hardware numbers end in June — before the September move in yields. If the Q3 release shows that $400 billion rolling over, this stops being a financing story and becomes a demand story, which would be considerably worse and would invalidate the most reassuring chart in this post.

Data appendix

Every figure in this post, with the date it belongs to. Each source is free and public, so you can check us — and recompute the averages with a different list of names if you disagree with ours.

Want to watch the same series we do? The market radar tracks the live version of several of these, and Learn explains the instruments from scratch.