Everyone Says AI Stocks Are On Sale. Our Data Says Five Are.
July 2026 took 40–60% off AI infrastructure stocks. We ranked 50 names against their own five-year valuation history: only five actually got cheap, and about a third of the complex still trades near five-year-high multiples.
July 2026 took 40–60% off AI infrastructure stocks in a month. The selloff narrative says the whole complex is on sale; the counter-narrative says the whole complex was a bubble finding fair value. We built a dataset to stop arguing from vibes: for 50 AI-infrastructure names (semiconductors, memory, networking, neoclouds, data centers, and the power chain) we computed trailing multiples weekly over ten years from SEC filings and market prices, then ranked today's multiple within each stock's own five-year range.
The result doesn't match either narrative. The "AI dip" is about five stocks. Much of the complex is still trading at the top of its own historical range.
If you only read one thing
The short version: July's selloff made almost every AI stock's price fall, but only a handful actually got cheap compared to their own past. Our data flags five: Nvidia, GE Vernova, Constellation, Vistra, and Super Micro (whose low price hides a business burning cash). Meanwhile the chip-equipment makers (KLA, Lam Research, Applied Materials) plus Eaton and Arista still cost about as much as they ever have. Cheap and falling are not the same thing.
The method, in one paragraph
A valuation percentile answers one question: compared to what this stock has cost over the past five years, what does it cost today? We compute TTM P/E, P/S, EV/Sales, and EV/EBITDA weekly from EDGAR XBRL fundamentals and split-adjusted prices (no look-ahead: a quarter only enters the series once it was actually filed), and rank the latest value in the trailing five-year distribution. 5th percentile = cheaper than ~95% of the stock's own last five years. It says nothing about whether the stock is cheap versus other companies, or whether the business deserves its history; those are the questions the number sets up, not answers. The full method is in our guide, what a valuation percentile is and how we compute it.
The takeaway: Every number here answers one question: is this stock cheap or expensive today compared to what it has cost over its own last five years? A low reading means near its own five-year low, not that the stock is a bargain versus anyone else.
What's actually compressed (and still growing)
| Ticker | P/E (TTM) | P/E 5y %ile | Δ vs 5y median | Rev TTM YoY | OCF/EV yield |
|---|---|---|---|---|---|
| NVDA | 34.3 | 5 | −45% | 71% ↑accel | 2.3% |
| GEV | 28.5 | 7 | −56% | 13% ↑accel | 5.5% |
| CEG† | 26.3 | 33 | −19% | 20% ↑accel | 3.6% |
| VST | 23.6 | 29 | −22% | 19% ↓decel | 7.2% |
| SMCI | 16.5 | 30 | −21% | 56% ↑accel | −26.9% |
Nvidia is the cleanest version of the story the bulls tell: fifth percentile of its own five-year P/E range, 45% below its own median, while trailing revenue growth is accelerating. Whatever you believe about AI capex, "the market is paying near a five-year-low multiple for accelerating earnings" is simply what the data shows. The surprise is who joins it: the power chain (GE Vernova, Constellation, Vistra) compressed alongside the chips, and Vistra carries a 7.2% operating-cash-flow yield on enterprise value per the August 7 readout. One honesty note: Constellation is the most marginal of the five. It sat at the 21st percentile when we first cut this data on August 7, has drifted to the 33rd since, and it is the row with the data-quality flag.
The takeaway: The data flags five names as actually cheap against their own history: Nvidia, GE Vernova, Constellation, Vistra, and Super Micro. Nvidia stands out most (near a five-year low on price versus earnings while sales growth speeds up), and Super Micro comes with a big warning: it looks cheap, but its operations are burning cash.
What isn't compressed at all
| Ticker | P/E (TTM) | P/E 5y %ile | Δ vs 5y median | Rev TTM YoY |
|---|---|---|---|---|
| KLAC | 52.7 | 95 | +91% | 12% ↓decel |
| LRCX | 54.1 | 94 | +127% | 26% ↓decel |
| AMAT | 50.7 | 98 | +145% | 3% ↓decel |
| ETN | 45.3 | 99 | +38% | 16% ↑accel |
| ANET | 59.5 | 96 | +37% | 33% ↑accel |
The takeaway: The companies that make chip-making equipment (KLA, Lam Research, Applied Materials) plus Eaton and Arista are still priced near the highest they have been in five years, even after July's crash. If you bought a basket of "AI stocks on sale," about a third of it was never on sale.
The middle, and the traps
Micron looks mid-range on P/E (47th percentile) despite 167% revenue growth, but its P/S sits at the 94th percentile, which is the market saying current earnings are cyclical peak, not trend. The July forced-liquidation names (CoreWeave, Nebius, IREN, Core Scientific) mostly have no meaningful P/E at all, and three of the four listed too recently to have a five-year rank; where the cohort can be ranked on P/S, none of it screens as historically cheap. The operating-cash-flow column (August 7 readout) splits the miners-turned-AI-hosts sharply: Core Scientific generates +5.6% OCF/EV while Riot and CleanSpark burn at −8% to −9%.
The takeaway: Micron looks reasonably priced against its earnings, but against its sales it is near a five-year high, a sign the market expects today's memory-chip profits to fade. And among the bitcoin miners turned AI hosts, Core Scientific brings in cash while Riot and CleanSpark burn it.
What this does and doesn't mean
A fifth-percentile multiple is not a buy signal; sometimes a stock is cheap against its own history because its future genuinely got worse. What the percentile does is separate two claims that July's tape blurred together: prices fell (true everywhere) and valuations compressed (true in a handful of places). The gap between those claims is where research starts. Every number above lives on its ticker page with the full ten-year band chart (NVDA's valuation history is the place to start), and the screen that produced this table runs live: the compression-without-deceleration screener.
The takeaway: Cheap against your own history is a reason to look closer, not a reason to buy; sometimes a stock gets cheap because its future really did get worse. What the data settles is simpler: prices fell everywhere in July, but only Nvidia, GE Vernova, Constellation, Vistra, and Super Micro actually got cheap by their own five-year standards.
Data: SEC EDGAR XBRL + market prices, computed by StockResearch. Percentiles and multiples were re-verified against the live dataset on 2026-08-10; operating-cash-flow yields come from our 2026-08-07 readout. Computed multiples are cross-checked against independent snapshots, with divergences flagged in-table. Nothing here is investment advice; do your own research.