AI Infrastructure Stocks: Where Each Sits in Its Own Valuation Range

As of October 5, 2026. Growth is trailing-twelve-month revenue from our dataset unless marked. † CoreWeave listed in 2025, so its percentile compares against 32 weeks, not five years. ‡ Dell's five-year window includes its 2021 VMware spin-off, which makes older readings less comparable. Our data-quality monitor flagged Broadcom's and Astera Labs' figures against an independent source, so we show their multiples but not a rank. IREN has 22 weeks of history, short of the 26 a rank needs. TSMC, ASML and Nebius file annual reports as foreign issuers and Tokyo Electron lists in Tokyo, so we have a current snapshot but no weekly history.

July's crash, and what came after

The Philadelphia Semiconductor Index fell 21% in July, its worst month since October 2008, and erased $2.2 trillion of market value, according to Fortune's account of the July chip selloff. The fears were about demand peaking, cheaper open-source models, and a July 27 report that China had started producing its own immersion lithography machines. Nvidia and Broadcom rose that month anyway.

Then the buyers spoke. Within eight days of each other in late July:

• Alphabet raised its 2026 capital spending plan to $195 to $205 billion and said 2027 would rise significantly (Alphabet's second-quarter call).
• Microsoft said it expected more than $50 billion of spending in the September quarter alone, with fiscal 2027 higher than fiscal 2026 (Directions on Microsoft's call summary).
• Meta narrowed its 2026 range upward to $130 to $145 billion (Meta's second-quarter release).
• Amazon moved to about $220 billion from about $200 billion, citing higher memory costs (a summary of Amazon's second-quarter call).

The suppliers followed. Nvidia's second-quarter revenue was $96.2 billion, up 106%, with $108 billion guided for the next quarter. Broadcom's AI chip revenue rose 221%. In September, AMD and Intel each rose about 29% (Yahoo Finance on the September chip rebound). Gavin Baker argued during the selloff that no hard metric had turned negative; we went through each metric Gavin Baker cited and how to check it.

The takeaway: July priced in a slowdown that the spending plans and supplier results did not show. By October the valuations had mostly recovered with the prices.

Priced near the top: equipment, networking, AMD

Applied Materials, Lam and KLA started 2026 at the 100th percentile of their own EV/Sales ranges. In the week of the July low they were at the 95th to 97th. On October 5 they were at the 97th and 98th. Lam trades at three times its own five-year median; Applied Materials and KLA at between two and three times theirs. The crash moved their prices, not their place in their own history.

Arista, at the 98.5th, is the most stretched networking name, and on P/E it is at the very top of its five-year range too. AMD sits at the 99.6th on EV/Sales: its multiple went from 10.3 at the start of the year to 25.2. Marvell made a similar jump, from 9.4 to 26.7. Dell is at the 98th, with the spin-off caveat above.

Optics is the one place where the earnings caught up. Ciena's P/E percentile fell from the 99th in January to the 65th, and Coherent's from the 95th to the 55th, because profits grew faster than prices. Their EV/Sales ranks stayed at the 89th or higher. Toggle Ciena's valuation history between P/E and EV/Sales to see the gap.

The takeaway: Most of the supply chain trades at two to three times its own five-year median EV/Sales. The market is paying now for spending that is guided, not yet booked.

The low end: Nvidia, Super Micro, CoreWeave

Nvidia sits at the 20th percentile on EV/Sales and the 3rd on P/E. Its trailing revenue grew 83% while the stock rose much less, so the multiple fell: P/E is about 30 against a five-year median of 57. That is the same pattern we noted in August, in what Nvidia's lowest forward P/E in a decade does and does not mean, and it survived a recovery and a record $150 billion buyback authorization. A low rank after fast growth can also mean investors expect growth to slow, so it describes the present, not a forecast. Nvidia's EV/Sales history shows the full band. Super Micro sits at the 24th with 78% revenue growth. It sells servers on thin margins, so a sales multiple under 1 is normal for it; the rank says it is below its own norm. CoreWeave sits at the 22nd, but against only 32 weeks of history, so the rank moves easily. Its EV/Sales fell from about 9.8 in its first weeks of data, in March, to 5.8.

The takeaway: Nvidia is the clearest case of compression without deceleration in this group: revenue up 83%, P/E near the bottom of its five-year range. Super Micro and CoreWeave are low too, with more specific worries attached.

Neoclouds are the hardest layer to measure

Of the five neocloud and data-center names, two cannot be ranked (Nebius files as a foreign issuer, IREN is too new) and CoreWeave has less than a year of data. What we can measure is the debt. CoreWeave's second-quarter interest expense was $640 million on $2.6 billion of revenue, and it raised more than $10 billion of new debt in the quarter, according to Investing.com's summary of CoreWeave's second-quarter results. Its backlog was $104.2 billion. Pricing is moving their way for now: Nebius raised GPU rental prices by roughly 17% to 21% from October 1 (24/7 Wall St. on neocloud pricing).

The takeaway: For neoclouds, the valuation percentile matters less than the balance sheet. Watch funding costs and GPU rental prices.

The bear case

• The build-out runs on debt. The big cloud companies issued $219 billion of investment-grade bonds in 2026 through August 20, and their spreads widened about 30 basis points while the broad market barely moved, per J.P. Morgan Asset Management's note on hyperscaler credit.
• Permits and power. New York paused state permits for data centers of 50 megawatts or more in July (Davis Polk on the New York moratorium). Power is the slowest input to add; our AI power stocks ranking covers that side.
• China builds its own tools. The output is small for now, but it targets the mature chips that drove much of the equipment makers' China sales.
• No cushion. At the 97th percentile and above, a single disappointing guide can move a stock a long way, as July showed.

The takeaway: The risk is less about demand than about who pays for it and how fast it can be built. Debt, permits and power are the pressure points.

Signposts

• 2027 spending plans. Alphabet and Microsoft have said spending will keep rising. Full-year numbers usually arrive with the January and February reports.
• GPU rental prices. Rising prices support the neoclouds and Nvidia's customers. A sustained drop would be the first hard negative.
• Hyperscaler bond spreads. Further widening raises the cost of the whole build-out.
• Nvidia's percentile. If revenue keeps growing and the multiple stays low, the gap with the rest of the chain widens.

The AI infrastructure theme page groups all 25 names by layer with live percentiles and an equal-weight basket against SMH, and you can open the same 25 names in the screener. Memory and storage names, including Micron, are on the AI memory theme page. If band charts are new to you, read how to read a valuation band chart first.

---

Data: SEC EDGAR XBRL filings and market prices, computed by StockResearch as of October 5, 2026. Earlier-date percentiles were computed from the same weekly series using only the five years available at each date. Latest-quarter growth for foreign filers comes from market-data snapshots. Company figures come from the linked releases and reports. Jake is the founder of StockResearch.app, where he writes data-first research on valuation dislocations. Nothing here is investment advice or a recommendation to buy or sell any security. Do your own research.