AI turned memory from a commodity into the scarcest part of a data center. Memory makers posted the biggest results in their history this year, and the stocks moved even faster than the revenue. We ranked ten memory and storage stocks against their own five-year valuation history as of October 5, 2026. Every ranked name except SanDisk sits above the 80th percentile of its own EV/Sales range. At the same time, the memory makers trade at 4 to 7 times next year's expected earnings. Those two numbers disagree, and the disagreement is the whole debate.
If you only read one thing
The short version: Memory is now more than 75% of an AI server's hardware cost, by one Google executive's count, and prices are still rising. The stocks priced it: Micron, Seagate, Everpure, NetApp, FormFactor, and Rambus all sit at the 90th percentile or higher of their own five-year EV/Sales range. But on forward earnings the memory makers look cheap, which means investors are paying up for revenue already reported while doubting that record profits last.
Why memory became the bottleneck
Every AI accelerator ships with stacks of high-bandwidth memory (HBM) beside it, and every AI server needs more ordinary DRAM and flash than the servers before it. Only three companies make HBM in volume: SK hynix, Samsung, and Micron.
The cost shows up everywhere. At SEMICON Taiwan in September, a Google supply-chain director said memory now makes up more than 75% of an AI server's hardware cost. Microsoft said about $25 billion of its record capital budget reflects higher component prices, chiefly memory. Even Zscaler, a security software company, told investors on its September earnings call that memory prices will keep its equipment spending high.
Prices are still going up. TrendForce expects fourth-quarter contract prices to rise another 10% to 15% for ordinary DRAM and 15% to 20% for NAND flash.
The takeaway: AI made memory the most expensive part of the server, and supply has not caught up. Buyers from Microsoft down to Zscaler are paying more for it.
How to read the numbers
Each percentile compares a stock with its own past. We compute trailing EV/Sales weekly from SEC filings and market prices, then rank today's value against the last five years. The 96th percentile means the stock was cheaper than today in about 96% of those weeks.
We use EV/Sales rather than P/E for this group on purpose. Memory profits swing from losses to records within a few years, so P/E reads lowest near the top of a cycle and goes missing near the bottom. Our post on why a low P/E is a warning sign in memory stocks shows the history. The general tradeoffs are in when to use P/E, P/S, EV/Sales, or EV/EBITDA.
The full table
| Ticker | Layer | EV/Sales | EV/Sales 5y %ile | Forward P/E | Revenue growth |
|---|---|---|---|---|---|
| P | Storage systems | 11.3x | 99.6 | 28.0x | 27% |
| NTAP | Storage systems | 6.0x | 99.6 | 18.0x | 12% |
| FORM | Test (probe cards) | 12.8x | 97.5 | 39.0x | 18% |
| MU | DRAM and HBM | 13.4x | 96.2* | 5.7x | 167% |
| STX | Hard drives | 16.1x | 92.5 | 14.7x | 34% |
| RMBS | Memory interface chips | 15.9x | 90.2 | 29.6x | 19% |
| WDC | Hard drives | 12.3x | 83.5† | 16.1x | 36% |
| SNDK | Flash | 12.9x | 52.0‡ | 6.8x | 175% |
| 000660.KS | DRAM and HBM | 6.6x | snapshot only | 3.9x | 257% |
| 005930.KS | DRAM and HBM | 3.4x | snapshot only | 3.9x | 130% |
Where these stand now
Current readings update separately from the article’s dated analysis.
Below its usual range
MU's P/E of 12.5x is lower than 70% of its weekly readings over the last five years.
Historically high
NTAP's P/E of 32.1x is higher than 99% of its weekly readings over the last five years.
Above its usual range
FORM's P/E of 73.2x is higher than 73% of its weekly readings over the last five years.
The memory makers: expensive on sales, cheap on earnings
Micron is the clearest case. Revenue in its September quarter was $54.2 billion, up 379% from a year earlier, and it guided to $61.5 billion for the next quarter. Its EV/Sales sits at the 96th percentile of its own range, about three times its five-year median of 4.3. Yet it trades at about 5.7 times forward earnings.
The Korean makers show the same split more sharply. SK hynix, the HBM leader, earned a 76% operating margin in its June quarter and trades at about 3.9 times forward earnings. Samsung, also at 3.9, sells phones and displays too, which helps explain its lower sales multiple. US investors can now buy SK hynix directly: its ADRs began trading on Nasdaq as SKHY in July.
A high sales multiple and a low earnings multiple at the same time is what you get when margins are at a record. Investors pay up for each dollar of revenue because each dollar is unusually profitable today, and they discount next year's earnings because they doubt the margins last. You can follow the run on Micron's EV/Sales history.
The takeaway: On sales, Micron is near the most expensive it has been in five years. On forward earnings, the memory makers look cheap. Which one matters depends on how long record margins last.
Storage: the slower businesses re-rated the most
The two storage-system vendors sit at the very top of their ranges. Everpure, the company formerly called Pure Storage, grew revenue 38% last quarter and signed a second top-five hyperscaler. NetApp grew 12% and still sits at the 99.6th percentile on every multiple we track. For NetApp, that is the market paying for the storage cycle rather than for its own growth rate.
The hard-drive makers sit a little lower, at the 92nd (Seagate) and 84th (Western Digital). Their position changed this month: on October 2, Toshiba said it would double its hard-drive capacity, and both stocks fell about 11% that day, as reported by ZeroHedge on the Toshiba announcement. We cover that in why HBM and commodity memory face different supply risks.
The takeaway: Everpure and NetApp are at the top of their own five-year ranges. NetApp's growth is modest, so most of its re-rating is the sector, not the company.
The suppliers are not cheaper
A common way to play a boom is to buy the companies that sell to everyone, so you do not have to pick the winner. In memory those are FormFactor, whose probe cards test HBM and DRAM wafers, and Rambus, whose interface chips sit on server memory modules from all three DRAM makers. Both are at the 90th percentile or higher. FormFactor, at 39 times forward earnings, is the most expensive name in the table on that measure.
The takeaway: The picks-and-shovels names priced the boom too. FormFactor and Rambus sit at the 97th and 90th percentile of their own ranges.
The bear case
- New supply is funded. Micron alone plans about $25 billion of capex in the first half of its fiscal 2027, close to its full fiscal 2026 total, per its September prepared remarks. China's CXMT and YMTC target new output in 2027.
- Demand outside AI is shrinking. Gartner expects PC shipments to fall 10.4% and smartphones 8.4% in 2026 as memory raises device prices.
- History. In the last downturn, Micron's fiscal 2023 revenue halved and it lost $5.8 billion.
Signposts
- Contract prices. The first quarter in which ordinary DRAM or NAND contract prices fall.
- HBM for 2028. Whether next year's HBM agreements are signed at higher or lower prices than 2027's.
- Cloud capex. Any large cloud company cutting its 2027 data-center budget.
- New fabs. Whether CXMT, YMTC, or the new Korean and Idaho fabs ship volume ahead of schedule.
The takeaway: The memory trade is priced for a long shortage on sales and for a short one on earnings. Contract prices will settle it first.
The AI memory theme page tracks all ten names by layer with live percentiles. The companies buying this memory are on the AI infrastructure theme page, and the power they run on is on the AI power theme page. New to percentiles? Start with what a valuation percentile is and how we compute it.
Data: SEC EDGAR XBRL filings and market prices, computed by StockResearch as of October 5, 2026. Forward P/E is consensus market data. SK hynix and Samsung figures are market-data snapshots in Korean won. 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.