Memory stock rotation scare: four facts behind the selloff

Tuesday's closing screen showed two different AI infrastructure markets: connectivity and custom-silicon stocks soared while memory and storage names were sold hard. Marvell lifting its long-term revenue target to $70-90 billion lit up the connectivity side as Western Digital fell 7% and Seagate 9%. At first glance it looks like money is fleeing memory. Yet the physical data on the ground says AI investment is not slowing at all; the bottleneck is simply changing address. This article walks through the four concrete developments behind the selloff and explains the difference between a capital rotation and a broken thesis.

The day's scoreboard captured the split cleanly. On the memory side Micron slipped 1.7% to around $1,045, SanDisk dropped hard, Western Digital lost 7% and Seagate 9%; SK Hynix ADRs fell 6.5% on Samsung earnings anxiety. On the connectivity side Marvell gained 6% to $287, Broadcom added 4%, Astera Labs 7.5%, Applied Optoelectronics 7%, Corning 6% and Arista 4%. The stock market is pricing the same AI spending two ways: connectivity as an unconstrained compounder, memory as a late-cycle commodity. That pricing produced a short-term rally in connectivity stocks . For investors the question is which side is backed by cash flow.

The catalyst was Marvell's investor day in New York. The company raised its fiscal 2028 revenue target to about $20 billion, clearly above the $18.2 billion consensus set compiled by LSEG. More striking was the fiscal 2031 band of $70-90 billion: the $80 billion midpoint is nearly double the $46.85 billion consensus from four analysts polled by Visible Alpha. According to the investor-day notes carried by ChannelNewsAsia, the company expects $12 billion in custom-circuit sales in 2029, up from a prior $10 billion goal. The shares have more than tripled this year, and the dollar-denominated valuation now demands flawless execution.

The most concrete support for that target is the Google contract. According to the August record reported by Reuters, Google took a warrant for 58.97 million Marvell shares at $206.58, worth about $12.2 billion, in exchange for co-developing custom AI circuits. If performance milestones are met, the deal could generate $120 billion in sales through fiscal 2033. As Reuters emphasized, Google could thus become one of Marvell's five largest investors. The agreement moves the company ahead of Broadcom in the custom-silicon race. The message for investors is plain: big cloud customers have opened their wallets to cut Nvidia dependence, and that money is now being priced in.

According to Fool analysis, Marvell's 2028 target has climbed through three straight updates: $16.5 billion in May, $18 billion in August, $20 billion now. Against $8.2 billion of sales in fiscal 2026, the 2031 band implies 8.5 to 11 times growth, aimed at a roughly $400 billion addressable market by 2030. As Fool pointed out, the stock trades at about 43 times estimated 2028 earnings; expensive, but reasonable if the targets hold. The pace of profit growth matters here: up 37% in the second quarter, with more than 50% expected this quarter. So the connectivity rally is fed not by an empty story but by upward-revised dollar guidance.

A multibillion-dollar race for TDK on the HDD front

The trigger for pressure on the memory side was TDK's magnetic-head business. According to Bloomberg's sourced reporting, Seagate entered a bidding contest worth several billion dollars against Japan's Toshiba for TDK's hard-drive read-write head unit; Toshiba sat down in the spring and Seagate showed up over the summer with a higher bid. TrendForce data puts 2026 share at 48% for Western Digital, 42% for Seagate and 10% for Toshiba, with TDK the only independent supplier serving all three. This account is drawn from TrendForce reporting and explains why the supply chain is so critical. Equities priced it harshly, because if head supply jams, the high-capacity drive ramp jams too.

The second leg of the selloff is regulatory risk. In a note from Wedbush analyst Matt Bryson, Seagate would struggle to win antitrust clearance, since third vendor Toshiba would be left solely dependent on a rival's technology. Toshiba's position is genuinely fragile: it relies entirely on TDK for magnetic heads, and its heat-assisted magnetic recording (HAMR) roadmap depends on that technology. Toshiba plans to double disk capacity by fiscal 2027 with a $300 million investment. Even though the Japanese company denied the story, the market keeps pricing the possibility; investors fear the supply-chain knot will show up on the balance sheet .

The third item is Micron clearing its legal overhang. Under the announcement carried by PRNewswire, Micron and Netlist signed a five-year license and settlement ending all suits: $30 million per quarter from the fourth quarter of 2026 through the third quarter of 2031, $600 million in total. The license covers server DIMMs and high-bandwidth memory patents, and Netlist chief C.K. Hong called the deal validation of its AI memory technology value. At $120 million a year, the bill is small next to billions in free cash flow. This detail comes from the PRNewswire release and removes an uncertainty discount from the shares.

Record profit at Samsung, but a margin question

The fourth item is Thursday's Samsung preliminary result. According to the expectations survey compiled by KoreaHerald, the company will report about 10.6 trillion won (roughly $7.9 billion) in third-quarter operating profit, nearly nine times last year and a fourth straight record quarter. But the LSEG SmartEstimate across 21 analysts has been cut 7.7% since August. This context comes from KoreaHerald reporting and sums up market unease. The shortage is thought to last into 2028, yet the pace of price gains has slowed; in earnings week investors now price not just growth but margin durability.

The margin worry has two structural roots. First, Samsung's weight in legacy PC and mobile DRAM plus NAND, where pricing is soft. Second, slower qualification and scaling of the highest-margin HBM3 and HBM4 bits versus SK Hynix and Micron. On top of that, long-term supply contracts cap price gains in exchange for guaranteed supply. TrendForce expects conventional DRAM contract prices to rise 10-15%; growth, but not the old frenzy. This forecast is drawn from TrendForce analysis. The lesson for portfolio managers: record profit and record margin are not the same thing.

Rotation is not substitution: two faces of one system

Through all this noise the host's core thesis holds: connectivity does not replace memory, it stands beside it. Google's custom accelerators and the big accelerator makers' racks want the same high-bandwidth memory, the same server DRAM and the same dense SSD storage. So Marvell's $80 billion midpoint describes a world where the system grows, not one where memory demand collapses. That is why Micron locking much of its 2027 output into high-margin long-term contracts matters. The shares closed with RSI near 25, deep in oversold territory, so a technical bounce is plausible. Post-Samsung pressure may persist into Friday, but the physical bottleneck is unchanged; memory stocks remain at the center of the AI system in investor portfolios.