The Storage and Memory Signal

Direction: Worsening.

This period: Three memory earnings reports in three days drew a sharp line between DRAM/HBM and NAND. SK hynix's CFO Kim Woo-hyun said October 28 that combined DRAM and HBM volume is "effectively sold out" through 2029, backed by 14 signed long-term agreements covering more than half of planned 2027 and 2028 capacity, on a quarter that cleared 100 trillion won in revenue for the first time at a record 80% operating margin. Samsung's October 29 results confirmed 2027 HBM4 supply is fully sold out at agreed pricing and raised 2026 capital expenditure 60% to a record 84.4 trillion won, with management expecting the shortage to persist through 2027 and possibly into 2028. Kioxia told a different story on October 30: NAND average selling price growth decelerated to the mid-20% range quarter over quarter from 70% the prior quarter, fourth-quarter guidance missed consensus, and shares fell roughly 11%. TrendForce's own October 28 data backs that up: the 512Gb TLC NAND wafer spot price fell 9.19% week over week, the sharpest one-week drop this newsletter has logged, as Chinese phone makers cut prices for Double 11 promotions.

Driver: Unchanged since Issue 03 for DRAM and HBM: HBM production still eats disproportionately into capacity that would otherwise make conventional DRAM, and that tightness is now locked into multi-year contracts, not just reflected in spot pricing. NAND doesn't have that same HBM-driven capacity drain, which is why it's the first segment showing real cracks once a consumer demand pull-forward fades.

Watch for: Micron's fiscal Q1 2027 results, due December 23, the next test of whether Micron's own DRAM and HBM commentary lines up with SK hynix's and Samsung's multi-year sold-out claims, or drifts closer to Kioxia's deceleration warning.

Start with the number that stopped a lot of people this week: 80%. That's SK hynix's operating margin for Q3 2026, reported October 28, on revenue of 104.6 trillion won, the first time any quarter in the company's history, or in this newsletter's running count of chip-industry superlatives, has cleared 100 trillion won. Revenue was up 328% year over year and 32% quarter over quarter. Operating profit hit 83.3 trillion won, up 632% year over year. Those are not the kind of numbers a company posts while still fighting for allocation; they're the kind of numbers a company posts when it has already stopped worrying about demand. CFO Kim Woo-hyun said as much directly on the earnings call: combined DRAM and HBM volume is "effectively sold out" through 2029, backed by 14 signed long-term agreement customers covering more than half of planned 2027 and 2028 capacity. That is not a forecast. It's a count of signed paper (see this issue's Uplevel Highlight on what an LTA actually locks in).

Samsung's October 29 results confirmed the same read from the other side of the HBM market. Operating profit hit a record 107.4 trillion won, up 783% year over year, slightly ahead of the preliminary guidance this newsletter covered in Issue 12. Memory chip revenue alone soared 427% year over year to 140.7 trillion won. HBM4 sales more than tripled quarter over quarter and now make up 60% of total HBM revenue, and Samsung said its entire 2027 HBM4 supply is already sold out at agreed pricing. The company is backing that confidence with money: 2026 capital expenditure is going up 60% year over year to a record 84.4 trillion won, with 76.8 trillion won of that earmarked specifically for the semiconductor division. Samsung's own language extends the shortage call through 2027 and "possibly into 2028."

Then came Kioxia, on October 30, and the tone flipped. The company's non-GAAP operating profit of roughly 1.99 trillion yen matched consensus, but the number that moved the stock was the deceleration underneath it: NAND average selling prices grew in the mid-20% range quarter over quarter, down sharply from 70% growth the prior quarter. Fourth-quarter guidance of about 2 trillion yen in operating profit missed the roughly 2.32 trillion yen analysts expected. Shares fell about 11%. TrendForce's own spot-market data backs up what Kioxia is describing from the inside: the 512Gb TLC NAND wafer spot price dropped 9.19% week over week, the sharpest single-week move this newsletter has logged in fourteen issues, as Chinese phone makers cut prices for Double 11 promotions and pulled NAND sellers into more aggressive discounting alongside them.

None of this is actually a contradiction, which is the part worth sitting with. DRAM and HBM are tight for a structural reason this newsletter has repeated since Issue 03: HBM production eats disproportionately into the wafer capacity that would otherwise make conventional DRAM, and that reallocation doesn't reverse just because a quarter ends. NAND doesn't have an HBM-sized sibling cannibalizing its capacity. Its demand is still heavily tied to consumer cycles, phones and PCs stocking up ahead of promotions, then pulling back once the promotion hits. SemiAnalysis's framing this week, that combined memory spend is heading toward 30% of AI system cost this year and past 40% in 2027 (this issue's Signal section), is really a story about DRAM and HBM specifically. NAND's AI-driven demand case, the one SanDisk is leaning on with its KV cache bet (also this issue's Signal), is real but newer and smaller than DRAM's, which is exactly why it's more exposed to a plain old consumer-electronics down-cycle showing up in the same spot prices.

Seagate's fiscal Q1 2027 results, reported October 27, were the direct test this newsletter flagged in Issue 13: does Western Digital's blowout quarter reflect an industry-wide hard drive story, or a company-specific one. The answer is industry-wide. Revenue came in at $4.2 billion, up 61% year over year, with GAAP gross margin at a record 57.3% and GAAP diluted EPS of $7.59, up from $549 million in net income a year earlier to $1.76 billion this quarter. The company shipped 220 exabytes total, up 21% year over year, with 198 exabytes going to data-center customers, a record 90% of total volume. Nearline capacity, per management, is now allocated through calendar 2028, with some cloud agreements stretching to 2030. Guidance for the December quarter calls for $4.8 billion in revenue, plus or minus $100 million, and non-GAAP EPS of $9.20.

Run the same math this newsletter applied to Western Digital in Issue 13: revenue grew roughly three times faster than exabytes shipped, 61% against 21%. Seagate isn't selling meaningfully more drives than a year ago, it's getting paid a lot more for the drives it does sell, exactly the pricing-power pattern Western Digital showed with its own 22%-exabytes-versus-28%-price-per-terabyte split. Two separate hard drive makers posting the same lopsided growth math in the same week is a stronger confirmation than either number alone.

And yet Seagate's stock fell roughly 7% that same week, with Western Digital down 8%, SanDisk down 9% despite its own beat and raised guidance (this issue's Signal section), and Micron off 6%, all against a broader Nasdaq retreat of about 3% on October 30 and a resurfaced rumor about Toshiba's capacity ramp, the same oversupply worry this newsletter flagged back in Issue 11. None of that rumor has new substance behind it this week. What changed is sentiment, not supply. Four companies posted some combination of record revenue, record margin, or beat-and-raise guidance in the same five-day window, and the market sold all of them anyway. That's not a verdict on the shortage thesis. It's a reminder that an earnings beat and a stock move are now answering two different questions.