Samsung's 1,814% Profit Jump, Kioxia's 30x, and a Memory Sector That Sold Off Anyway
Samsung reported second-quarter revenue of roughly $118.1 billion, up 130% year over year and a shade under the $118.9 billion consensus, alongside operating profit of about $61.46 billion, up 1,814% against a $60.68 billion estimate. Revenue missed. Profit did not. The distance between those two facts is the entire memory cycle compressed into a single print: sales slightly more than doubled while operating income went up eighteen-fold, which is only arithmetically possible when the incremental bit sells at a price the cost base was never built to justify. That is not operating leverage in the ordinary sense. It is a shortage clearing through price with a fixed asset base underneath it.
Kioxia makes the same point from the other end of the balance sheet. A company that was cash-constrained in 2023, kept alive by a shareholder structure that spent years looking for an exit, is now looking at a roughly thirty-fold increase in operating income for the first quarter of fiscal 2026. Nothing about Kioxia’s fab footprint changed enough in three years to explain a 30x. NAND pricing did.
And on the same tape, memory and memory-adjacent equities were being sold. Disappointing results out of SK Hynix set off a session in which Sandisk fell more than 7.5%, Arm more than 6.5%, and AMD more than 6.5%. Two of those three names have no meaningful memory revenue at all, which tells you the selling was factor-driven rather than fundamental. The market was not repricing NAND demand. It was reducing exposure to a cohort.
The gap between record earnings and falling multiples is the oldest pattern in this industry and it is not evidence of investor confusion. Memory equities peak before memory earnings peak, reliably, because the earnings are a lagging print of a spot price and the multiple is a forward bet on the supply response. A stock trading at eight times a peak year is not cheap if the next year is a trough. The 1,814% number is precisely the kind of figure that historically appears one or two quarters before the derating, because it advertises the size of the profit pool that new capacity is being built to attack.
What is genuinely different this cycle sits in the wafer math rather than in the price. High-bandwidth memory consumes multiples of the die area per delivered bit compared with conventional DRAM, so every incremental HBM allocation removes conventional supply from the market rather than adding to it. A supply response that would normally arrive as a wave of commodity DRAM instead arrives as HBM capacity that keeps the commodity market tight. That is the structural argument for a longer cycle, and it is the only argument that matters, because everything else in the bull case is just extrapolated pricing.
The competitive positions inside this are not symmetric. Samsung is the only participant that owns memory, foundry, and advanced packaging in one company, which should be decisive and so far has not been, because the same integration that gives it breadth has kept its HBM qualification behind SK Hynix at the customer that matters. Its moat is scale and process breadth. Its problem is that in HBM the moat is the qualification, and qualifications are granted by buyers, not earned by capex. Kioxia’s position is narrower and cleaner: pure NAND, no HBM exposure, no diversification to dilute the price move, which makes it the levered expression of exactly one variable. Sandisk sits in the same box and got sold off anyway.
The name nobody in this session was pricing is CXMT, which closed at roughly $464 billion in market capitalization on July 28, above every Western memory franchise combined on some measures. A domestic Chinese DRAM maker with that equity value can fund capacity at a cost of capital no Korean or American incumbent can match, and it does not need Western customers to justify the build. That is the supply response, and it is being financed in a market that does not answer to the same return discipline.
The number to reconcile is the divergence between conventional DRAM contract pricing and HBM allocation share over the next two quarters. If conventional pricing holds while HBM share rises, the wafer cannibalization thesis is live and the derating is early. If conventional pricing softens while HBM ramps, the cycle already turned and the print was the top.