The Thesis

Samsung’s preliminary Q2 2026 results confirm the memory supercycle thesis in full: operating profit of roughly 89.4 trillion won (about $58.4 billion) surged 19-fold year-over-year and beat consensus estimates of around 86 trillion won by roughly 6%. Revenue more than doubled year-over-year to 171 trillion won. On operating income, Samsung has now posted the highest quarterly profit ever recorded by a technology company, ahead of Nvidia’s most recent quarter. And yet Samsung shares fell as much as 6.8% in Seoul on the news — a reminder that in this cycle, “beat” and “priced in” are not the same thing.

The joint venture just did the thing it does best: ship a genuine engineering advance and wrap it in a press release that says more than the underlying event quite supports. Kioxia and SanDisk announced the “start of production” of their 10th-generation 3D flash — BiCS10 — at the K2 fab in Kitakami. The technology is real and competitive. The framing is doing some work. And the business underneath it is the most cyclical, least-cushioned corner of the memory complex, which matters a great deal for how much of this belongs in a SanDisk valuation that has already run over 750% this year.

The memory trade finally blinked. Micron and SanDisk each fell roughly 10.6% on Wednesday, July 1, with Western Digital and Seagate dropping 6.3% and 5.2%. Thursday brought a second leg down: SanDisk lost another 11%, Seagate 7%, Micron 4%. The Roundhill Memory ETF (DRAM) — the cleanest sector proxy, launched only in April — shed nearly 11% Wednesday and another 5% Thursday. The Philadelphia Semiconductor Index posted a 7.9% weekly decline, its worst since April 2025.

Micron broke ground this week on a roughly $9 billion HBM fab inside its existing Hiroshima campus, with first shipments targeted for the summer of 2028. Strip away the ribbon-cutting and the strategic logic is genuinely sound: HBM is the most constrained component in the AI supply chain, Micron is the number-three player trying to close the gap on SK Hynix and Samsung, and the Japanese government is covering a large slice of the bill. Every part of that is defensible. The problem isn’t the decision — it’s the arrival date. This capacity lands in 2028, which is precisely the year the supply-glut argument that drove this week’s memory selloff says the cycle rolls over. The same event is the bull’s bottleneck-reliever and the bear’s Exhibit A, and which one it becomes won’t be knowable for two years.

The memory bull case rests on a single assumption, and it is worth stating plainly because everything else follows from it: every incremental dollar of AI compute requires proportionally more high-bandwidth memory. GPUs pair with HBM, HBM is scarce and expensive, and that scarcity is precisely what handed Micron and SK Hynix gross margins near 85% and market caps north of a trillion dollars. If the assumption holds, memory demand scales with the AI buildout indefinitely. The SRAM wildcard is the possibility that a meaningful slice of AI demand quietly stops needing the memory these companies sell.