The Memory Supercycle
The memory supercycle is the shift of DRAM and NAND from a boom-bust commodity cycle into a structural growth layer of AI infrastructure. High-bandwidth memory (HBM) sits beside the accelerator rather than behind it, consumes far more wafer capacity per bit than conventional DRAM, and cannot be brought on fast enough to match demand — which is why the three companies that make it are now among the market's most consequential earners.
What it is
Memory used to be the textbook commodity cycle: prices rose, fabs added capacity, supply overshot, prices crashed, and the cycle repeated on a two- to three-year clock. AI breaks that pattern for one segment of memory. HBM is stacked vertically next to the accelerator and wired directly into it, so it behaves less like a commodity input and more like an architecture component — the accelerator cannot run faster than the memory feeding it allows. HBM production also eats a disproportionate share of wafer capacity relative to conventional DRAM, and building it requires complex die-stacking and packaging steps that ordinary DRAM lines do not need. The result: when demand rises, supply cannot follow at commodity speed, and the shortage shows up directly in the earnings of the three companies that make it.
Why it matters
The scale of the shift is now visible in the numbers. Samsung's chip division flagged that 2026 alone should out-earn the division's entire roughly 40-year chip history — brokerages track around ₩320–330 trillion for 2026 against under ₩300 trillion cumulative from 1985 to 2025, a single year rivaling four decades. Samsung's Q2 2026 operating profit came in at ₩89.4 trillion (roughly $59 billion), up 19x year over year, briefly making it the most profitable company in the world for the quarter — ahead of Nvidia. The stock still fell 7% on the print, a reminder that a single quarter's number, however large, does not settle the market's debate about how durable that earnings level actually is once new supply eventually arrives.
The three ways in
The supercycle is showing up differently across the three producers. Micron reached a $1.12 trillion market cap in July 2026, roughly 10x its 52-week low, while still trading near 6.6x forward earnings. SK Hynix, the current HBM share leader, trades at 4.8x versus an industry median near 29.8x (LSEG) — a valuation gap the market has started calling the Korea discount. SK Hynix underlined its own scale in July 2026 with a $26.5 billion Nasdaq ADR listing at $149 per ADR, the largest US listing by a foreign company, surpassing Alibaba's 2014 debut. Per Futurum Group, SK Hynix's HBM share is expected to move from roughly 57% in 2025 to roughly 50% in 2026 and into the low 40s over time as Samsung and Micron gain ground — but the binding constraint through the decade looks like total capacity, not who owns which slice of it.
How to read it
The supercycle framing does not mean the cycle risk has disappeared — memory has surprised bulls and bears before, in both directions. It means the demand driver has changed character: HBM is now consumed as a fixed architectural input to every new accelerator generation rather than as a discretionary commodity purchase, and the industry's own announced capacity expansions do not obviously cover the demand path through the rest of the decade. That is a different question from "is the price high" — it is "can supply catch up," and on the evidence so far this cycle, it has not, which is why the three producers keep printing numbers that look structural rather than cyclical, and why the debate has shifted from timing the top to sizing the duration.