The number-one supplier of high-bandwidth memory and the primary HBM source for NVIDIA. It has sold out its entire 2026 output and captures the largest slice of the AI-memory scarcity rent, yet still trades like a Korean memory-cyclical.
Our Position
Research
The verdict
Captures the LARGEST slice of the HBM scarcity rent (~62% share, ~2/3 of NVIDIA's HBM4, 72% op margin, ~61% ROE) yet is valued like a Korean memory-cyclical (~6–9× forward) — and management's own $14B US listing is an explicit re-rating bet. Best HBM rent-capture in the cluster. BULLISH / MEDIUM-HIGH / 1Y.
World's #2 memory maker by revenue and #1 in HBM — the stacked-DRAM that gates every AI accelerator. DRAM ~70–75% of revenue (incl. HBM, server DRAM, DDR5, LPDDR), NAND/eSSD ~20–25%, HBM the margin engine within DRAM. Memory crown jewel of SK Group (SK Square holds 20.07%; Chairman Chey Tae-won). Defining fact: sold out its entire 2026 output — DRAM, NAND, HBM — to customers including NVIDIA, tight into 2027, contracted a year-plus ahead with prepayments — a structural inversion of normal memory economics. NVIDIA is the anchor; primary HBM supplier since Hopper/Blackwell.
A double chokepoint — gated upstream by ASML/TSMC, itself the binding constraint downstream on NVIDIA. Upstream: ASML EUV (~$8B of tools committed by 2027); for HBM4 the logic base die moved off SK Hynix's DRAM process onto TSMC's logic node (a formal "One Team" alliance — HBM4 16-Hi uses a TSMC 12nm base die), coupling SK Hynix's roadmap to TSMC capacity + CoWoS. Downstream: NVIDIA → hyperscalers; HBM must match TSMC CoWoS interposer capacity, so the effective ceiling is the intersection of SK Hynix stacking yield and TSMC packaging. Chokepoint stack: HBM sev 9, CoWoS sev 9, EUV sev 7.
Lead is real; durability rests on yield and trust, not architecture. HBM share ~58–62%, Micron now #2 (~21%), Samsung third (~17%). NVIDIA Vera Rubin HBM4 allocation: SK Hynix ~60–70%, Samsung mid-20s%, Micron ~20%. The twist: Samsung passed NVIDIA's HBM4 qual and mass-produced first (Feb 2026) while SK Hynix was still optimizing to clear 11 Gb/s — yet SK Hynix kept the dominant allocation, because the moat is yield-at-volume + a multi-generation NVIDIA relationship (Samsung's HBM4 on 1c DRAM reportedly yielded ~50%), which is more durable than a process lead. Bargaining power: sold-out + prepaid + ~20% HBM3E price hikes. Vulnerability: a 2/3-vs-1/4 split, not a monopoly — Samsung's qual win is the wedge; allocation can shift at the margin.
DRAM (incl. HBM) ~70–75% ↑; NAND/eSSD ~20–25% ↑ sharply (fully booked 2026); HBM within DRAM ↑↑ but not separately disclosed (n/a) — HBM3E 12-Hi 36GB ≈ ~2/3 of 2026 HBM shipments, HBM4 16-Hi 48GB ramps H2→Q4 2026. HQ Korea (Icheon/Cheongju); China NAND fabs (Wuxi/Dalian); new Yongin (KR) + planned Indiana (US) packaging.
Record. Revenue ₩52.58T (~$35.5B) — first quarter ever above ₩50T, +60% QoQ, +198% YoY. Operating profit ₩37.61T, op margin ~72% (record); net ₩40.3T. Capex to rise "significantly" above 2025's ~₩30.2T — analysts model ~₩50T (+20–30%), prioritizing M15X, Yongin, ASML EUV. Stock reaction muted on the day; the violent move came in early June (Lens 8).
Tone escalated from "tight" to "structurally short into 2027/2030." Threads: (1) "sold out" — DRAM+NAND+HBM booked through 2026, demand exceeding capacity for three years; (2) HBM4 + the TSMC One Team pivot, HBM4E mass production 2027, move into client-specific custom HBM; (3) heavy capex with a balance-sheet guardrail (goal: net cash >₩100T while raising returns) — "fund the ramp from the cash flush." CFO: memory has entered "a new paradigm."
| Co | Ticker | Mkt cap (USD) | EV/EBITDA | P/E (TTM) | Fwd P/E | Div yld | ROE |
|---|---|---|---|---|---|---|---|
| SK Hynix | 000660.KS | ~$880B–$1.0T (crossed $1T late May) | ~15.7x | ~18x | ~6–9x | ~0.2% | ~61% |
| Micron | MU | ~$450–460B (Mar ref) | ~26x | n/a | ~8.5–9x | ~0.1% | n/a |
| Samsung | 005930.KS | ~$1.30T | ~6.6x (memory SOTP proxy) | n/a | n/a | ~4.7% | n/a |
| Nanya | 2408.TW | ~$38B | n/a | n/a | ~6x | n/a | n/a |
| ; EV/Sales n/a; Samsung figures are conglomerate/SOTP, not clean. The comps read (core thesis): the structurally-short oligopolist trades at ~18× trailing / single-digit forward — the scarcity rent shows up in earnings (72% margins, 61% ROE), not the multiple. That gap is exactly why management is pursuing a US listing (Lens 9). |
High-beta memory-cycle + AI-narrative + Korea-macro instrument. All-time high ₩2,407,000 on Jun 2, 2026 (cap ~$100B → $1T in ~16 months). Jun 6, 2026: −9.92% (≈ −12% over five sessions) on a soft Broadcom outlook that spooked the AI-chip complex — note it fell ~10% on a Broadcom wobble, not even NVIDIA, showing how tightly coupled the equity is to one demand narrative. 2022–23 trough: swung to a widening operating loss, then re-rated violently on HBM/AI. Catalyst taxonomy: memory-pricing/contract data, NVIDIA HBM4 allocation headlines, Samsung/Micron qual news (share-shift fear), Korea macro.
CEO Kwak Noh-jung; CFO Kim Woohyun/Kim Jae-joon (romanization varies — a hygiene note). Controlled via SK Square (20.07%) under the SK Group pyramid; Chairman Chey Tae-won drives strategy (at Computex 2026 pledged to double capacity in 5 years, citing shortage through 2030). Foreign ownership ~56%. The headline governance event: a confidential US (Nasdaq) listing, H2 2026, ~2–3% float, raising ~$14B, with the explicit stated rationale of forcing a re-rating by putting SK Hynix in direct comparison with Micron* — the trade thesis voiced by the company. Red flags / chaebol tells: investor groups oppose the dilutive new-share issuance (want buybacks); a ₩100T shareholder/employee reward "burden"; SK pyramid cross-shareholding governance discount.
(1) Peak-cycle accounting — 72% op margin / 61% ROE are cycle-peak; memory mean-reverts violently (losses in 2023); forward P/E ~6–9× is the market saying it won't last; watch inventory build + contract-price roll-over (Q1 balance-sheet detail n/a — pull primary). (2) NVIDIA concentration — HBM ~2/3 to one buyer's roadmap; the Jun-6 −10% on a Broadcom wobble shows the coupling. (3) HBM not separately disclosed — the #1 value driver is modeled, not reported. (4) Capex vs durability — ~₩50T + ~$8B EUV + Indiana/Yongin underwritten by "sold out 3 years"; the net-cash guardrail mitigates, but a cycle turn mid-build is the canonical memory trap. (5) Chaebol/related-party — SK pyramid, the ₩100T reward, dilutive listing opposed by minorities.
Bottom-up; calendar-year; KRW. Q1'26 already ~₩40T net at ~77% net margin annualizes high — dispersion is the memory signature.
| Scenario | 2026 rev / EPS | 2027 | 2028 | Assumptions |
|---|---|---|---|---|
| Base | ~₩210T / ~₩290k | ~₩250T / ~₩340k | ~₩220T / ~₩240k | sold-out 2026 holds; HBM4 share ~60%; 2028 mild softening as HBM4 capacity broadens |
| Bull | ~₩220T / ~₩310k | ~₩300T / ~₩430k | ~₩340T / ~₩500k | HBM4/HBM4E lead holds; custom-HBM premium; US listing re-rates multiple |
| Bear | ~₩200T / ~₩260k | ~₩180T / ~₩150k | ~₩150T / ~₩70k | Samsung yields converge → allocation shifts; 2027 HBM glut; Korea macro/won |
Log: npx tsx scripts/research/forecast.ts create --topic hardware --question "SK Hynix 2026 op margin >= 65%" --p 0.6 --resolves 2027-02-28 --tags sk-hynix,deep-dive |
Bull. SK Hynix captures the single largest slice of the AI scarcity rent — ~62% of HBM, ~2/3 of NVIDIA's HBM4, sold out and prepaid into 2027, at 72% operating margin and ~61% ROE — and the market still values it as a Korean memory-cyclical at ~6–9× forward. Management is so convinced of the mispricing it is taking the company to Nasdaq specifically to force the re-rating. That is a rare alignment: the best fundamentals in the cluster and the most explicit, dated re-rating catalyst. Bear. (1) The lead is yield-and-trust, not a monopoly — Samsung out-qualified it on HBM4 timing and got ~1/4; if Samsung's yields converge, NVIDIA wants a credible second source and allocation shifts. (2) ~2/3 of HBM rides one customer's roadmap. (3) It's a memory cyclical at peak margins, with a chaebol governance discount the US listing may not erase. Pre-mortem (2027): Samsung's HBM4 yields hit parity, NVIDIA rebalances Rubin/Feynman allocation toward two credible suppliers, and a capex pause coincides with HBM4 capacity from all three landing — SK Hynix's margins fall from 72% to 40%, and the US listing prices at the old cyclical multiple anyway. Contrarian view: the market is fighting the last memory war (assuming mean-reversion) while the contract structure + the One-Team TSMC coupling have made SK Hynix more like a contracted infrastructure supplier than a spot-price cyclical — but unlike consensus, the variant view here is also that the US listing is the catalyst that forces the multiple to admit it.
The 2/3 allocation is not a moat — it's a procurement decision NVIDIA re-makes every generation, and Samsung already passed the qual. A −10% day on a Broadcom outlook proves the equity is a leveraged bet on one customer's capex, dressed as an oligopoly. The chaebol will steer capital to group priorities; the dilutive US listing (opposed by minorities) and the ₩100T reward "burden" both transfer value away from the shares. 72% operating margin is the most mean-reverting number in hardware. If AI capex disappoints 20–30% while HBM4 capacity floods in 2027, op margin halves and the "structural" story reverts to "Korean cyclical at the top of the cycle" — exactly what the 6–9× forward multiple always said.
Number-two in high-bandwidth memory and the only US maker of both DRAM and NAND.