Forward-looking commentary for a general audience, not personalised advice. Figures below are sourced to company IR / SEC filings, TrendForce, Counterpoint, and Yahoo Finance quotes; prices are as-of market close 4 Sep 2026 unless noted. Horizon: 2–5 years.
Lead
The market is pricing memory as if the AI infrastructure build is multi-year and binding. Operating margins in the 50–80% range are not “cyclical recovery” numbers — they are scarcity numbers. The tension is simple: if those margins mean-revert before CapEx pays for itself, optically cheap forward multiples become expensive overnight.
The cycle
This is an AI/HBM + server DRAM supercycle, not a broad PC/mobile bounce.
TrendForce (press, 30 Jul 2026) frames 2026 DRAM as still short (sufficiency roughly −1% to −2%), with the gap potentially widening in 2027 as demand outruns supply — driven by HBM wafer cannibalization, AI/server DRAM, and related form factors. NAND remains tight near-term in that same read, but sufficiency is expected to turn positive later, with downward price pressure possible in 2H27 as high-layer capacity and new fabs ramp. Consumer/PC replacement demand is the soft side of the ledger; AI/server is the hard side.
HBM revenue share in Q2 calendar 2026 (Counterpoint / Aju secondary reporting): SK Hynix ~50% / Samsung ~33% / Micron ~18%. Conventional DRAM in the same quarter: Samsung ~38% / SK Hynix ~25% / Micron ~24% / CXMT ~10%.
Micron’s management commentary (June 2026 call coverage) has been more bullish on both DRAM and NAND staying tight beyond 2027 than TrendForce’s NAND-easing path — a useful disagreement to monitor, not to paper over.
The business in plain English
All three sell the scarce ingredient that feeds GPUs: HBM and high-end DRAM, plus NAND. How you own the theme differs.
Access note: Yahoo’s SKHY market-cap print can be inflated by the ADR premium. Prefer Korea listing fundamentals for enterprise value sense-checks.
Quality scorecard
Comparability warning first: Micron’s latest print is fiscal Q3’26 (quarter ended 28 May 2026). Korean names’ latest prints are calendar Q2’26 (ended 30 Jun 2026). Currencies differ (USD vs KRW). Samsung’s consolidated margins dilute a semiconductor P&L that is currently carrying almost all group operating profit. Treat cross-name margin and P/E ranks as directional, not precise apples-to-apples.
SK Hynix (000660.KS / SKHY)
From company IR (29 Jul 2026, Q2’26 preliminary K-IFRS): revenue KRW 79.32T (+51% QoQ, +257% YoY); operating profit KRW 60.54T (operating margin 76%); net income KRW 93.92T. End-Q2 cash & equivalents KRW 88T, total debt KRW 18.6T, net cash KRW 69.4T. CapEx guided in the high KRW 40T range for 2026 with acceleration at M15X / Yongin.
Yahoo key-stats (as shown around the 4 Sep 2026 quote): trailing / forward P/E roughly 7.0 / 3.3; EV/EBITDA ~4.8x. Prefer IR for the balance sheet — Yahoo cash/debt fields conflict with the IR print.
HBM share ~50% in Q2’26 (Counterpoint). Management tone: AI infra demand structural; customer demand above supply; multi-year LTAs; HBM4 mass shipments began in Q2’26.
Micron (MU)
From SEC Exhibit 99.1 (FQ3’26): revenue $41.46B (+74% QoQ, +346% YoY); GAAP / non-GAAP gross margin 84.6% / 84.9%; operating margin ~80–81%; diluted EPS GAAP / non-GAAP $24.67 / $25.11; adjusted free cash flow $18.3B. Company cash/marketable/restricted context ~$30.2B against ~$5.7B debt (net cash on the order of ~$24B per company commentary).
Product mix cited on the call (secondary summaries of company remarks): DRAM roughly ~$31.3B (~76%), NAND ~$9.9B (~24%). Guidance for FQ4’26: revenue $50.0B ± $1.0B, gross margin ~86%, non-GAAP EPS $31.00 ± $1.00. Next print flagged around 30 Sep 2026.
Contract durability: coverage of the call describes 16 strategic customer agreements, with 14 of 16 representing on the order of ~$100B minimum revenue at contracted minimum prices over remaining terms, plus large customer deposits/commitments. HBM share ~18% in Q2’26 (Counterpoint). Strategy commentary emphasises keeping HBM share nearer DRAM share to limit wafer cannibalization.
Yahoo (4 Sep 2026 close ~$1,016.59): trailing / forward P/E roughly 21.7 / 6.2; EV/EBITDA ~15.6x. Wide consensus target dispersion on Yahoo (low hundreds to low thousands) is itself a signal of narrative disagreement, not a precision tool.
Samsung Electronics (005930.KS)
From Samsung IR (Q2’26): consolidated revenue KRW 171.5T; operating profit KRW 89.5T (margin 52.2%); gross margin 69.6%. Device Solutions (semiconductors) revenue KRW 127.5T and operating profit KRW 89.2T — essentially all of group OP in the quarter. Cash ~KRW 190T vs debts KRW 22.4T → net cash ~KRW 168T* (IR definitions).
HBM share ~33% in Q2’26, up sharply from Q1 (Counterpoint / Aju). Management: H2 still undersupplied in server DRAM, eSSD, and HBM despite mobile/PC moderation. Preferred shares (005935.KS) trade at a discount to common with lower liquidity — a separate expression, not a substitute for the thesis.
Yahoo (4 Sep 2026 common close KRW 255,500): trailing / forward P/E roughly 11.1 / 3.4; P/S ~3.4x (diluted by non-memory mix).
The differentiated insight
Three things a screener will not hand you cleanly:
-
Margin regime vs multiple regime. Forward P/Es in the low-to-mid single digits (Korea listings) and ~6x (Micron) look “cheap” because earnings power is at historic extremes. The correct question is durability of scarcity, not whether the multiple is below a historical average computed on mid-cycle earnings.
-
HBM share is not the same as memory P&L quality. Hynix has the purest HBM torque. Micron is buying durability with long-term agreements and a US listing. Samsung is reclaiming HBM share while you still own a handset/TV conglomerate whose device margins are pressured when memory BOM costs scream.
-
Expression risk on SKHY. The ADR can trade at a material premium to Korea with a hard conversion cap. US price action can diverge from Korea fundamentals without the underlying business changing. Prefer 000660.KS when the vehicle is the thesis; treat SKHY as a liquidity/access wrapper with basis risk.
What is priced in
Base. Multi-year HBM and server DRAM tightness continues. CapEx rises, but ASPs stay elevated enough that 2026–27 earnings power is not a one-quarter mirage. NAND may soften earlier than DRAM (TrendForce path).
Bull. Agentic AI keeps expanding HBM attach. NAND stays tighter than the 2H27 easing path. Long-term agreements keep pricing floors. Share winners in HBM4 compound free cash flow into fortress balance sheets without a classic crash. In that world, today’s “cheap” forwards still understate mid-decade cash generation.
Bear (steelmanned). CSP CapEx pauses or FCF stress hits. HBM packaging/yield misses. CXMT and new industry capacity bite conventional DRAM. NAND ASPs crack first. KRW strength hurts exporter optics for foreign investors. For SKHY, the ADR premium collapses toward Korea. Margins mean-reverting from 70–80% operating profit is a double hit — earnings and the multiple.
What would change our mind
Breaks the thesis: sustained ASP decline in HBM/server DRAM while CapEx stays high; credible evidence that AI server build is pausing for multiple quarters; export-control or China-fab shocks that strand advanced capacity without offsetting pricing power.
Changes the expression, not the thesis: HBM4 share rotating from Hynix toward Samsung; Micron SCA terms proving more valuable than incremental HBM share; NAND easing while DRAM stays tight — that reshuffles which name you want overweight inside the sleeve, it does not by itself kill AI-memory scarcity.
Relative ranking (by HBM torque / purity of expression)
- SK Hynix — purest scarcity expression; highest HBM leverage. Prefer Korea listing over chasing ADR premium.
- Micron — best liquid US pure-play; contract book is the durability tell. Extended prints argue for patience on timing, not for abandoning the theme.
- Samsung — fortress balance sheet and HBM recovery story, with conglomerate dilution. Better as mega-cap ballast than as max HBM torque.
Ranking by expected risk-adjusted return can differ from ranking by purity if entry price and ADR basis dominate — especially on Micron after a vertical run and on SKHY when the premium is wide.
What we are watching
- Micron FQ4 print / guide (~30 Sep 2026) — ASP, HBM revenue disclosure, SCA commentary.
- TrendForce / industry sufficiency updates — especially whether NAND truly loosens in 2027 while DRAM stays short.
- HBM4 share prints (Counterpoint and peers) — share rotation vs industry growth.
- CSP CapEx commentary from the hyperscalers — the demand-side falsifier.
SKHY premium vs 000660.KS — basis risk for US-only accounts.- China fab licensing / CXMT HBM progress — supply-side falsifier on the margin.
Bottom line
This is a scarcity trade on AI memory, not a value screen on mid-cycle semiconductor multiples. Among the three, Hynix is the purest HBM torque, Micron is the cleanest US vehicle (respect extension on timing), and Samsung is quality conglomerate ballast with a memory rocket underneath. We would rather be early on structural HBM shortage than late on peak-margin “cheapness” — and we would size the sleeve as if the cycle can still break.
Primary sources
- SK Hynix Q2’26 results: https://news.skhynix.com/en/q2-2026-business-results/
- Micron FQ3’26 Ex. 99.1: https://www.sec.gov/Archives/edgar/data/723125/000072312526000013/a2026q3ex991-pressrelease.htm
- Samsung Q2’26: https://news.samsung.com/global/samsung-electronics-announces-second-quarter-2026-results
- TrendForce DRAM/NAND sufficiency (30 Jul 2026 press)
- Counterpoint DRAM/HBM share (early Sep 2026 reporting)
- Quotes: Yahoo Finance
000660.KS, SKHY, MU, 005930.KS (close 4 Sep 2026)