Korea AI Memory Complex — Post-Crash Re-underwrite (Kr Ai Memory Sector)

Executive Summary & Action Plan

Verdict

SELL (Reduce, even at crash prices) — The −9% index day removed froth, not overvaluation. All three names still price peak-cycle margins as permanent: SK Hynix's record 71.5% operating margin — the highest any memory maker has ever printed — back-solves at the crash close to 2.4x our normalized free cash flow held in perpetuity, and even our bull-case DCF sits below the tape. No name in this complex earns a pick; the honest anchor is SK Hynix, where the street's lowest target (₩1.85M) now equals the crash close to the won.

Price ₩1,845,000 | Market cap ₩1,310T (~$877B) | Target ₩1,060,000 | Upside -42.5% IWANNAVY Fair Value (peak-margin-normalized) · price as of 2026-07-13 close · Street consensus ₩3,258,502 (n=37, reference) — we are 67% below Street because the street applies ~8.4x to PEAK-year EBITDA where memory history pays 4–6x; 94% of brokers rate it favorably on extrapolated record margins

Abstract

We re-underwrote Korea's AI-memory complex at Monday's crash closes with full cyclical discipline: peak-margin normalization on every name, Korean-broker cross-checks on every target, and today's HBM evidence embedded. The parabola is real but concentrated in the most cyclical line — conventional DRAM contract prices (industry revenue +81% QoQ in Q1), not HBM, whose contract prices decline through 2026 with the HBM4 ramp deferred toward 2027. SK Hynix earns ₩1.06M (log-median of a ₩740K mid-cycle DCF, 9x normalized EBITDA ₩1.35M, and 16x normalized FCF ₩1.06M): the crash pulled it into the historical peak-multiple band (~4.7x FY26E peak EBITDA) but leaves nothing for the mid-cycle. Samsung earns ₩150K — its 42.8% consolidated OM implies memory margins near 75%, half again above the 2018 peak — with the foundry option (2nm yield >60%, Tesla AI5) credited but unable to bridge the gap. Hanmi earns ₩61K: a 52x normalized-EBITDA multiple on a collapsed order book facing Hanwha undercutting and hybrid-bonding disruption. Ratings: all three SELL/Reduce; re-entry belongs at mid-cycle prices, and the 7/22 SK Hynix print plus Samsung's 7/30 HBM4 disclosure are the first honest tests.

Action Plan

Reduce into any post-crash squeeze — the 7/22 print against a washed-out tape is genuine squeeze fuel, which is a reason for discipline, not ownership. Re-entry ladders sit at normalized fair value: SK Hynix ₩1.06M/₩740K, Samsung ₩150K/₩130K, Hanmi ₩61K only.

• Prob-weighted 12M returns: SK Hynix −38% (Bull 25% × +8.4% + Base 45% × −42.5% + Bear 30% × −70.2%) · Samsung −39% · Hanmi −58% | Confidence: Medium-High on the normalization framework, Medium on cycle timing — LTA-backed HBM demand can extend the peak longer than history suggests.

---

What Happened / State of the Sector

Monday's crash (KOSPI −8.95%, circuit breaker, SK Hynix −15.37% worst day ever) was positioning mechanics — see our companion Korea macro report. This report answers the question the crash forces: what are these businesses worth at mid-cycle?

The earnings are historic. SK Hynix (000660.KS) printed Q1'26 revenue ₩52.6T (+198%) at 71.5% OM; Samsung's (005930.KS) 7/7 preliminary showed Q2'26 OP of ₩89.4T (~52% OM, the largest quarterly profit of any private company globally outside Aramco) — yet its shares fell >6% that day because revenue missed. The composition is the tell: the surge is overwhelmingly conventional-DRAM contract pricing (Samsung DS revenue +86% QoQ on essentially flat bits; SK Hynix blended DRAM ASP +mid-60s% QoQ), while HBM — the structural story — has contract prices declining through 2026 and SK Hynix's own ASP growth now lags the market because of its higher HBM mix (Korea Investment & Securities, 7/13; TrendForce, 7/3). The HBM4 ramp "does not appear to have materialized at scale" (KED Global, 7/13); SK Hynix deliberately kept capacity on DDR5 rather than converting. Meanwhile Samsung claims first HBM4 mass production/shipment for NVIDIA Vera Rubin — its HBM share recovered to ~21% while SK Hynix's slid from 69% to 58% in a year (Counterpoint, Q1'26).

Concentration cuts both ways: NVIDIA is 14.8% of SK Hynix revenue and a second unnamed >10% whale appeared (12.4%, press: Microsoft or Google) — two buyers ≈ 27% of revenue. Hanmi (042700.KS) shows what the equipment tail of the cycle looks like: revenue collapsed −65.5% as SK Hynix froze TC-bonder orders between HBM3E fleet completion and HBM4 tool decisions, while Hanwha Semitech undercuts pricing — yet it still trades at 40.6x sales.

What to Watch Now

• SK Hynix Q2'26 print (7/22; some calendars 7/29): consensus OP ~₩63.1T (range ₩62–68T); KIS warns ~8% miss risk on HBM-mix ASP drag. The number matters less than the HBM4 ramp commentary — and a beat into a washed-out tape is squeeze fuel against our rating.
• Samsung full Q2 results (7/30): the DS/HBM4 segment detail quantifies the Samsung-vs-Hynix share shift — the single most thesis-relevant disclosure of the month.
• SKHY ADR–Seoul parity: ~25% premium with one-way conversion (Seoul→ADR only) means Seoul float permanently migrates to Nasdaq; TSMC's steady-state ~16% premium is the convergence anchor — implying either ADR downside or a persistent Seoul discount, both keeping the long-ADR/short-Seoul pressure on.
• Seoul policy response: F4 macro-finance meeting imminent, FSC chair at cabinet 7/14; single-stock leveraged-ETF curbs under discussion. A stabilization package is the tape-turn candidate; none arrived by Monday's close.
• Margin-debt burn-down: ₩38T of 신용융자 with liquidation ratios in the panic regime — capitulation completes when forced selling exhausts (historically 3–7 sessions).

What to Watch Next (Technology)

• HBM4 2027 LTA repricing: executed long-term agreements reset HBM ASPs upward in 2027 — this is what separates a 2027 plateau from a classic crash in our DCFs; evidence arrives via customer qualification news through H2'26.
• Hybrid bonding vs TC bonding (HBM5/6 horizon): the technology transition that decides whether Hanmi's franchise (71% HBM TC-bonder share) has a second act — its in-house hybrid bonder targets HBM6 by end-2027 against Hanwha and BESI/AMAT.
• Samsung foundry option: 2nm GAA yield >60% (from ~20% in H2'25), Tesla AI5 tape-out at Taylor under the $16.5B deal — the only non-memory leg in the complex that could re-rate independently of the cycle.

Fair Value Notes

• SK Hynix (000660.KS) ₩1,060,000 — log-median of DCF ₩740K (WACC 11.6% on KTB 10Y 4.15%; normalized rev ₩220T × OM 33% → FCF ₩45T; explicit FY26–30 path harvests the boom then mean-reverts), 9x normalized EBITDA ₩103T → ₩1.35M, 16x normalized FCF → ₩1.06M. Bull ₩2.0M (HBM as LTA annuity: rev ₩290T, OM 42%), bear ₩550K (deep ₩470K). Supersedes our 6/23 mark (₩2.65M): that mark predated the ADR listing, HBM-price-decline evidence, HBM4 slip, and applied weaker normalization. Reverse-DCF: the crash close still capitalizes ₩110T perpetual FCF = 2.4x normalized.
• Samsung Electronics (005930.KS) ₩150,000 — log-median of DCF ₩130K (normalized blended OM 16.5% segment-up: memory ₩135T rev at 30% through-cycle OM + foundry floor + DX/SDC/Harman; WACC 10.0%), 6.0x normalized EBITDA ₩145T → ₩150K, 18x normalized FCF ₩50T → ₩155K, all on the total share base (common+pref, ₩1,671T cap basis). Bull ₩240K (foundry inflects + memory plateau), bear ₩90K. Note the freshest domestic bear (Kiwoom ₩390K, cut 7/8) is still 2.6x our base — the entire street underwrites peak persistence.
• Hanmi Semiconductor (042700.KS) ₩61,000 — log-median of DCF ₩42K (recovery-path rev ₩1.0T by 2027, normalized OM 35%, WACC 11.5%), 15x normalized EBITDA ₩0.40T → ₩65K, 20x normalized FCF ₩0.28T → ₩61K. Bull ₩95K, bear ₩15K. Current EV = 52x normalized EBITDA and 83x peak-year FY25 OP — a peak multiple on hoped-for peak earnings; bulls' own FY26 forecasts (~₩0.80T rev) require the exact HBM4 order wave the 7/13 ramp caution defers.

---

Catalysts & Risks

---

References

• Korea Investment & Securities — Q2 miss risk, FY26/27 OP cut 9%/11%, target maintained (2026-07-13, via E-Today) · KED Global — HBM4 ramp caution (2026-07-13) · TrendForce — 3Q26 DRAM +13–18% QoQ; HBM contract prices declining (2026-07-03)
• Samsung Newsroom — Q2'26 preliminary: rev ~₩171T / OP ~₩89.4T (2026-07-07) · Counterpoint — Q1'26 HBM share: SKH 58%, Samsung 21%, Micron 21% · SK Hynix Q1'26 분기보고서 (2026-05-15) — NVIDIA 14.8% + second >10% customer 12.4%
• Bloomberg — SKHY ADR conversion likely one-way, Seoul float migration (2026-06-25) · ZDNet Korea — Hanmi HBM4 TC-bonder contracts ₩42.8B/₩44.2B (2026-06-08) · InvestChosun — SK Hynix target dispersion ₩1.85M–4.2M (2026-07-09)