SK Hynix Investment Thesis
100% concentrated position. Goldman-primary forward model + bottom-up product research + verified primary sources. Memory supercycle / AI infrastructure. Updated June 2026.
Live KRX:000660 (delayed) — SK Hynix in KRW. If the widget is blocked offline, recent level ≈₩2,150,000 (~$1,593) after the Jun 5 pullback (May 29 ATH ₩2,379,000). USD throughout at ~1,350/$.
At a Glance — what matters first
Price (live)
≈₩2.15M
$1,593 · ATH ₩2.38M
PEG / PEGY
~0.07
P/E ÷ ~107% EPS CAGR
Op margin
72%
Q1’26 · record
Revenue Q1’26
₩52.6T
+199% YoY
2026E revenue
₩318T
$235B cons.
Base / Bull Price Targets (% from ≈₩2.15M)
End 2026
Base₩2.7M ($2,000) +26%
Bull₩3.9M ($2,889) +81%
End 2027
Base₩4.0M ($2,963) +86%
Bull₩5.7M ($4,222) +165%
End 2028
Base₩4.9M ($3,630) +128%
Bull₩7.2M ($5,333) +235%
Scenario ladder from the thesis model (Goldman OP × multiples). Goldman-implied 2028 FV ~₩5.1M sits between Base and Bull. Targets deliberately lead the ₩2.08M sell-side consensus — that’s the thesis bet.
2028E revenue mix
HBM — 32% ($143B)
Server DRAM — 20% ($88B)
CPU / SOCAMM — 6% ($26B)
PC DRAM — 15% ($67B)
Mobile/LPDDR — 9% ($40B)
NAND — 19% ($86B)
Analyst forecasts (12-mo PT)
| Goldman Sachs Buy | ₩3.5M +63% |
| Street high KB/Mirae/KIS | ₩3.8M +77% |
| Morningstar FV | ₩4.11M |
| Consensus 12-mo | ₩2.08M ~spot |
Goldman OP path ₩271T → ₩401T → ₩454T (FY26/27/28) · 2026E rev ₩318T · EPS ₩281K
Technical analysis (live)
Latest — Computex 2026, Taipei (Jun 2)
• Chey Tae-won pledged to DOUBLE SK Hynix wafer capacity over 5 years — a reversal from March, when he said new capacity "couldn't be added on demand." Cited memory shortage persisting through 2030. Exec-credibility read: a chairman reversing capacity guidance in 90 days is a high-conviction demand signal, not a hedge.
• Jensen wrote "Please Make More" on an HBM4E wafer and "Love SOCAMM" on a 192GB SOCAMM2 module at the booth — 3rd Chey-Jensen meeting in 7 months. The top buyer publicly asking the top supplier to accelerate.
• Vera Rubin now in full-scale mass production (MSFT+Foxconn bring-up done); SK Hynix holds ~58-70% of Rubin HBM4 orders.
• Flow: Janus Henderson (Richard Clode, Global Technology Leaders Fund) plans to add SK Hynix, citing outsized earnings growth as multi-year supply contracts reprice higher in 2027. Korean media now floating ₩4M targets (~$2,960, +72%).
• Market cap surpassed $1T; YTD +260%+. Competitive watch (SemiAnalysis): Samsung Foundry producing HBM4 base/logic dies on SF4 — qualification is the item to track.
Vera CPU — the x86→Arm shift is a DRAM share-capture win (GTC Taipei, Jun)
Intel −6% / AMD −5% on the Vera unveil — the market priced x86 displacement. The mirror image: the memory house with the SOCAMM2 lock is the structural winner of that same shift.
• Verified content: each Vera CPU carries up to 1.5TB LPDDR5X via SOCAMM; a Vera Rubin NVL72 rack = 36 Vera → ~54TB LPDDR5X + 20.7TB HBM4 on the 72 Rubin GPUs. SK Hynix is lead/exclusive SOCAMM2 supplier.
• Same bytes, higher share, better terms: that CPU DRAM used to be DDR5 RDIMM on x86 hosts — a commodity socket split ~3 ways, repriced quarterly into spot. On Vera it’s SOCAMM2, annually contracted, SK Hynix lead/sole — in the same rack as its HBM4. NVIDIA pitches Vera at 2× bandwidth, 3× per-core vs x86+DDR5.
• De-risks the bear case: the one Server-DRAM worry was “x86 server units fall.” This flips it — x86 decline = DRAM migrating to a socket SK Hynix wins harder. Also ties to Jun 5: the “SOCAMM 55→28TB” scare was about this LPDDR tier; NVIDIA’s live spec confirms 1.5TB/Vera.
Honest weighting: raises conviction on the Server-DRAM + multi-socket leg, not the price targets (SOCAMM was already modelled). Moat-type tailwind, not TAM-type — HBM4 is still the bigger $ driver. Kicker: Vera also runs standalone, so it can eat general Xeon/EPYC sockets, each carrying SOCAMM DRAM — incremental TAM not yet in models.
Room to Run — upside from ≈₩2.15M (post-Jun 5 pullback; live ticker above)
Goldman PT ₩3.5M (~$2,590) = +63%
Street high ₩3.8M (~$2,815) = +77% · Korean media floating ₩4M (~$2,960) = +86% · now ≈₩2.15M (~$1,593)
Goldman-implied 2028 FV ~₩5.1M (+137%) · Bull 2028 ₩7.2M (+235%) · Bull 2030 ₩9.0M (+319%) · consensus ₩2.08M (raised from ₩1.79M) ~at spot
Market Context — June 5 Selloff (KOSPI circuit breaker)
SK Hynix fell ~9% (Samsung -6-7%, KOSPI -6% intraday). Read: crowded-position reset, not a thesis break — but with one genuine fundamental watch-item.
Triggers:
• Broadcom guide: Q3 AI rev $16B vs $17.2B expected, FY $56B vs $57.6B; $100B FY2027 target not raised. Still +200% YoY — an expectations miss, not a demand collapse.
• Rubin memory rumor (the one that matters): reports NVIDIA may cut Rubin NVL72 SOCAMM from 55TB → 28TB/rack — hit memory names directly (Micron -9.5%, SanDisk -11%). SK Hynix is the exclusive 192GB SOCAMM2 supplier, so this is the real watch-item. Caveats: unconfirmed rumor; SOCAMM is not HBM4 (the bigger value driver); 28TB still large; a 20-30% HBM4 shipment shift on Rubin timing was already known.
• Crowded unwind: foreign net selling, leveraged single-stock ETFs -20%, Korean FinMin flagging retail leverage.
What's UNCHANGED (the disconnect): 2026 revenue consensus RAISED ₩266T → ₩318T and EPS ₩218K → ₩281K as the stock fell; MSFT/Meta/Google booked most 2026 HBM capacity; US-listing feedback "tremendously positive." Forward P/E now ~7x (vs ~23x Korean semi avg).
Honest read: fundamentals improved while price dropped — classic crowded-trade reset. But the stock was ~25% above the (raised) ₩2.08M consensus at the peak, and the SOCAMM rumor (if confirmed) is a real per-system content headwind — so some reset was rational, not pure irrationality. Watch: SOCAMM confirmation, foreign-flow reversal, Samsung HBM4 (unveiled Jun 1).
Live Charts & Figures — TradingView
KRX:000660 (SK Hynix) — live/delayed, pulled directly from TradingView. Updates on open.
Figures sourced live from TradingView (KRX:000660). All financials in KRW unless the widget states otherwise. USD comparisons elsewhere on this page use ~1,350/$.
🧠 Governing Dynamic: The AI platform transition generates its own demand. More AI → more tokens → more compute → more memory → more optics. This self-reinforcing cycle is bounded only by physical manufacturing capacity (ASML machines, TSMC packaging, fab space) — which is constrained through 2030.
SI-CTRL analysis: This is not a product cycle. It is a platform transition. Revenue models based on gradual growth curves understate the Rubin step-function.
Research Pages
Live Intelligence Tools (Claude AI)