SK Hynix
Vertically-integrated IDM: designs + fabricates + sells memory; HBM sold via multi-year allocation contracts to AI accelerator makers, commodity DRAM/NAND priced spot+contract.
The thesis on this name
Memory Super-Cycle
The HBM share leader and Nvidia's lead HBM supplier — the tightest, highest-margin node of the memory super-cycle.
Earnings, margins, COGS & capex
Record-shattering AI memory supercycle. FY25 revenue ₩97.15T / op profit ₩47.21T (49% margin), then 1Q26 broke ₩50T in a single quarter at a 72% op margin — HBM and high-cap server DRAM are sold out for years. Balance sheet de-levered to net cash ₩35T.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~30¢ is cost of goods and ~0¢ operating expense, leaving ~70¢ of operating profit (~77¢ net).
Revenue trend
Margins
sharply rising on HBM mix + memory price spike
rising
rising
elevated, cycle-peak
COGS structure
Dominated by silicon wafers, EUV/litho tool depreciation, and TSMC-fabbed HBM4 logic base dies; HBM4 base-die outsourcing to TSMC adds cost pressure but defends performance lead.
Capex
Rising in 2026 vs 2025 — M15X (Cheongju) HBM fab ramp, Yongin mega-cluster infrastructure, ASML EUV procurement, 1c-nm DRAM + 321-layer QLC NAND. Exact KRW figure not disclosed.
Latest earnings
In line / modest beat — record set; CNBC noted 'in line with estimates as memory prices climb'
FY26 investment 'significantly above' FY25; HBM demand already exceeds planned capacity for next 3 years; HBM4E sampling 2H26, MP 2027. No precise revenue/capex guide given.
- Op margin
- 72% (record)
- Net cash
- ₩35T
- HBM share
- ~62% (Q2 2026)
- Cash & equiv
- ₩54.3T
Growth drivers
- HBM for AI accelerators (Nvidia Rubin/HBM4 ~2/3 allocation)
- High-capacity server DRAM (DDR5, 192GB SOCAMM2, LPDDR6 1c-nm)
- Enterprise SSD / 321-layer QLC NAND (Solidigm)
- Memory price super-cycle (DRAM+NAND ASPs rocketing on shortage)
- HBM4E sampling 2H26, mass production 2027
Bull & bear
The single best-positioned beneficiary of the AI memory supercycle — HBM is structurally short for years, SK Hynix owns ~62% of it with Nvidia's largest HBM4 allocation, earning 72% op margins on a net-cash balance sheet, yet trades at a mid-single-digit forward P/E because the market refuses to believe memory cyclicality is broken.
- HBM demand exceeds planned capacity for the next 3 years — visibility unusual for memory
- ~2/3 of Nvidia Rubin/HBM4 orders locked; Micron absent from some Rubin allocations
- Forward P/E ~5-6x vs ~70% net margin and ~44% ROE — extreme cyclical discount
- Net cash ₩35T funds the M15X/Yongin/EUV capex without dilution risk
- Nasdaq ADR listing is a re-rating catalyst (US index/ETF eligibility, broader buyer base)
This is peak-of-cycle memory: 72% margins and 198% YoY growth are the definition of a top, not a baseline. The ~5x forward P/E isn't 'cheap' — it's the market correctly discounting a coming glut as all three players pour capex into HBM4/HBM4E, with a single end-market (AI training) and a single dominant customer (Nvidia).
- Memory always mean-reverts — the low forward multiple reflects expected ASP collapse, not value
- Samsung + Micron HBM4 ramps add supply into 2027; HBM4E base-die shifts to TSMC for all
- Demand is concentrated in a few hyperscaler AI capex budgets that could pause
- HBM4 logic-die outsourcing to TSMC raises cost and cedes part of the value chain
- Massive 2026 capex commits at the cycle peak — classic memory over-build risk
- Korea/China geopolitics + FX add non-operational tail risk
What it is worth
Comps + reverse-DCF sanity check on a cyclical-peak earnings base
2027 HBM glut as all three ramp
ASPs roll over; op margin reverts toward 25-30%; the 'cheap' multiple was correct → earnings halve and the stock de-rates with them.
Memory normalizes gradually; HBM premium cushions the down-leg; multiple stays depressed (5-8x) as the market keeps discounting cyclicality even as earnings hold → range-bound to modest upside.
HBM scarcity holds 2026-27, margins stay 50%+, ADR re-rating lifts multiple toward Micron's ~10x → meaningful upside on un-derated earnings.
At ~$1.24T cap on ~₩97T FY25 / ~₩200T+ annualized 1Q26 run-rate, the forward P/E is only ~5-6x (vs Micron ~10x, semis median ~37x). The mid-single-digit multiple means the market is implicitly assuming this is peak-cycle: it prices a sharp ASP/margin reversion (op margin from 72% back toward a 25-40% mid-cycle) within ~2 years, i.e. it does NOT capitalize current earnings as a baseline. Reverse-DCF reading: to merely justify today's price on a no-growth basis you only need ~2-3 normalized cycles of mid-teens-to-20% margins — a low bar if HBM scarcity persists, a trap if a glut hits. Analysis only; non-US name, NOT a buy recommendation.
SWOT
Strengths
- #1 in HBM at ~62% share with multi-year sold-out book
- ~2/3 of Nvidia's next-gen HBM4 (Rubin) allocation
- Record 72% op margin + net-cash ₩35T balance sheet
- Process lead: 1c-nm DRAM, 321-layer QLC NAND, TSMC HBM4 base-die partnership
Weaknesses
- Pure memory exposure — historically the most cyclical, lowest-moat semi segment
- HBM4E logic die outsourced to TSMC adds cost + supplier dependence
- Single-customer concentration: Nvidia/AI hyperscalers dominate the HBM book
- Korea-listed; capital-return yield thin (~0.2%) vs US peers
Opportunities
- AI memory supercycle multi-year demand > supply
- Nasdaq ADR ('SKHY', July-2026, ~$29B raise) broadens US investor access + funds capex
- HBM4/HBM4E generational upgrade cycle into 2027
- Custom HBM (logic-integrated base die) deepens lock-in with accelerator designers
Threats
- Memory is cyclical — a supply glut (Samsung/Micron HBM ramp) could collapse pricing
- Samsung racing to close HBM4 gap; Micron has overtaken Samsung on some allocations
- Customer in-housing / HBM commoditization compresses the premium
- Geopolitics: China demand exposure, US/Korea export controls, FX
Moats, dependencies & bottlenecks
Moats
generational; must re-win each node (HBM4→4E) ~62% share, best yields, first to qualify on Nvidia platforms
Multi-year sold-out HBM book; ~2/3 of Nvidia HBM4
Only 3 firms can fund leading-edge memory fabs; net-cash funds capex
Custom logic-integrated HBM deepens lock-in but cedes die to TSMC
Dependencies
Customer concentration Largest HBM buyer; Rubin/HBM4 allocation is the core growth engine
HBM4/4E logic base die 3nm/N12 base die outsourced; cost pressure + single advanced foundry
EUV lithography EUV tools gate 1c-nm DRAM; sole-source, long lead times
HBM demand rides a few hyperscaler training budgets
ASPs are the swing variable; glut would crush margins
Advantages
- Largest HBM share (~62%) + best yields
- Net-cash balance sheet at cycle peak
- Deepest Nvidia HBM4 allocation
- Full DRAM+NAND+HBM portfolio (Solidigm enterprise SSD)
Weaknesses
- Pure-play memory cyclicality
- Customer + end-market concentration (AI/Nvidia)
- Dependence on TSMC for HBM4E logic dies
- Thin shareholder yield vs US large-cap semis
Bottlenecks
- HBM/advanced-DRAM wafer capacity (sold out for years)
- EUV tool availability from ASML
- TSMC advanced-node base-die capacity for HBM4E
- Yongin cluster build-out timeline / power + infrastructure
Top signals & trends
Top signals
Demonstrates pricing power at cycle peak
Bullish (re-rating) / dilutive (17.79M new shares) · Broadens US buyer base + funds capex; some equity dilution
Bearish signal embedded · Market pricing a downcycle, not value
Confirms intensifying 3-way HBM4 race
Trends
Multi-year visibility, rare for memory
SK Hynix samples 4E 2H26, MP 2027; TSMC takes base die
Negative (longer-term) · Glut risk into 2027-28
SK Hynix Nasdaq ADR widens capital access
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
EUV lithography — gates 1c-nm DRAM; sole-source
HBM4/HBM4E logic base-die foundry (3nm/N12) — supplier AND co-designer
Deposition/etch wafer-fab equipment
Etch/deposition tools, key for 3D NAND + advanced DRAM
Process control / inspection metrology
Largest HBM customer; ~2/3 of HBM4 (Rubin) allocation
Instinct AI accelerators use HBM
Custom AI ASICs (hyperscaler) consume HBM
AI data-center demand for server DRAM + HBM-bearing accelerators
#2 memory + racing to close HBM4 gap; also a foundry rival
#3 in HBM but fully sold out 2026; the US-listed pure-play comp; fwd P/E ~10x
Broadest memory portfolio + scale
NAND competition, not HBM