Layer 3 · 000660 — SK Hynix
One-line thesis
SK Hynix is the cleanest HBM monopoly trade in the global equity market — Q1 26 printed revenue +198% YoY and operating margin at 72% with management explicitly guiding "HBM demand for the next 3 years far exceeds capacity," and the NVDA Rubin design win plus the HBM4E samples in H2 26 makes this the highest-conviction Layer-3 memory position in the cycle.
What SK Hynix physically does
SK Hynix is the global leader in High Bandwidth Memory (HBM) — the stacked DRAM that sits adjacent to every AI GPU and is the single binding constraint on every AI training cluster and inference fleet. The Icheon M14 and M16 fabs in South Korea, plus the Cheongju M15 fab, plus the dedicated HBM advanced-packaging lines that surround them, collectively produce roughly 50-55% of the world's HBM volume in 2026 — and the share at the leading HBM3e and HBM4 generations is closer to 60%. SK Hynix is the sole HBM3e supplier to NVIDIA Blackwell and Blackwell Ultra; the company is the lead-customer launch partner on HBM4 for NVIDIA Rubin shipping H2 26 into 2027.
The physical process at the HBM stacking step: 12-Hi or 16-Hi stacks of DRAM dies (each die ~30µm thick after TSV grind) bonded vertically with MR-MUF (mass reflow molded underfill) or TC-NCF (thermal-compression non-conductive film) using Hanmi or ASMPT TC bonders, with thousands of through-silicon vias per die etched by Lam Sense.i Plus tools, inspected by Camtek Eagle / Onto Dragonfly tools, with film metrology done by KLAC, ONTO, and Hitachi tools. The HBM die itself is a 1z or 1a-process DRAM at roughly 10-15nm half-pitch; the next-generation HBM4 base die moves to a 5nm logic process (produced at TSMC for SK Hynix specifically) for additional buffer logic and I/O scaling.
The HBM4 generation marks a structural shift in the supplier-customer relationship. The base die — historically a low-margin DRAM-process buffer die produced internally — moves to a 5nm logic process produced at TSMC, with custom logic IP designed by SK Hynix and integrated I/O for customer-specific interface variants. NVIDIA Rubin gets an SK Hynix-specific HBM4 SKU with optimised base-die logic; AMD and Broadcom get different SKUs. This customisation increases HBM ASPs by roughly 25-35% versus HBM3e and structurally raises the gross margin on the HBM line — a key reason the FY26 operating margin printed at 72% and the company can sustain near that level through the cycle.
Outside HBM, SK Hynix is the second-largest DRAM maker globally (after Samsung) and a meaningful NAND player through the Solidigm subsidiary acquired from Intel. The DRAM standard product line (DDR5 server, LPDDR5x mobile) is the cycle-driven cash flow ballast; HBM is the structural growth and margin driver. The Solidigm acquisition — completed in 2021 for $9B — has gone from a problematic integration through 2022-23 into a contributor as AI-inference NAND demand has materialised in 2025-26. The data-center NAND pricing surge captured by SanDisk Q3 FY26 (+130% QoQ NAND ASP) and Kioxia FY26 (record operating profit) is also flowing through Solidigm at smaller scale.
The financial print
Q1 2026 (reported April 23 2026) printed revenue of KRW 52.6 trillion — +198% YoY, +60% QoQ — operating profit of KRW 37.6 trillion (operating margin 72%, +405% YoY), and net income of KRW 40.3 trillion (net margin 77%). These are the strongest semiconductor margins ever recorded by a memory maker in cycle history — and they are not a one-quarter anomaly. Management guided that HBM demand for the next three years "far exceeds capacity," and that HBM4 mass production is in qualification with HBM4E samples shipping in H2 26 ahead of mass production targeted for 2027.
The Q1 print also confirmed that 2026 HBM, DRAM, and NAND capacity is "essentially sold out" — a statement that has been used by SK Hynix's IR team in every quarterly call going back to Q3 2024 and that the market has now priced into the 12-month forward equity revisions. The stock at KRW 1,820,000 sits roughly 9% below the all-time high of KRW 1,995,000 set May 15. 1m +35%, 3m +77% — the strongest 90-day rally of any name in this Tier A list, and the principal reason the position-sizing discussion below is more nuanced than the conviction call.
Q2 2026 prints July 24. The consensus revenue is approximately KRW 50-53 trillion with operating profit around KRW 35-38 trillion. The setup is constructive but the bar is now extraordinarily high — any sequential softening in HBM3e ASPs or any HBM4 qualification slippage at NVIDIA Rubin can produce a multi-week sell-off even into a beat-and-raise print.
Customer mix
NVIDIA is the single largest end-customer for HBM and represents approximately 60-65% of SK Hynix HBM revenue through 2026 — sole-supplier on HBM3e for Blackwell, lead-customer launch partner on HBM4 for Rubin. AMD is the second-largest at 15-20% (MI355X and the rumoured MI400). Google (TPU v6 / v7) and Broadcom-designed custom silicon collectively account for another 10-15% of HBM through 2026. The customer concentration is real and the principal financial risk — NVIDIA's HBM share-of-wallet is the single largest determinant of SK Hynix's quarterly margin.
The contracting structure has shifted materially in this cycle. Historical DRAM commercial structure was quarterly spot pricing with limited forward visibility. The 2024-25 HBM cycle has moved to long-term-agreement (LTA) pricing — multi-quarter or multi-year fixed-price commitments at the wafer or chip-stack level, principally between SK Hynix and NVIDIA/AMD. SK Hynix has publicly disclosed that 2026 HBM capacity is sold out under LTA commitments and that 2027 HBM capacity is "essentially committed." This is a structural shift away from cyclical pricing and is the principal reason consensus equity models attribute lower cycle-beta to SK Hynix than would be implied by historical memory-cycle analogues.
Outside HBM, the DRAM standard-product customer base is the broader cloud / data-center / handset cohort, while NAND through Solidigm goes to the data-center storage market (a meaningful tailwind given the AI-driven storage demand discussed under WDC / SNDK / Kioxia in Part 2 of this Tier A series). The DRAM standard-product mix is approximately 60-65% server / data-center (DDR5 RDIMM, MRDIMM) and 30-35% mobile (LPDDR5x for premium smartphones), with the balance in client PC and graphics DRAM. The server mix tilt is the principal exposure to the hyperscaler capex cycle — which is the broader AI thesis but adds cyclicality.
Competitive context
Samsung Memory and Micron are the two direct competitors. Samsung lost the HBM3 / HBM3e qualification cycle at NVIDIA in 2024-25 and is now publicly committed to closing the gap on HBM4 — the PyeongTaek HBM4 line is in qualification with NVIDIA targeting H2 26, but the trade press through Q2 26 (Korea Economic Daily, Trendforce, Korea Herald) consistently suggests Samsung is still 1-2 quarters behind SK Hynix on HBM4 yield. Micron has fully qualified HBM3e at NVIDIA Blackwell (a 2025 win) and is ramping the Boise / Idaho IPF1 / IPF2 capacity through 2027 — Micron's share of NVIDIA HBM has expanded from <5% in 2024 to roughly 20-25% in 2026, and is forecast to expand further on the Idaho ramp.
The structural moat for SK Hynix is the multi-decade HBM-specific R&D base (SK Hynix invented commercial HBM in 2013 with AMD), the customer-co-development lock-in at NVIDIA, the advanced-packaging-line capacity that surrounds the M14/M16 fabs, and the cost-position advantage at the Korean operating base. The moat is real but not absolute — Samsung is closing the HBM4 gap, Micron is taking marginal share, and the duopoly+ economic structure is more competitive than the SK-Hynix-monopoly narrative often presented in sell-side material.
Terminal risk
Two terminal risks. First, the cycle: HBM is still a memory product, and memory products cycle. The current cycle has been extraordinary in both peak ASPs and capacity tightness, but every prior memory super-cycle (1995, 1999, 2007, 2017, 2021) ended with a sharp ASP correction and a 12-18 month equity drawdown of 40-60%. The bull case requires that the 2026-27 HBM cycle is structurally different — driven by long-term-agreement (LTA) pricing visibility through 2027 and the binding-constraint nature of HBM supply — but a Samsung HBM4 qualification breakthrough or a Micron capacity ramp surprise can compress HBM ASPs faster than the consensus 2027 EPS model anticipates.
Second, the AI-capex tail: if hyperscaler AI capex moderates in 2027-28 (the bubble warning recycled every six months), HBM unit volumes compress as the marginal NVIDIA Rubin / AMD MI400 build slows. SK Hynix's equity exposure to AI capex is the highest in the memory cohort.
Bull case
The base case is FY26 revenue of KRW 195-210 trillion (versus FY25 KRW ~85T), operating profit at KRW 140-155 trillion, operating margin holding above 70%, and FY26 net income at KRW 130-145 trillion (~$95-105B equivalent). At 13-15x forward earnings — historically generous for a memory maker but appropriate for the HBM-monopoly tilt — that prints the stock at KRW 2,200,000-2,400,000 in 2027, +21-32% from spot.
The upside case is HBM4 qualification at NVIDIA Rubin landing clean H2 26 and Samsung HBM4 qualification slipping into 2027, which preserves SK Hynix's HBM monopoly economics for a full additional generation. Combined with NAND pricing surge through Solidigm (read-through from SanDisk Q3 NAND +130% QoQ, see Part 2) and the HBM4E samples landing on schedule, FY27 net income prints toward KRW 170-180 trillion. At 15x forward, the stock prints KRW 2,800,000-3,000,000 in late 2027, +54-65% from spot.
Gap / bear case
Three concerns. First, the +77% three-month rally has compressed the entry-level R/R — the stock has rerated to peak HBM-monopoly multiple and a Samsung qualification breakthrough produces an immediate 15-25% pullback. Second, the operating margin at 72% is mathematically near-impossible to maintain through a full cycle — even at peak DRAM 2017 the operating margin topped out at ~50%, and the consensus FY27 model assumes margin moderation to 55-60% which is already a meaningful step down. Third, the equity is approaching a market cap of KRW 1,300 trillion (roughly $950B equivalent), which puts SK Hynix in the global top-25 by market cap — at that scale the institutional bid is broader but the marginal flow is harder to sustain.
Optionality
Four options. The NVIDIA Rubin design-win on HBM4 — currently the base case — confirmed publicly during H2 26 removes the principal qualification-risk overhang and triggers the next equity re-rating leg. HBM4E mass production targeting 2027 with a one-generation lead on Samsung extends the monopoly economics for another year. The NAND optionality through Solidigm: if NAND ASPs continue to surge on AI-inference storage demand (the SanDisk / Kioxia / WDC theme), Solidigm contributes incremental margin that is currently modelled as roughly break-even.
The fourth option is hybrid-bonding leadership through the HBM5 / HBM6 transition. SK Hynix has been the most aggressive HBM maker in committing to hybrid-bonding integration (with Hanwha Semitech as the local equipment partner and BESI as the global partner). If SK Hynix delivers production hybrid-bonded HBM5 ahead of Samsung and Micron — which is the consensus path on 2027-28 timelines — the monopoly economics extend through another full generation. This is the long-tail terminal-value preservation that distinguishes SK Hynix from its competitors and that is not fully reflected in the current FY27-28 consensus.
The trade
000660 (SK Hynix) — BUY 9/10. Entry: trim into spot at KRW 1,820,000; full size on pullback to KRW 1,600,000-1,700,000 if the cohort consolidates. The +77% three-month rally has compressed the entry — do not chase here. Position size: 4-5% of NLV — paired with Micron in Part 2 as the HBM bilateral exposure. The Korean equity carries a 10-15% structural geopolitical / governance discount versus a US-listed analogue (this is the reason MU is sized comparably even at smaller fundamental margins). Stop: daily close below KRW 1,350,000 (the early-March base and the 200-day moving average). Catalyst date: July 24 Q2 print; NVIDIA Q1 FY27 print May 20 (HBM commentary read-through); September SEMICON Korea (HBM4 yield disclosure); Samsung Q2 earnings late July (HBM4 qualification status). Trim/exit triggers: operating margin guide below 65%; any verifiable Samsung HBM4 qualification breakthrough at NVIDIA; FY27 HBM ASPs softening on Trendforce monthly data. Conviction: the cleanest HBM monopoly trade in the cycle, with the strongest single-quarter print of any name in this Tier A list. The reason this is 9/10 not 10/10 is the cycle-peak margin question — HBM is the best memory product ever invented, but it is still memory, and memory cycles. Size at 4-5% with full awareness of the +77% three-month context.