SK Hynix (000660.KS) Deep Research

Publication date: 2026-06-15 Method: Serenity supply-chain bottleneck research framework Nature: research support only, not a trading instruction

Bottom Line First

SK Hynix is earning from two stacked bottlenecks: (1) the “exclusive time window” in HBM3E 12-Hi, where it passed NVIDIA qualification 9-12 months before Samsung and was almost dominant through the Hopper/Blackwell generation; and (2) the “turnkey supply model” that integrates HBM4 base die with TSMC CoWoS2, turning HBM from a component into a module that can be placed directly on the interposer and bypassing one OSAT step.

But both bottlenecks are being diluted by time in 2026. Samsung already passed NVIDIA HBM3E 12-Hi validation in September 2025, and its HBM4E samples reportedly shipped six months ahead of Hynix. NVIDIA Vera Rubin has officially moved to three suppliers: Hynix at 60-70%, Samsung at 25-30%, and Micron taking the remainder. That means the 72% operating margin in Q1 2026, unprecedented in memory manufacturing history, is very likely a cycle peak. The question is not “can it keep rising?” but “how many quarters remain in the high-profit window?”

More importantly, the memory cycle and the AI bottleneck cycle are running on top of each other in SK Hynix. Once AI HBM price momentum slows, and DDR5/NAND cycle normalization happens at the same time, the drawdown can be larger than in a pure “AI bottleneck” company.

This is a research-priority judgment, not a trading instruction.

1. Market Narrative to System Change

Market narrative: SK Hynix is the “memory crown jewel” of the AI compute era: HBM leader, core NVIDIA supplier, and beneficiary of the HBM3E to HBM4 to HBM4E cycle.

What actually changed in the system:

1. The accelerator bottleneck shifted from computing to feeding. In Hopper/Blackwell/Rubin, GPU FLOPS are rising faster than bandwidth. The new constraint is keeping the GPU fed. Rubin uses 8 stacks of HBM4 per GPU, 288GB capacity, and more than 22TB/s of bandwidth. HBM is not an accessory; it is an organ of the GPU.
2. HBM has moved from DRAM derivative to advanced semiconductor manufacturing. HBM4 base die has shifted from memory makers’ internal processes to TSMC N5 logic process. Hynix intentionally gave up that step in exchange for better integration with NVIDIA’s overall design.
3. HBM unit economics have changed completely. In Q1 2026, Hynix DRAM ASP rose in the mid-60% range QoQ and NAND ASP rose in the mid-70% range. That was a side effect of AI HBM lifting the whole DRAM/NAND price deck, not an ordinary cycle. A 72% operating margin comes from that structure.
4. The customer structure moved from PC/phone/server to NVIDIA plus the three ASIC giants. Google TPU, AWS Trainium, and Broadcom custom ASICs all use HBM3E, and ASIC HBM demand is expected to rise sharply in 2026.

The constraint moved from “compute chips” to “AI memory supply.” The real bottleneck in AI memory supply is not the DRAM die alone, but TSV drilling, 12/16-layer stacking, MR-MUF packaging, and base-die integration.

2. Required Components, Supply-Chain Layers, and Bottlenecks

The real bottleneck ranking:

1. TSMC CoWoS capacity — the hardest bottleneck. In 2026, it is barely enough for NVIDIA alone. All three HBM vendors must queue through this step.
2. HBM3E 12-Hi yield and production ramp — the core of Hynix’s moat; Samsung only recently passed validation, and Micron remains smaller in volume.
3. HBM4 base-die and TSMC integration know-how — Hynix moved first; Samsung is following.
4. HBM backend packaging capacity (TSV + MR-MUF) — Hynix Cheongju P&T7, around $13B, and the U.S. packaging plant, around $3.9B, both target this layer.

Hynix leads in bottlenecks 2 and 4, but it does not monopolize them. In bottleneck 1, TSMC CoWoS, Hynix is constrained rather than the constrainer.

3. Company Evidence, Ranked by Strength

Strong Evidence: Official IR / Earnings / Exchange Disclosures

• Q1 2026 results, record level: revenue of ₩52.6T, up 198% YoY and 60% QoQ; operating profit of ₩37.61T, up 405% YoY and 96% QoQ; 72% operating margin; net income of ₩40.35T.
• Q1 2026 DRAM ASP up mid-60% QoQ; NAND ASP up mid-70% QoQ, with NAND shipments down 10% but pricing more than offsetting the decline.
• HBM sold out for all of 2026, with some customers reportedly extending to three-year visibility.
• NVIDIA Vera Rubin HBM4 three-supplier qualification, announced 2026-06-05 by Jensen Huang: Hynix 60-70%, Samsung 25-30%, and Micron taking the remainder.
• 2025 full-year performance: HBM revenue more than doubled YoY, and HBM3E 12-Hi began scaling in Q4.
• Capital expenditure: Cheongju P&T7 advanced packaging plant around $13B; M15X Yongin fab around ₩19T; U.S. HBM packaging fab around $3.9B.
• HBM share in Q3 2025: Hynix at 57% by revenue and 62% by shipment volume; about 70% in HBM3E specifically.
• NVIDIA represented 27% of H1 2025 revenue, about ₩11T, versus 16% or ₩10.9T in 2024. NVIDIA revenue share rose about 70% in a year.

Medium Evidence: Industry Research and Cross-Checks

• Hynix was the first to complete HBM4 production readiness in September 2025.
• Hynix-TSMC strategic cooperation: HBM4 base die uses TSMC N5 logic process, with HBM and CoWoS2 integration.
• Samsung passed NVIDIA HBM3E 12-Hi validation in September 2025, about 9-12 months after Hynix.
• Samsung HBM4E samples reportedly shipped six months ahead of Hynix in May 2026.
• Samsung supplies more than 60% of Google TPU HBM3E and remains primary supplier in 2026.
• Hynix reportedly won first HBM3E slots for Google v7p/v7e TPU and AWS HBM3E 12-Hi.
• ASIC HBM demand is expected to rise sharply in 2026.
• HBM3E pricing for 2026 reportedly rises around 20%, with Samsung and Hynix negotiating from strength.

Weak Evidence / Items to Check Against DART

• Exact FY2025 full-year line items differ across summaries and should be reconciled against original DART filings.
• U.S. fab location and tax-credit details require checking 20-F/F-1 equivalents and company ESG/IR disclosures.

4. What the Market May Be Missing

1. A 72% operating margin is almost certainly a cycle peak. That level has never appeared in DRAM history; even the 2017-2018 cycle peak did not reach 60%. The question is not how much more margin can rise, but how many quarters the high-profit window can last. I would use 2027 H2 as the base-case timing for visible operating-margin decline.
2. The HBM3E “exclusive window” has closed, and the HBM4 window may be shorter than the market thinks. Samsung passed HBM3E 12-Hi validation in September 2025, and its HBM4E samples were reportedly ahead of Hynix. This means Hynix may not keep 60%+ share in HBM4E during 2027-2028.
3. NVIDIA revenue share rising from 16% to 27% in one year is fragility, not only strength. The market sees core-customer binding, but NVIDIA’s simultaneous qualification of all three HBM suppliers is a dependency-reduction move. When NVIDIA is your largest customer and is actively cultivating your rivals, bargaining power sits with NVIDIA.
4. The real constraint on Hynix is TSMC CoWoS, not Hynix constraining everyone else. No matter how expensive HBM becomes, it still queues through TSMC CoWoS. Hynix’s “turnkey HBM + interposer” ambition is an attempt to reduce that dependence, but success needs 2027 data.
5. DRAM and NAND price increases show AI lifted the whole storage cycle, but the reverse also holds. Once AI HBM pricing slows, the broader DRAM/NAND cycle can normalize at the same time. That is stacked downside risk, not diversification.
6. The capex peak, including Cheongju, Yongin, and the U.S. plant, lands in 2026-2027 and releases depreciation and interest pressure in 2027-2029. That overlaps exactly with the period when HBM4E competition should be most intense, creating a double pressure point for margins.
7. MR-MUF process advantage is being copied. The old “Hynix MR-MUF versus Samsung TC-NCF” story is much less central in HBM4, because both vendors are moving toward multiple process paths.

5. Bear Case / What Would Prove This View Wrong

Ranked by what is most likely to happen first:

1. 2026 H2 HBM ASP rises less than 20%, or ASP flattens early. That would support the cycle-peak view.
2. Samsung’s NVIDIA Rubin HBM4 share rises from 25-30% to above 35%. That would break the assumption of a durable Hynix bottleneck.
3. DDR5 / NAND prices start falling sequentially after 2026 Q3. The larger memory cycle would be reverting.
4. Hynix Q3/Q4 2026 operating margin falls from 72% to below 55%. The market would recognize the cycle high.
5. NVIDIA moves Rubin Ultra / Feynman HBM allocation toward 45/35/20 across the three suppliers. Hynix’s bargaining power would drop.
6. Samsung HBM4E production and yield outcompete Hynix. The structure becomes a three-way HBM market rather than Hynix-led.
7. CoWoS capacity remains larger than HBM supply through 2027. HBM would stop being the true bottleneck, forcing a valuation reset.

On the other side, the bullish view strengthens if Hynix shows clearly superior HBM4 ramp yield versus Samsung, sold-out visibility extends from three years to four years, the U.S. packaging plant starts on time in 2027, TSMC turnkey integration receives first-order confirmation, and ASIC customers diversify further toward Anthropic or xAI custom chips using HBM4.

6. Next Checks, in Priority Order

One-Sentence Summary

SK Hynix is the cleanest and highest-margin bottleneck company in this AI memory cycle, but much of today’s profit comes from the stacking of two temporary forces: the HBM3E exclusive window and peak-cycle ASP. Q1 2026’s 72% operating margin is a peak signal, not a new normal. It belongs at the top of the research list, but the research direction should shift from “how much longer can it earn?” to “which defense line loosens first when the cycle normalizes?” The key variables are HBM3E ASP, Samsung’s HBM4 share, and TSMC CoWoS.

Companion names that should be monitored together are Samsung (005930.KS), Micron (MU), TSMC (TSM), and ASML (ASML). They sit at different bottlenecks in the same chain; studying Hynix without them means seeing only half the system.

This is a research-priority judgment, not a trading instruction. Final decision-making remains with you.