SK hynix — Research Report (2026-07-17)

Scope: 000660 KS. This is personal research, not investment advice. Reported results are separated from broker estimates; price references carry their source date because this report does not use a live quote.

TL;DR

• What it is: SK hynix is a DRAM and NAND maker whose HBM stacks are a critical input to AI accelerators.
• Take: HBM leadership has turned a historically cyclical memory producer into the scarce supplier in the AI stack — but the current earnings power is also unusually cycle-sensitive.
• ₩97.1tn FY2025 revenue, +47% YoY (record year; profit grew faster than sales).
• ₩47.2tn FY2025 operating profit, 49% margin (mix and pricing, not merely volume).
• ₩52.6tn Q1 2026 revenue; 72% operating margin (a new quarterly record, but also an extreme base to defend).
• ₩35tn net cash at Q1 end (funds capacity expansion without balance-sheet stress).
• HBM4 prepared for mass production first (qualification, yield, and packaging execution are the moat).
• Strongest bull: KB Securities’ ₩4.5m scenario assumes higher memory prices and ~71% ROE.
• Strongest bear: KB’s ₩2.0m scenario assumes HBM4 share loss at major customers and slower DRAM/NAND pricing.
• Watch next: Q2 results, HBM4 customer ramp, and whether operating margin stays above 70% while capex rises.

1. What SK hynix sells — and why it matters

So what: SK hynix is not simply “a memory company” this cycle; it sells the bandwidth and capacity that stop expensive AI processors from waiting for data.

SK hynix (Korea Exchange: 000660) makes DRAM — working memory — and NAND flash, used for storage. Its important product is HBM (high-bandwidth memory): several DRAM dies stacked vertically beside an AI processor. That physical integration delivers far more bandwidth than ordinary DRAM, so an AI accelerator can be fed rather than idle.

The company’s 1Q26 revenue mix was 78% DRAM, 21% NAND and 1% other, according to KB Securities’ company-data compilation. Customer-by-customer revenue is not disclosed in the sources reviewed, so no customer concentration figure is inferred here. The strategic customer set is nevertheless clear: AI accelerators, server vendors, and hyperscale data centers pull HBM, server DRAM, eSSD and newer AI-server modules together.

The immediate product bridge is HBM4. SK hynix announced that it had completed development and prepared the first mass-production system in September 2025; it says the generation doubles I/O terminals to 2,048, doubles bandwidth versus the prior generation and improves power efficiency by more than 40%. Those are product claims, not an independent benchmark, but they explain why timely qualification matters more than merely adding wafer capacity. Company HBM4 announcement

2. Financials — the memory cycle, then the AI step-change

So what: the 2023 loss shows this is still a memory-cycle business; the 2025–26 margin step shows that HBM and supply discipline can radically change its economics.

Sources: FY2022 company release, FY2023, FY2024, and FY2025.

FY2025 revenue rose 47% and operating profit doubled from FY2024. The company says HBM revenue more than doubled, high-performance DRAM expanded, and NAND shifted toward eSSD and other higher-value products. This is the important distinction: revenue grew, but the operating margin climbed from 35% to 49%, showing that each won of sales became more valuable. FY2025 results

Cash confirms, rather than contradicts, the profit. A Mirae Asset Securities table built from company data reports ₩53.373tn operating cash flow, ₩27.374tn PP&E investment, and ₩25.854tn free cash flow for 2025. Treat the FCF definition as the analyst’s presentation convention, not a company-issued non-GAAP KPI. Mirae Asset, Apr. 1, 2026, p. 6

The latest reported quarter went further. In 1Q26, revenue was ₩52.576tn, operating profit ₩37.610tn (72% margin), and net income ₩40.346tn. Cash and cash equivalents reached ₩54.3tn, interest-bearing debt fell to ₩19.3tn, and the company reported ₩35tn net cash. These quarterly figures are preliminary and stated under K-IFRS. 1Q26 company results

3. The operating metric that matters: qualified HBM output

So what: the economic engine is not “bits shipped”; it is the number of HBM stacks that pass customer qualification at high yield, multiplied by their price premium.

HBM consumes advanced DRAM, stacking, thermal management and packaging capacity. A stack is only valuable if it meets a customer’s performance, reliability and volume requirements. That makes yield + qualification + dependable supply the practical unit economics, even though SK hynix does not publish a per-stack gross profit.

The visible proxy is operating margin: 49% in FY25 and 72% in 1Q26. A rising margin in a capital-intensive manufacturer is evidence of price/mix power, but it is not proof that such margins are permanent. The company is now extending the same AI-memory basket beyond HBM: 1c-process DRAM, 192GB SOCAMM2 AI-server modules, and 321-layer QLC enterprise SSDs. 1Q26 results

4. Value chain — a bottleneck, but not a monopoly

So what: SK hynix occupies a bottleneck layer between AI-chip designers and the equipment/packaging ecosystem, yet Samsung and Micron remain real alternatives.

ASML / materials / wafer tools → SK hynix DRAM fabs → HBM stacking & test → AI accelerator / server → cloud or enterprise inference
                                      ↑
                              Samsung and Micron compete here

HBM benefits from a high barrier: a customer cannot instantly switch a qualified, yield-proven memory stack into a demanding accelerator design. SK hynix strengthened that position by preparing HBM4 mass production ahead of the transition and by collaborating with TSMC on the HBM4 base-die logic process. SK hynix–TSMC announcement

But this is not ASML-style monopoly economics. Samsung and Micron can gain qualifications, and ordinary DRAM/NAND remain cyclical commodities. The company itself plans substantially higher 2026 investment around M15X ramp-up, Yongin infrastructure and EUV tools. That protects future supply but eventually risks converting scarcity into oversupply. 1Q26 results

5. Balance sheet, management, and capital allocation

So what: the balance sheet is unusually strong for a memory maker, which lets management fund capacity and still return cash — the question is whether it spends peak-cycle cash at the right point of the cycle.

At its March 2026 AGM, SK hynix said it had delivered ₩14.3tn of shareholder returns from 2025 results through dividends and treasury-share cancellation, while targeting ₩100tn of net cash over the long term. The same AGM framed 2026 around AI-memory breadth, production infrastructure, and expansion across Yongin, Cheongju P&T7 and U.S. advanced packaging. AGM release

KB Securities lists SK Square and affiliates at 20.5% and Korea’s National Pension Service at 7.5% as major shareholders as of its May 2026 report. The sources reviewed do not provide a clean, current Korean equivalent of U.S. Form 4 data, so this report does not make an insider-buy/sell claim. KB Securities, May 29, 2026, p. 1

The relevant M&A legacy is Solidigm, which gives the group a differentiated enterprise-SSD / QLC position rather than leaving it dependent on commodity NAND alone. The execution question is whether that NAND portfolio earns durable returns, not whether it makes the company an AI-memory pure play.

6. Valuation — cheap on estimates, expensive on certainty

So what: the shares look optically inexpensive on forward earnings because forecasts assume extraordinary margins; the real valuation risk is that those estimates are too high, not that the P/E formula is wrong.

The most recent accessible broker price reference is ₩2.58m on June 24, 2026. Mirae Asset then estimated 2026 P/E at 7.8×, 2026 P/B at 5.1×, and a ₩4.2m target price, using a 6.7× P/B multiple — a 10% discount to the average of Micron and Kioxia. This is a dated broker estimate, not a live quote or a target endorsed by this report. Mirae Asset, June 25, 2026

There is no honest single “normal semiconductor multiple”: memory earnings are cyclical, and current forecasts are unusually high. A more useful cross-check is KB’s May peer table, which put SK hynix at 6.6× 2026E P/E versus Micron at 10.8×, but also projected SK hynix’s 2026E ROE at 96.3% — a warning that the denominator contains peak-cycle economics. KB Securities, May 29, 2026, p. 2

7. Bull / base / bear — named, sourced views

So what: the published disagreement is mainly about duration of memory pricing and HBM4 share, not whether AI memory exists.

• Bull — KB Securities: ₩4.5m. Its bull case assumes DRAM ASP rises more than 194% and NAND ASP more than 244% in 2026, yielding 70.7% ROE and 5.9× implied P/B. KB, pp. 1–2
• Base — KB Securities: ₩3.8m target. It assumes high-performance DRAM demand grows with AI and NAND profit normalizes; its P/B–ROE model uses 60% average forward ROE and 5.0× target P/B. KB, pp. 1–2
• Bull, independent cross-check — Mirae Asset: ₩4.2m target. It raised the target on an expectation that more than half of supply volume will sit under long-term agreements and on higher 2027 HBM pricing. This is an estimate, not contracted revenue disclosed by SK hynix. Mirae, p. 1
• Bear — KB Securities: ₩2.0m. Its explicit downside case is higher HBM4 vendor share for competitors at major customers and a slowdown in DRAM/NAND price gains. That is the correct bear case to watch: qualification/share loss plus a normal memory-cycle reversal. KB, p. 2

8. Catalysts and what would falsify the thesis

So what: do not watch an AI headline; watch the three measurable links between the headline and cash.

Catalysts: 2Q26 margin and net-cash update; evidence of HBM4 volume qualification; DRAM/NAND contract-price direction; capex detail for M15X/Yongin; and any disclosed change in long-term customer commitments.

The thesis is wrong if:

1. HBM4 qualification or allocation materially shifts to Samsung/Micron;
2. DRAM/NAND pricing stops rising while operating margin falls for two quarters;
3. capex rises faster than operating cash flow and reverses the net-cash position; or
4. AI customers cut accelerator/server orders rather than merely changing chip vendor.

📚 What this company teaches

1. A commodity can become a bottleneck. DRAM is cyclical, but qualified HBM combines memory, packaging, yield and customer approval; that temporarily changes who has pricing power.
2. Margins tell you whether the story is real. Revenue can grow from volume. A move from 35% to 49% annual operating margin, then 72% in one quarter, says price/mix changed — and sets a very high bar for the next result.
3. Net cash buys options, not immunity. It lets SK hynix fund M15X, Yongin and packaging without emergency financing, but it cannot protect a memory maker from building too much capacity at the peak.

Self-test: If SK hynix reports another record quarter but operating margin falls from 72% to 60%, what are the three possible explanations? Which one is good (mix/capex timing), which is neutral (customer mix), and which would damage the thesis (pricing/qualification loss)?

📖 Glossary

• HBM (high-bandwidth memory) — stacked DRAM that transfers data far faster than ordinary server memory and feeds AI processors.
• DRAM — working memory that holds data while a processor uses it.
• NAND / eSSD — non-volatile flash storage / an enterprise-grade solid-state drive built with that storage.
• HBM qualification — the customer’s technical and reliability approval allowing a memory stack into a specific accelerator or server platform.
• ASP — average selling price; it separates price growth from unit-volume growth.
• Operating margin — operating profit divided by revenue; a direct view of price/mix power after production and operating costs.
• OCF (operating cash flow) — cash generated from normal operations.
• Capex — cash invested in fabs, packaging lines and equipment.
• FCF (free cash flow) — operating cash flow less capital expenditure; cash remaining after maintaining and expanding the business.
• Net cash — cash minus interest-bearing debt; the opposite of net debt.
• P/E — price paid for one won of earnings; it becomes misleading if earnings sit at a cycle peak.
• P/B — price divided by book value; analysts often use it for capital-intensive, cyclical businesses.
• ROE — return on equity; profit generated from shareholders’ capital.

Sources

• SK hynix FY2025 results
• SK hynix 1Q26 results
• SK hynix HBM4 mass-production preparation
• SK hynix 2026 AGM
• Mirae Asset Securities, June 25 2026
• KB Securities, May 29 2026