000660
The HBM Leader, After the Q2 Reset

Snapshot: Executive Summary

SK hynix is the world's largest maker of high-bandwidth memory (HBM) and the #2 DRAM maker, headquartered in Icheon, South Korea. HBM is the vertically stacked DRAM bolted beside every AI GPU, and SK hynix is the clear leader: the lead NVIDIA HBM supplier, first to HBM4 mass production, and holder of a majority of NVIDIA's Rubin-generation HBM4 allocation⁴. On the library's "own the toll, not the traffic" frame, HBM leadership is as close to a toll-road as memory gets: NVIDIA's accelerators cannot ship without it, and SK hynix has been the first through NVIDIA validation each generation.

The price of that leadership has just been re-set. On book value, the durable anchor for a deeply cyclical business, SK hynix now trades at ~3.9x current book (book value per share ~₩360,000, lifted by the Q2 gains), down from a ~10x June extreme but still above a ten-year 0.75x-2.9x band³. The spot P/E is distorted both ways (trailing collapses on the record quarter, forward ~7.5x normalizes to ~4.5x on a FY26E operating-EPS anchor of ₩315,736), which is exactly why memory is valued through the cycle, not on a snapshot multiple. And the tape has broken: after a near-vertical run to a ₩2,987,000 all-time high on June 25, the stock has fallen ~52% to ₩1,422,000⁵, including a ~9.6% drop on the July 28 Q2 print itself, after the June 29 unveiling of a nearly $600bn Samsung / SK hynix capacity plan had already revived out-year oversupply concerns⁹. The HBM revenue-share lead is also normalizing, from ~70% a year ago to ~57% at end-2025, as Samsung and Micron qualify at NVIDIA (Counterpoint)⁴.

Q2 2026 (reported July 28) set records: revenue ₩79.32tn (+257% YoY, +51% QoQ) and operating profit ₩60.54tn (a ~76% operating margin), with first-half revenue crossing ₩100tn for the first time². But it missed lofty expectations (~₩84tn revenue / ~₩64tn operating profit) because HBM4 shipments slipped into later periods⁸, and reported net profit of ₩93.92tn was flattered by ~₩63.3tn of one-time gains (chiefly the Kioxia stake sale; core net ~₩45.5tn), on a net-cash balance sheet¹. We rate SK hynix Buy, 12-month PT ₩2,100,000 (bull ₩2,700,000 / bear ₩900,000) via a through-cycle EPS x normalized-multiple model. That ₩2.1M sits about +48% above spot yet still well below the Street's median (₩3,300,000) and mean (₩3,334,250) and its Strong-Buy consensus (34 Buy + 4 Overweight / 2 Hold / 1 Sell)⁶: a deliberately disciplined target. We anchor fair value on the record fundamentals, not the crashed price; the ~52% sell-off overshot on cycle-peak fears, so at ₩1,422,000 this is a buy-the-dip. The one real caveat is cycle-top risk, which the ₩900,000 bear frames as a normalized mid-cycle roll. Note the library's memory peer Micron (MU) is also rated Buy at similar forward multiples.

Tactical: SK hynix is trading at ₩1,422,000, ~48% below our ₩2,100,000 12-month target (a disciplined figure that still sits well below the ~₩3.2M Strong-Buy Street). We rate the HBM leader a Buy: the clear #1 in the memory that matters, with net cash and a record Q2 2026, sold off ~52% from its ₩2,987,000 ATH on cycle-peak fears. We anchor fair value on the record fundamentals, not the crashed price: the sell-off overshot, a buy-the-dip. The key caveat is cycle-top risk (memory is cyclical, and out-year oversupply risk builds as Korea funds a nearly $600bn capacity expansion), which the ₩900,000 bear frames as a normalized mid-cycle roll. Rating Buy.

Investment Thesis

Rating: Buy. The probability-weighted scenarios (50% base / 25% bull / 25% bear) blend to ~₩1.95M, about +37% above the ₩1,422,000 spot and a touch (~7%) below the ₩2.1M base as residual cycle-top risk tempers the weighting. We set the headline at the ₩2.1M base, which still sits well below the Street's median target (₩3,300,000). We anchor fair value on the record fundamentals, not the crashed price: the ~52% sell-off overshot on cycle-peak fears, so at ₩1,422,000 this is a buy-the-dip. Our ₩2.1M is deliberately disciplined, far below the Street's Strong-Buy consensus (34 Buy + 4 Overweight / 2 Hold / 1 Sell, mean ₩3,334,250) that has been slow to mark targets down (Micron, at similar multiples, is also a Buy in this library). The one real caveat is cycle-top risk, which the ₩900,000 bear frames as a normalized mid-cycle roll. What would pull us off Buy: signs of an HBM price rollover or an NVIDIA Rubin cut. What would push us toward the Street: confirmation the HBM4 ramp is back on schedule, the upcycle extends toward 2028 (as management and some analysts argue), and the ~57% share lead stabilizes.

Business Overview

SK hynix makes memory chips: the volatile, high-volume commodity that stores and feeds data across every computer and phone, and now every AI accelerator. Its core is DRAM, where it is the #2 global maker and, critically, the #1 in HBM, the high-bandwidth stacked DRAM that the AI build-out is pulling forward. A weaker, more commoditized NAND flash leg rounds out the model, supported more by price hikes and supply discipline than by volume. The whole franchise is deeply cyclical: memory prices boom and bust, and SK hynix's earnings swing with them. The franchise's standing has re-rated with the cycle: in mid-2026 SK hynix overtook Samsung to become South Korea's most valuable company, both memory makers topped $1tn in market value, and beyond NVIDIA the company was selected by Microsoft to supply memory for its custom AI chips. It also listed American depositary receipts on Nasdaq in July 2026 (SKHY), a ~₩45.5tn ($29bn) offering over existing shares⁷.

Where SK hynix sits in the memory chain

• DRAM (the profit core): dynamic memory for servers, PCs, and phones. SK hynix is the #2 maker globally, and the operating profit runs almost entirely through DRAM, now increasingly HBM.
• HBM (the growth engine): high-bandwidth memory stacked beside AI GPUs. SK hynix is the clear #1 (~57% revenue share), the lead NVIDIA HBM supplier, and first to HBM4 mass production. HBM has exceeded ~40% of DRAM revenue and is still rising.
• NAND (the weaker leg): flash storage for SSDs and mobile. More commoditized and lower-margin; supported by price hikes and supply cuts rather than volume growth.
• The cycle: memory is a boom/bust capex business. FY2025 and 2026 are a supercycle peak; the durable through-cycle anchors are book value and normalized mid-cycle earnings.

HBM: The AI Memory Moat

The single most important section in this report. AI accelerators are gated not by compute alone but by memory bandwidth: how fast data reaches the GPU. HBM solves this by stacking DRAM dies vertically (16-high on HBM4) with a wide interface, delivering far more bandwidth per watt than standard DRAM. Every NVIDIA, AMD, and custom-silicon accelerator ships with HBM bolted alongside, and SK hynix is the supplier that has been first through NVIDIA's validation each generation. It moved to HBM4 mass production ahead of Samsung and Micron and holds a majority of NVIDIA's Rubin-generation HBM4 allocation⁴.

Why SK hynix, and why the moat is real

HBM is hard: stacking 16 DRAM dies with through-silicon vias, keeping thermals and yields in line, and clearing NVIDIA's exacting validation is a multi-year capability. SK hynix has been the first mover every generation (HBM2E, HBM3, HBM3E, now HBM4), which earns it the majority allocation and the pricing power. That is the "route through, not past" test from the library's synthesis: the AI build cannot avoid HBM, and it cannot avoid SK hynix as the lead supplier today. The record FY2025 margins (48.55% operating) are the moat cashing out. Management frames the current run as a "structural shift" rather than a normal cycle: with customers prioritizing procurement over price and HBM demand exceeding capacity over the next three years, it argues the upcycle can extend into 2027 and possibly 2028⁸. Independent sell-side voices echo the framing: Goldman Sachs' Alvin So says this is "not the typical memory cycle" and sees "a structural change" in the industry¹¹.

The wrinkle: the lead is normalizing, and the cycle can still turn

Two things temper the bull narrative. First, the revenue-share lead is compressing: from ~70% a year ago to ~57% at end-2025, as Samsung and Micron finally qualify HBM3E/HBM4 at NVIDIA (Counterpoint)⁴. A duopoly-plus-one erodes the pricing power a near-monopoly held. Second, and more structural, the memory cycle can turn: Korea's nearly $600bn Samsung / SK hynix capacity plan is set to double national DRAM capacity within five years, and Morningstar cautions that such build-outs "often lead to oversupply in tail years" if AI spending slows⁹. The franchise quality is not in question; the durability of the peak is, and even at ~4x book (still above the historical band) after the reset, that durability is what the multiple must earn.

The earnings backdrop

DRAM, HBM & NAND: The Segments

Outside the HBM headline, SK hynix runs a three-product memory business whose economics diverge sharply. DRAM (now HBM-led) carries almost all the profit; NAND is the weaker, more commoditized leg. The bear case is less about the HBM moat than about the cycle riding this base: memory prices boom and bust, and SK hynix's record margins are a peak, not a plateau.

DRAM and HBM: the profit core

• DRAM (HBM-led): the profit engine. HBM has exceeded ~40% of DRAM revenue and keeps rising, dragging blended DRAM margins to record levels as AI demand outruns supply. This is where the FY2025 48.55% operating margin comes from.
• This core is also the source of the peak-cycle risk: any HBM or DRAM price rollover reverses record margins fast, because operating leverage cuts both ways.

NAND: the weaker leg

• NAND flash: storage for SSDs and mobile, a more commoditized and lower-margin business. Supported through 2026 by price hikes and supply cuts rather than volume growth.
• NAND is the drag on the blend: a weak leg that dilutes the DRAM/HBM strength and offers less through-cycle protection.

The risk lens

The base case is that 2026 capacity (sold out across HBM, DRAM, and NAND) fills at strong prices while HBM4 ramps; new wafer capacity only adds volume in 2027. The downside is specific: an HBM or DRAM price rollover as capacity expands, arriving while the multiple still pays ~4x book (above the historical band). That single dynamic (peak margins reversing after a broken parabola) is the difference between the base and bear cases.

Financial Health & Trends

The trough-to-record arc

SK hynix's revenue collapsed in the 2023 memory downturn (a full-year operating loss), then re-accelerated violently on the AI/HBM cycle. FY2025 closed at record levels: revenue ₩97.15tn (+47% YoY), operating margin 48.55%, net margin 44.18% (net ~₩42.9tn), EPS ₩60,378, ROE 44.15%². Q2 2026 (reported July 28) extended the run to fresh records: revenue ₩79.32tn (+257% YoY, +51% QoQ) with operating profit of ₩60.54tn (a ~76% operating margin), and first-half revenue crossed ₩100tn for the first time⁸. Reported net profit was ₩93.92tn, but ~₩63.3tn of that was one-time investment gains (chiefly the Kioxia stake sale); core net was ~₩45.5tn. That gain lifted book value per share to ~₩360,000, and the balance sheet remains net cash. Consensus models FY26E operating EPS at ₩315,736; even so, the print missed the ~₩84tn revenue / ~₩64tn operating-profit bar on HBM4-shipment timing.

Q2 2026 print highlights

Note: Q2 2026 figures are from SK hynix's July 28, 2026 newsroom release; reported net profit includes ~₩63.3tn of one-time gains (chiefly the Kioxia stake sale), so NM marks the YoY comparison as not-meaningful. FY26E operating EPS is consensus. Book value per share ~₩360,000 (est.) reflects the Q2 equity gain; net cash was ~+₩32.5tn at Q1 2026 and was further boosted by the Kioxia sale proceeds. Values are ₩ (won).

Capital Allocation & Returns

SK hynix has inflected to a net-cash balance sheet: cash of ₩54.36tn against debt of ₩21.83tn, ~+₩32.5tn net, after years of cyclical net debt¹. Roughly 50% of accumulated free cash flow is pledged to shareholder returns for 2025-2027, and the dividend was raised to ₩3,000/share. But the equity story here is not income: the capital return is real, yet you own SK hynix for the HBM leadership and the through-cycle book value, not the ~0.2% dividend yield at this price.

The reinvestment question

SK hynix's best investment is its own roadmap: HBM4, 16-Hi stacking, and new fabs, funded from record operating cash. Capex was ~₩30.2tn in FY2025 and is guided to the high-₩40tn range for 2026 to fund HBM4, the M15X fab (pulled forward) and Yongin Fab1, including an ~$8bn order for ASML EUV lithography systems. At ~4x current book, the value question is not capital allocation (the balance sheet is now a fortress); it is whether the peak-cycle earnings that fund it are durable, and whether the HBM4 ramp that slipped this quarter lands on the revised schedule. That is the work of the next two sections.

Valuation Overview

Memory cannot be valued on a spot P/E, because a deep cycle distorts it both ways: SK hynix's trailing P/E has collapsed on the record trailing-twelve-month earnings (and is further distorted by a one-time Kioxia gain), while its forward P/E is ~4.5x on a FY26E operating-EPS anchor of ₩315,736. Neither is the anchor. Through the cycle, the durable anchors are price-to-book and normalized mid-cycle earnings. On book, after the ~52% derating and a Kioxia-boosted equity base, SK hynix trades at ~3.9x current book (BVPS ~₩360,000), off the ~10x June extreme but still above a ten-year 0.75x-2.9x band. The valuation question is a durability question: does the supercycle hold and the HBM4 ramp re-rate the multiple back up, or does the cycle turn first?

The bull's strongest valuation counter is that this rally was earnings-driven, not multiple-driven: the Kospi trades at roughly 8.6x forward earnings (versus ~21.2x for the S&P 500), SK hynix lower still at ~4.5x forward, and some managers (Robeco) call Samsung and SK hynix "quite cheap" on an earnings basis. We take the point, but anchor on book: forward EPS is a peak-cycle number (and flattered by the Kioxia one-off in reported terms), and it is the price-to-book, now ~4x after the reset but still above its historical band, that we watch¹¹.

Why the scenario model

A single multiple obscures the two things actually in play: what a normalized mid-cycle EPS is worth, and what multiple the market pays on it. We value FY26E / normalized EPS × a through-cycle P/E → implied PT, cross-check against price-to-book, run three scenarios, and weight them. The base (FY26E operating EPS ₩315,736 at ~6.65x) reconciles to ₩2,100,000, cross-checked at ~5.8x current book; the bear uses a normalized mid-cycle EPS (~₩90,000). Note the library's memory peer Micron (MU) is also a Buy at similar multiples. That work (with an interactive tool) is the next section.

Scenario Model: EPS × Normalized Multiple

The model is the heart of the report. Each scenario values memory through the cycle: EPS × a normalized P/E = implied PT, cross-checked against price-to-book. Toggle Base / Bull / Bear to load the anchor inputs; drag the EPS and P/E sliders to see the implied PT update. The base (FY26E operating EPS ₩315,736 at ~6.65x) reconciles to ₩2,100,000: still well below the Street median (₩3.30M). The bear deliberately uses a lower normalized mid-cycle EPS (~₩90,000), not the peak FY26E figure, because a bear case is a turned cycle.

The two live levers are FY26E/normalized EPS (₩) and the through-cycle P/E multiple. The P/B cross-check reads current book value per share ~₩360,000 (so ₩2.10M ≈ ~5.8x current book). Memory is a book-value and mid-cycle-earnings story, so both the EPS and the multiple are the swing variables.

Sensitivity grid: EPS × P/E (PT in ₩)

PT calculator (EPS × multiple)

Risk / Reward calculator

Note: The assistant reasons from the dashboard's data snapshot and thesis sections: it does not browse the web or access real-time fundamentals beyond what's in data.js. Treat its responses as scenario-modeling support, not primary research. Author judgments on rating, PT, and probabilities remain with the analyst.

Upcoming Catalysts

Risk Factors

• Peak-cycle operating leverage. The pre-eminent risk. Record margins (~76% operating in Q2 2026) are a cyclical peak; any HBM or DRAM price rollover reverses them fast, because leverage cuts both ways. Memory has never sustained margins like these, and the whole bear case is this peak turning.
• NVIDIA concentration. HBM revenue leans heavily on NVIDIA, the primary customer for SK hynix's Rubin-generation HBM4. A cut to NVIDIA's build plans, or a shift of allocation to Samsung and Micron as they qualify, would hit the demand the whole thesis depends on.
• The HBM share lead is normalizing. From ~70% a year ago to ~57% at end-2025 (Counterpoint), as Samsung and Micron qualify HBM3E/HBM4 at NVIDIA. A duopoly-plus-one erodes the pricing power a near-monopoly held.
• Capacity-driven oversupply. Korea's nearly $600bn Samsung / SK hynix plan is set to double national DRAM capacity within five years; Morningstar warns such build-outs "often lead to oversupply in tail years" if AI spending slows, since fresh fabs take two to three years to come online and can arrive as demand tapers⁹.
• Memory-efficiency software. Techniques such as Google's TurboQuant that reduce the memory an AI model needs could, at the margin, soften demand growth; SK hynix argues they instead expand the market by improving AI economics.
• HBM4 ramp timing and valuation. The Q2 revenue miss came from HBM4 shipments slipping into later periods; a further delay would push revenue and the ramp out again. At ~4x current book (still above a 0.75x-2.9x band), the multiple has already compressed off its June extreme but is not yet cheap: any further pricing, demand, or timing disappointment compresses it toward the historical band, which is most of the bear-case move.
• Market frothiness and retail leverage. SK hynix and Samsung together now make up more than half of the Kospi, and the Korea Exchange delayed the June 2026 launch of single-stock weekly options (SK hynix among them) after record volatility, including an ~8% single-day drop followed by an ~8% rise the next day and a market volatility gauge at an all-time high. Regulators voiced "regret" over leveraged single-stock ETFs tied to the two chipmakers, and national margin debt sits at a record ₩37tn ($24bn) with about half of adults now holding brokerage accounts. A momentum- and leverage-heavy, retail-driven tape can amplify a drawdown if the AI narrative cools, independent of the fundamentals¹⁰.
• China / geopolitical exposure. SK hynix runs fabs at Wuxi and Dalian in China on annual US export licenses; a licensing or policy change could disrupt a meaningful part of the footprint.
• NAND weakness. The NAND leg is commoditized and lower-margin, supported by price hikes and supply cuts rather than volume; it offers less through-cycle protection than the DRAM/HBM core.
• Cyclicality of the memory base. Memory is a boom/bust capex business; the 2023 downturn (a full-year operating loss) is a recent reminder that the trough can be severe even with the HBM moat intact.

Scenario Stress Tests

All stress-test PTs are derived from the same EPS × normalized-multiple framework used in the interactive tool: each shock changes only the inputs noted. Deltas are vs the ₩2,100,000 base PT. The bear cases deliberately use a normalized mid-cycle EPS (not the peak FY26E figure), because a bear case is a turned cycle. After the ~52% reset the spot (₩1,422,000) sits ~48% below the base, which is why the probability-weighted blend (~₩1.95M) lands a touch below the base and about +37% above spot.

Bull vs Bear Debate

Technical Analysis

Trader's view

• Price < 50-DMA but > 200-DMA: the parabola broke, so the short-term trend has rolled over while the longer-term uptrend (the 200-DMA) still holds beneath spot.
• Key support: the ₩1,300,000 area (the post-drop shelf below spot). A close below would signal the reset has further to run.
• Key resistance: the broken ₩2,000,000 shelf, then the ₩2,987,000 all-time high far above. A decisive re-rating needs a fresh catalyst (HBM4 shipments back on schedule, or confirmation the upcycle extends into 2028).
• Momentum: washed out after the post-Q2 collapse, consistent with a Buy whose ₩2.1M target sits ~48% above a spot left oversold by a ~52% drawdown from the ATH. The drop also tracks a market-wide Kospi volatility spike that tripped an unprecedented number of trading curbs and pushed the index's volatility gauge to a record¹⁰.

Glossary & Methodology Notes

HBM (High-Bandwidth Memory)

Vertically stacked DRAM (16-high on HBM4) with a wide interface, delivering far more bandwidth per watt than standard DRAM. Bolted beside every AI GPU. SK hynix is the clear #1 (~57% revenue share) and the lead NVIDIA HBM supplier.

DRAM vs NAND

DRAM is fast, volatile working memory (the profit core, now HBM-led); NAND is slower, non-volatile flash storage (the weaker, more commoditized leg). SK hynix is #2 in DRAM globally and #1 in HBM.

The memory cycle

Memory is a boom/bust capex business: prices and margins swing violently with supply and demand. FY2025-26 is a supercycle peak; the 2023 downturn was a full-year operating loss. This is why spot multiples mislead.

Price-to-book vs P/E for cyclicals

For a deep cyclical, a spot P/E is distorted (trailing looks expensive as earnings lag; forward looks cheap as the market prices the jump). Price-to-book and normalized mid-cycle earnings are the durable anchors. SK hynix is at ~4x current book vs a 0.75x-2.9x band, down from a ~10x June extreme after the ~52% derating and a Kioxia-boosted book.

HBM4 / 16-Hi stacking

The next HBM generation, stacking 16 DRAM dies. SK hynix began HBM4 mass production in Q2 2026 ahead of Samsung and Micron; the volume ramp into NVIDIA Rubin runs through H2 2026 after Q2 shipments slipped a quarter (the cause of the Q2 revenue miss).

NVIDIA Rubin

NVIDIA's next-generation accelerator platform (after Blackwell), which uses HBM4. SK hynix holds a majority of the Rubin HBM4 allocation and is NVIDIA's primary HBM supplier; a hypothetical Rubin build cut is the central demand risk in the bear case.

"Memflation"

Shorthand for the sharp memory-price inflation of the AI cycle: DRAM and HBM contract prices rising strongly QoQ as demand outruns supply. It drives the record margins and, if it rolls over, the bear case.

Net cash

Cash and equivalents minus total debt. SK hynix inflected to ~+₩32.5tn net cash (cash ₩54.36tn vs debt ₩21.83tn) in Q1 2026 after years of cyclical net debt: a genuine balance-sheet de-risking.

Methodology

• Snapshot anchor: August 7, 2026. Live price patches via the Cloudflare-Worker quote proxy (Yahoo 000660.KS, in KRW) on page load.
• KRW-native: all market data and operating financials are in Korean won (the KRX listing); no FX conversion is applied.
• FY26E figures are consensus; reported actuals are from SK hynix's Q2 2026 (Jul 28, 2026) and FY2025 results (SK hynix newsroom / stockanalysis.com). Reported Q2 net profit includes ~₩63.3tn of one-time gains (chiefly the Kioxia stake sale); the model uses an operating EPS anchor. Scenario inputs and segment estimates are the author's view, flagged where used.
• The scenario model values FY26E / normalized EPS × a through-cycle P/E, cross-checked against price-to-book. The base reconciles to ₩2,100,000; the bear uses a normalized mid-cycle EPS.
• Conclusions are the author's view. Illustrative, not investment advice.