Data as of 2026-10-28; market action reflects intraday Seoul-time readings on October 28 (around 12:40 KST), with official closing values to be registered in a subsequent amendment. This is piece four in the public website series and piece five in the broader research sequence, following our September 16 SK hynix note, "The Strongest Results, the Most Conservative Pricing"; the Micron, SanDisk, and ChangXin notes serve as companion references, and their frameworks are not repeated here. Earnings figures are the company's preliminary results (not yet externally audited).
On July 29, SK hynix delivered what was then the best quarterly report in its history, and the stock closed down 9.61% that day. On the morning of October 28, the company delivered an even better one — and this time, the stock opened higher and kept climbing. Between those two "best-ever earnings reports" sits a three-week, -34% drawdown, and that drawdown is where this piece begins: "expectations are fully priced in" was never a static label — it is a function of the path. First the numbers, then the path.
3Q26 revenue came in at 104.6155 trillion won, up 32% quarter-over-quarter and 327.9% year-over-year — the first quarter in company history above 100 trillion won, with a single quarter's revenue exceeding full-year 2025 (97.1467 trillion). Operating profit was 83.3338 trillion won, for an 80% operating margin, up another 3 percentage points quarter-over-quarter; it was the sixth consecutive quarter of record revenue and record operating profit. Against consensus: revenue beat the LSEG SmartEstimates figure (98.34 trillion) by 6.4%; operating profit beat the LSEG figure (76.72 trillion) by 8.6% and the Yonhap Infomax figure (76.26 trillion) by 9.3%, and also exceeded the most aggressive pre-print sell-side estimate (Citi's 81.3 trillion).
Net profit of 39.7756 trillion won (38% net margin) fell 58% quarter-over-quarter, and that number has to be unpacked: the second quarter's 93.9 trillion net profit carried a high base of one-off investment gains including the Kioxia stake; the third quarter booked a net non-operating loss of 32.6 trillion won — a net foreign-exchange loss of 11.6 trillion from the stronger won, and 22.9 trillion of investment-asset valuation losses — all below the operating line. EBITDA was 87.7 trillion won, an 84% margin. On the balance sheet: cash and equivalents ended the quarter at 114.5 trillion won (up 26.5 trillion quarter-over-quarter), borrowings fell to 15.8 trillion, net cash reached 98.7 trillion, and the debt ratio stood at 5%.
The volume/price split: DRAM bit shipments grew about 10% quarter-over-quarter (on plan) with ASP up in the high-20s percent — conventional DRAM price increases and a rising HBM4 mix both contributing; server DDR5 shipments rose more than 40% quarter-over-quarter; NAND bits declined slightly (deliberate volume discipline, focusing on high-value products) with ASP up in the mid-20s percent (321-layer TLC enterprise SSD scaling). The stronger won was an operating-level headwind, fully overwhelmed by pricing.
Our core judgment: this income statement beat consensus by 9%, but the real state change this quarter is not in the income statement — it is in three structural facts: a contract book that effectively sells out capacity through 2029, roughly 100 trillion won of shareholder returns, and HBM4 yields above plan with the fourth-quarter crossover into the mainstream product. All three point to the same proposition: this memory cycle is being rewritten into an order-book business, while the market is still bidding it as a cyclical peak. The gap between those two grammars is the main storyline for the next twelve months.
Market Action on the Day
Note: this section uses intraday readings as of roughly 12:40 KST on October 28 (Yahoo Finance daily bars); official KRX closing values (15:30 KST) will be registered through a subsequent amendment and do not affect this note's judgment structure.
On earnings day, 000660 opened higher and kept strengthening: as of intraday Seoul time on October 28 it traded at 1,469,000 won, up about 9.5% from the prior close of 1,341,000, with an intraday high of 1,470,000 and volume of about 2.44 million shares (intraday cumulative); the KOSPI traded at about 5,985, up about 6.1%, triggering a buy-side sidecar (the program-buying halt mechanism) for the second consecutive session. Intraday flow readings (morning basis): foreign and institutional investors each net bought about 300 billion won — for foreigners, the first single-day turn to buying after more than a month of net selling since September 18 — while retail investors net sold about 600 billion won. In the prior US session, the Nasdaq ADR (SKHY) closed regular trading up 1.92% at $145.36 and rose another 6.11% after hours to $154.24.
That move has to be read inside October's decline. The stock was still around 1,841,000 won at the start of the month (October 6 close), then fell for three straight weeks: double-digit daily drops on October 13 and 15, a 1,223,000 close on October 20 (the first break below 1.2 million won in six months), and an intraday low of 1,171,000 on October 23 — a 61% drawdown from the June all-time high of 2,987,000. Two days of rebound followed, with the stock closing at 1,341,000 the day before earnings. For that -34% October drawdown, the brokerage community's attribution was remarkably uniform: "an excessive decline with fundamentals undamaged." All four drivers sat outside the fundamentals — the 40 trillion won buyback announced in August finished executing about a month ahead of schedule (around October 21), removing a mechanical daily bid of 400,000-650,000 shares; the Solidigm listing controversy introduced a governance discount (the financial regulator's October 19 remarks); foreign investors had been net sellers since September 18 (5.68 trillion won across October 19-20 alone); and the global AI-hardware deleveraging overlay (the US 10-year yield near 4.9%, position cleanup ahead of the November 3 US midterm elections).
Hence the opening contrast: the best-ever report on July 29 falling 9.6% was the textbook sample of "expectations fully priced in"; an even better report rallying on October 28 is the same label after being reset by a -34% non-fundamental drawdown. The state of expectations is path-dependent, and the label must be read against the price level it hangs on.
Three Sold-Out Years: This Time the Evidence Is Contract Text
The "supercycle versus just another cycle peak" debate received adjudication-grade evidence this quarter — and the evidence takes the form of contracts, not outlooks.
The CEO's words on the call: combining secured long-term agreement volumes with demand currently in concrete negotiation, DRAM and HBM volumes for the three years through 2029 are "effectively all sold out"; the company's stated view is that the supply shortage persists through 2030. The structure behind that sentence came with numbers attached: long-term agreements (LTAs) now span 14 customers (including leading global cloud service providers, versus roughly 10 in the second-quarter framing), covering more than 50% of total 2027-2028 production capacity; the agreements typically run five years, with some customers discussing extensions; pricing reflects market conditions within a band agreed with customers — a floor and a ceiling — supplemented by deposits and similar mechanisms that ensure purchase plans convert into actual orders. We wrote in the September 16 note that SK hynix had the thinnest contract disclosure of the three incumbents — "framework but no numbers." That gap was half-closed this quarter: coverage was disclosed for the first time, and the deposit mechanism was confirmed in the company's own words.
The demand-side framing is just as extreme: customers are currently asking for more than twice what the company can supply. The "spec downshift" circulating in the market (some customers moving from 12-high to 8-high HBM) was addressed head-on during the call: it is shortage behavior — customers maximizing the number of deployable HBM stacks (cubes) out of limited supply — not demand weakness; the supply plan already reflects the increase in base-die and back-end packaging demand that more cubes entail. HBM4E samples were delivered in the second quarter, and custom-HBM projects with major customers are progressing. Achievable 2027 demand growth (supply-constrained basis) is around 25% for DRAM and high-teens for NAND; 2027 revenue and profit are guided to be "significantly higher than 2026" in absolute terms.
Place this in history: the memory industry has never sat near a cycle peak with incumbents holding an order book covering three years, with price floors and deposits attached. The falsification conditions of the cycle change shape accordingly — a quarter-over-quarter downturn in contract prices is no longer sufficient on its own (movement within the band is by design); what matters is whether any agreement customer cuts volumes, defaults, or renegotiates deposit terms. That is the starting point for this note's falsification criteria below.
HBM4: Yields Above Plan, and 2027 Volume and Price Already Signed
HBM4 entered full mass production and began shipping this quarter, with 12-high product delivered on the schedule agreed with key customers; the M15X fab in Cheongju is ramping at speed toward full capacity by year-end; and the company guides that in the fourth quarter HBM4 will surpass HBM3E on both shipment volume and revenue to become the mainstream product. On yields (the HBM marketing head's own words): both front-end and back-end yields are at industry-best levels, with back-end yield measured across the full bump/stack/KGSD process chain; HBM4 has already exceeded planned yield and the target has been raised, while HBM3E has surpassed its originally scheduled mature yield.
More important is 2027: the CFO confirmed that next year's negotiations with major customers are complete, covering HBM volumes and prices — "the negotiations faithfully reflect the enhanced product value of HBM4, and we therefore expect the profitability gap between HBM and conventional DRAM to narrow meaningfully." That sentence deserves two readings, because it implies a counterintuitive fact: within 2026, HBM's booked margin is not higher than conventional DRAM's — legacy contract pricing is competing against spot-driven conventional DRAM price increases, making the "premium product" a drag on booked margins; otherwise there would be no gap to "narrow." Micron's late-September results gave the same directional reading, so two incumbents have now confirmed the same structure in sequence: HBM's repricing repair happens in 2027, and the intuitive "HBM premium" narrative does not match the 2026 books. For estimating 2027 profit elasticity, this framing matters more than any single quarter's numbers — SK hynix has the highest HBM mix of the three incumbents, hence the heaviest legacy-contract drag and the greatest repricing elasticity, and both the volume and the price ends of that elasticity are now signed.
The marginal change in the pricing environment should also be recorded faithfully: TrendForce's late-September guidance for fourth-quarter contract prices was conventional DRAM up 10-15% quarter-over-quarter, NAND up 15-20%, and enterprise SSD up 23-28% (the only accelerating category), with mid-October industry estimates putting overall DRAM at about +20% — still double digits, but the slope has been decelerating from +80% in the second quarter to +40% in the third and +20% in the fourth, and some suppliers' increases have begun lagging the market because of LTA price caps. The band mechanism cuts both ways: it caps in a bull market and floors in a bear market. Where prices actually land within the band is from here on a more important observable than the headline increase itself.
A Hundred-Trillion-Won Vote: Returns, Solidigm, and the Sequencing of Capex
Management's capital allocation is a more honest cycle signal than its rhetoric, and this quarter's signal is historic in scale: approximately 100 trillion won (roughly $75 billion) of total shareholder returns for fiscal 2026 — cash dividends of 30 trillion won or more, and buybacks-with-cancellation of 60 trillion won or more — fulfilling the commitment to return 50% or more of free cash flow. Within that, the 40 trillion won buyback announced in August completed execution last week (24.07 million shares, 3.3% of shares outstanding, about a month ahead of plan) and will be cancelled shortly; this quarter the board newly resolved a 14 trillion won dividend (20,000 won per share = 10,000 won on operating results + 10,000 won special dividend from Kioxia investment gains) plus a new 20 trillion won buyback (10 trillion earmarked for return-purpose cancellation + roughly 10 trillion for employee compensation, the latter not counted in the return total); buybacks will continue every quarter from here, sized by the board against the share price, valuation, and cash flows. The CFO's words: the current share price does not adequately reflect the profits and cash flows the company generates — it is undervalued.
Some dimensioning: 100 trillion won of returns is roughly 10% of the current market capitalization (about 976 trillion won on October 27, derived from roughly 728 million shares, pending verification against disclosed share count), and roughly 2x 2026 capital expenditure (previously guided at the high-40-trillion-won range); the return commitment sits on top of 98.7 trillion won of net cash. Contrast the 2018 cycle peak: incumbents raised returns near the top then too, but there were no LTA price floors, no deposit structures, and the return-to-free-cash-flow ratio was far below this 50%-plus commitment. Putting peak-cycle profits, an order-book structure, and returns-first capital allocation in the same frame is a first in memory-industry history.
The same day carried an easily overlooked move: the CEO opened the call with a dedicated explanation for discontinuing the review of the Solidigm IPO — given financial-market volatility and the pullback in semiconductor shares, an IPO could not fully achieve the intended outcome. This directly answers the governance discount triggered by the regulator's October 19 "duplicate listing" remarks, and it is also a capital-cycle reading: peak-cycle management behavior is to monetize assets near the top, and refusing to sell assets cheaply into a "share-price correction window" is the opposite signal. Most of the Kioxia investment gains are being returned to shareholders through special dividends rather than recycled into capacity.
On capex, 2027 spending will "increase significantly versus 2026" (advanced-node equipment, the Yongin fab 1 ramp, and accelerating infrastructure), with no absolute figure disclosed. But the sequencing of the language is itself the signal: infrastructure first (accelerating building shells: the Yongin cluster plus M17), cleanrooms opened in phases, and the genuinely supply-adding resources — equipment move-in and production ramps — adjusted to demand; and in assessing demand, the company looks not only at customers' capex plans but audits their financial soundness, financing capacity, and supply-chain readiness. In other words, supply discipline has been structurally written into the investment process — cleanroom-scale new supply does not pressure profits until late 2028 into 2029, and the hand on that valve belongs to the incumbents.
Trend and Implied Pricing: Which Path Is the Market Paying For?
Lay out the full price trajectory: the June all-time high of 2,987,000 → July's -46% (including -9.61% on the July 29 best-ever earnings day) → the September 22 rebound high of 1,925,000 (a lower high; that day closed at 1,840,000) → the October 23 intraday low of 1,171,000 (a lower low, -61% from the high) → 1,341,000 at the pre-earnings close → the earnings-day rally. The downtrend sequence (lower highs plus lower lows) was provisionally broken in late October by three consecutive up days plus the earnings-day surge, but no key platform has been reclaimed: 1.6 million (the mid-October breakdown level), 1.84 million (the early-October platform), 1.93 million (the September rebound high). The current position is roughly half off the high and about a quarter above the October low (on intraday October 28 readings) — the switching day between "basing" and "early repair."
Let implied pricing speak through reverse arithmetic. The forward 12-month P/E is 2.7-2.9x (2.89x per fnguide, October 27; 2.7x per Citi, October 20), the forward P/B is 1.4x (Citi, October 20), against a sector P/E of 24.84x. Run one calculation: at a normalized 12x P/E, the earnings-day intraday price of 1,425,000 implies a market-endorsed "sustainable EPS" of about 119,000 won, corresponding to annual net profit of about 86 trillion won — roughly a quarter of the net profit implied by the sell-side 2027 operating-profit consensus (405-435 trillion won, representing 55-61% profit growth in 2027). In other words, the market is pricing the path "this quarter is the peak, and profits fall 70%-plus after 2027"; the sell-side consensus path is "another 55-60% growth in 2027, then a high plateau." Both paths cannot be right.
The sell-side's loss of anchor deserves its own entry: the pre-earnings target-price consensus was roughly 3.249-3.394 million won (range 1.48 million to 4.7 million), implying +142% upside to the October 27 close; through October's decline, Citi, Nomura and others maintained targets the market simply ignored. This is the mirror image of what we saw in the Palantir note, where price ran above the sell-side's models — here price sits at roughly forty percent of the models, and the marginal pricing power lies with macro and flows, not fundamental models. There is also a mechanical anchor: 100 trillion won of returns against a 976 trillion market cap is roughly a 10.2% total return yield (about 3.1% dividends plus 6.1%-plus cancellation), and the cancellation portion is a permanent lift to per-share earnings — if the market keeps refusing to re-rate, the company mechanically raises per-share metrics by retiring roughly 6% of the share base per year. Re-rating does not depend on multiple expansion; the returns themselves are the path.
We disagree with both sides, and we will write it plainly. Against the bull consensus (targets of 3.25 million and up): the sell side extrapolates 2027 profits linearly into targets while underpricing the structural fragility October exposed — the flow vacuum between buyback windows, the governance discount from the listing controversy, the amplification of moves by index-level leveraged money in the Korean market — all of which put the timing of any re-rating outside the earnings calendar's control. Against the bears ("peak profits are unsustainable"): mean reversion is mathematically certain, but a trade has to answer "when, and from what level" — the price floors, deposits, and three sold-out years push the "when" past 2029 and lift the "level" far above historical peaks, while a 2.7x multiple has already prepaid the worst reversion path. The bears' only real remaining weapon is demand-realization risk (debt-financed AI capex, power bottlenecks, rates) — macro-level variables, not signals the memory industry itself can produce.
The chain's time-zone offset is this note's most practical reading for portfolio purposes: the KRX is this memory cycle's leading time zone — it completed the -61% de-rating first and has now started re-rating first; the analogous contract-book re-ratings in US-listed memory (Micron, SanDisk), the equipment chain feeding on capex from both sides, and the spillover of crowding-out effects into niche memory are the same logic's later time zones. The verification criteria in our Micron and SanDisk notes (September 7) run on their original timetables; SanDisk's November 6 report and Micron's late-December guidance are the next two checkpoints.
The Bears' Fair Hearing: Why the Re-Rating Could Be Late, or Absent
Take the other side's arguments seriously; there are five. First, the gravity of an 80% operating margin: normalization requires no extra assumptions — the bears need only wait for any one demand-side item (cloud capex deceleration, power delays, high rates) to land; PCs and handsets are already being crowded out by price increases, and if that crowding-out spreads to enterprise IT budgets, the "AI increment covers everything" framing comes under pressure for the first time. Second, the reflexivity of the buyback floor: 100 trillion won of returns ties the share price to the company's cash flows, and at the next earnings stall, a shrinking buyback would compound fundamental disappointment — October already demonstrated how deep the flow vacuum gets between buyback windows (-34% in three weeks). Third, the band-as-ceiling scenario: TrendForce has already noted some suppliers' price increases lagging the market because of LTA caps — in a bull market the band is a floor, but after prices peak it is also a ceiling; if contract prices keep landing near the lower edge of the band from the first quarter of 2027, "contracted = locked in high" gets rewritten as "contracted = capped upside." Fourth, two competitive fronts: Samsung reports this week, and progress on HBM4 yield recovery and major-customer tie-ups could change the share narrative at any time; Chinese suppliers' qualification progress in both HBM and niche DRAM is the longer-horizon supply variable. Fifth, Korean market structure: sidecars triggering in both directions signals index-level leveraged money in charge, with volatility mechanically amplified, and the macro windows — the US midterms and the Fed meeting — are still open.
The Signal Watchlist: What Observable Data Would Change the Judgment
Our directional conviction keys off the most recent signal — this quarter's report is an upward signal (a comprehensive beat, three sold-out years, hundred-trillion returns), and the direction holds until the next signal flips; it does not decay with the passage of time. Five watch items, with baselines and flip conditions:
| # | Watch item | Current baseline | Flips up | Flips down |
|---|---|---|---|---|
| 1 | HBM4 ramp delivery | Yields above plan, targets raised; Q4 dual-metric crossover vs HBM3E guided | Earlier crossover confirmation / further yield upgrades / customer add-ons | Delivery delays, customer timeline changes, crossover shortfall |
| 2 | 2027 pricing lands | Q4 contract prices still double-digit, but slope decelerating 80→40→20; LTA caps starting to bind | Double digits hold into 1Q27 / in-band landings near the upper edge | Increases shrink to single digits / landings hug the lower edge / spot prices turn down first |
| 3 | Return execution | 14 trillion dividend + 20 trillion buyback resolved this quarter; 40 trillion bought, awaiting cancellation | Faster execution / additional returns / on-schedule cancellation | Smaller buyback resolutions, delayed cancellations (the earliest cash-flow stress signal) |
| 4 | Flow structure | Foreigners net sellers since September 18; single-day turn to buying on earnings day (not yet a trend) | Weekly-scale foreign buying + price self-support between buyback windows | Continued foreign selling the buyback cannot offset (October pattern repeats) |
| 5 | Competitive landscape | Company claims first-supplier status on HBM4, industry-best yields; Samsung reports this week | Rival ramps fall short / custom-HBM positioning widens | Rival volume beats accompanied by price-cutting / Chinese supplier qualification breakthrough |
Items 2 and 3 are the primary tests — the two legs of a contract-book re-rating are prices landing and cash arriving; item 4 is the daily sentinel for whether October's pattern (flow-led decline) repeats; on item 5, distinguish "share redistribution" (no judgment change) from "share war with price cuts" (judgment change).
Falsification Criteria
These are after-the-fact audit anchors for reconciliation, not conditions to wait on before acting — when a watchlist signal flips, the judgment updates first, without waiting for this section:
- TrendForce quarterly contract-price increases shrink to single digits or turn negative (next checkpoint: the 1Q27 guidance, around January 2027);
- A single quarter's operating margin falls below 70% without a one-off explanation from the company (the high-plateau floor gives way);
- Any long-term-agreement customer publicly cuts volumes, defaults, or renegotiates deposit terms (direct counter-evidence to the contracted thesis);
- Quarterly buyback resolutions come in below 10 trillion won for two consecutive quarters, or the year-end special dividend falls short of the commitment (return execution stalls);
- A close below 1,171,000 won (the October 23 intraday low) — the basing structure fails;
- A Chinese supplier achieves HBM3 mass-production qualification before 2027; or Samsung retakes share with HBM contract prices falling alongside (carrying over this series' existing watch parameters, unchanged).
Verification Calendar
| When | Event | What to read |
|---|---|---|
| This week | Samsung Electronics Q3 report | HBM4 shipment and yield framing; the share-redistribution vs price-cutting discriminator |
| November 6 | SanDisk earnings | The next checkpoint for the contract-book re-rating in US-listed memory |
| Late December | Micron's next-quarter guidance | Conventional DRAM spot pass-through and the HBM repricing framing |
| Around January 2027 | TrendForce 1Q27 contract prices | Double digits holding or single-digit dulling; first auditable in-band landings under the LTAs |
| Around end-January 2027 | SK hynix year-end dividend resolution | The Kioxia special-dividend balance delivered; second reconciliation of the return commitment |
| Quarterly | Buyback resolution amounts | The highest-frequency reading of management's confidence in forward cash flows |
To close. On September 16 we wrote "the strongest results, the most conservative pricing," and the core was a fight between two pricing grammars in two time zones; six weeks later, the results are stronger still — 104.6 trillion won of revenue, an 80% operating margin, 98.7 trillion of net cash — and, more importantly, the form of the evidence has changed: three sold-out years are a contract, the price floors and deposits are contracts, and 100 trillion won of returns is resolved cash. For the first time, the cycle has been written into contract text, while the market still bids 2.7-2.9x peak grammar — the gap is wider than in September and more auditable. The earnings-day rally is the first reading of that gap closing, but only the first: from here, watch the execution of the quarterly buybacks, where 1Q27 contract prices land, and the price's ability to hold itself up through every buyback gap window. The contracts are signed; the pricing's concession still has to be delivered one quarter at a time.
This note represents Lunartulip Lab's research views, based on analysis of publicly available information. It does not constitute investment advice. Markets carry risk; decisions require independence.