SK hynix.

Fair value against price. On 16 October 2026 the Seoul price of ₩1,272,000 is 0.2% below our central value of ₩1,275,044. The ADS close of $146.11 on 15 October is 53% above the same value converted, $95.73. The gap is the ADS premium, not the business.

The business. SK hynix leads HBM and earned a 76% operating margin in Q2 2026.^S3S14 Memory remains a cycle: the ten-year mean margin is 24.7% and 2023 was a loss.^S1 The balance sheet is strong, with ₩127tn of net non-operating assets after marking KIOXIA to 16 October. The open question is the margin after 2028, when ₩61.5tn of committed capacity arrives and a Chinese entrant scales.^S2S26

What moved since the 15 October build. Three inputs were closed from dated sources. The discount rate fell from 12.25% to 10.24% (regression beta, dated ERP and country premium); on its own that moves the central value from ₩1,063,976 to ₩1,275,044. Marking the KIOXIA bond to 16 October cuts ₩44,087 per share after tax. The 2026 tax law raises the statutory rate to 27.5%; with credits, the rate used is 23.0%.

Review status: provisional. The four-round cap was reached. Rounds 1–3 each found a resolvable fact and returned the work to phase A; round 4 changed nothing, but one unchanged round does not meet the two-round rule. Open items: post-2028 margin: 32% used, ten-year mean 24.7% gives 1,075,461 KRW; beta index and window: four regressions give 1,100,094 to 1,275,044 KRW; tax credits after 2028: statutory 27.5% from 2029 gives 1,219,808 KRW; H2 2026 pricing until the 28 October Q3 report; terminal RONIC 15% (judgement).

What would change the view. Q3 results and 2027 contract terms on 28 October; any shipment of CXMT 4F² server modules; a change to the ADS conversion limit; the beta choice (section 09).

No event since the H1 report changes the 2026–27 cash flows. The October fall matches the market move times beta: a de-rating plus rate repricing and flows. CXMT and the AI-funding reset challenge the margin after 2028, which is where the cases differ. They are tested as sensitivities, not yet as facts.

Window: from the H1 2026 report on 18 August to 16 October 2026. Each event is graded by source level and assigned to one channel. An event without a mechanism is sentiment and does not enter the value.

SK hynix makes DRAM, HBM and NAND in Korea and China. It lists in Seoul (000660) and, since July 2026, as ADSs on Nasdaq (SKHY).^S1 H1 2026 revenue was ₩131.9tn and operating profit ₩98.2tn.^S2

• Validated data. Financial figures come from SEC filings and company releases. Each figure states its source, period and whether it is reported or calculated by us. Conflicting sources are printed side by side.
• Classification first. The company is routed by sector, sub-sector and the shape of its profit history before any valuation. Metrics that do not fit the class are named and not computed.
• Facts before judgement. Inputs that a source can settle were closed first: the statutory tax rate, the risk-free rate, a regression beta, a dated equity risk premium and country premium, and the KIOXIA look-through. Items no source can settle sit in sensitivity tables and do not drive the review.
• Deterministic model. One script computes every value in this report from stated inputs and stored source files. Prose numbers are read from its output.
• Blind multi-reviewer challenge. Reviewers saw the report without the conclusion and looked for omissions, errors and the strongest opposing case. The review is stable when two rounds in a row keep the test result, move the central value by less than 1%, and find no new resolvable fact.
• Event scan. Events since the last filing are split by channel: estimate, multiple, discount rate, flows and read-across. Only events with a mechanism enter the model.
• Signal strength. Strong: filings and official data. Medium: industry trackers and dated academic datasets. Weak: press reports, columns and platform data, kept in chapter 15.
• Your weights. The calculator in the Summary lets you set case weights. The conclusion shows which judgement it rests on.
• Compared with typical AI-written research. Typical output blends sources and states one target. Here every number has a source grade, the cycle is normalised by rule, and the price is reversed into the margin and the return on equity it requires.

How much of today's profit is a cycle peak, and how much is a new level that HBM and long-term agreements can hold?

SK hynix files with the SEC as a foreign private issuer under SIC 3674, semiconductors.^S13 It reports one segment.^S2 The classifier routes SIC 3674 to the semiconductor class with the memory sub-route. Operating profit turned negative in 2023 and rose to ₩47.2tn in 2025, so the cycle test applies inside that route.^S1S15S16 We classify on operating profit, not net profit, because H1 2026 net profit includes ₩62.9tn of revaluation gains on financial instruments.^S2

Screen and check. The screen is price to book against the return on equity the business can sustain, shown in section 12 beside the cash-flow value. The check is the memory cycle-indicator panel in section 13. The class rule caps the terminal margin at the ten-year mean unless a structural shift is shown; section 10 states the evidence we use and the value at the mean.

Metrics we do not apply. Faustmann is not computed: it belongs to the profitable-industrial class. The current quarter is not a norm, and peak margin is not capitalised. One multiple for all semiconductor sub-sectors is not used. Average return on equity over five years is not used as the norm, because one memory cycle is longer. Analyst forecasts are not used to date the cycle.

DRAM was 77.3% of Q1 2026 revenue and NAND 22.0%.^S1 SK hynix ranked first in HBM with a 56.4% share of Q1 2026 HBM market revenue, according to the tracker cited in the prospectus.^S1 The company states long-term supply agreements with about ten customers.^S45S46

Two trackers disagree on DRAM share

We do not average them. The periods differ, and the second tracker says SK hynix's ASP gain was limited by its high HBM bit share in Q2.^S28 We use the TrendForce figure for competitor shares, because it covers all named suppliers in one period.

Customers

The largest customer was 23.9% of 2025 revenue. The two largest were 14.8% and 12.4% in Q1 2026.^S1 In H1 2026 they were 13.35% and 13.03%, against 27.31% for one customer in H1 2025.^S2 Concentration has fallen. The filings do not name them. The prospectus says sales concentrate in the United States and China.^S1 Reports describe SK hynix as an Nvidia supplier.^S45 That the largest customer is Nvidia is our inference and is Unverified.

The annual operating margin moved from −23.6% in 2023 to 74.4% in H1 2026, about three years. Normalisation therefore uses margin, not profit in won.

We anchor the low case on the simple mean, because the weighted figure gives the HBM years most of the weight and so assumes part of the answer. The weighted figure appears as a sensitivity in section 10.

All values are reported by the company; margins are calculated by us. The 2020 operating profit is read from a release that states +84% on 2019; treat it as Unverified to the second decimal.^S21

Q2 2026 net profit of ₩93.9tn exceeded operating profit of ₩60.5tn. Finance income was ₩65.9tn, including ₩53.2tn of valuation gains and ₩10.0tn of dividend income.^S2 The prospectus attributes the valuation gains primarily to the KIOXIA interest.^S1 We value operations from operating profit and add the investments once.

Cash includes cash, short-term instruments, short-term investments and long-term instruments at 30 June 2026.^S2 We deduct the ₩7.8tn deferred tax liability and the ₩32.0tn of current tax payable on H1 profit, because the cash flows start in H2.^S2 The ADS proceeds of ₩39.9tn arrived on 14 July, after the balance sheet date.^S4

KIOXIA: look-through from the filings

SPC 1 sold its remaining KIOXIA equity in June 2026.^S1 The remaining exposure is the SPC 2 convertible bond. The audited notes value it on KIOXIA's equity value and SPC 2's 14.30% ownership of KIOXIA, less a liquidity discount of 8.95% at 31 December 2025.^S1 In H1 2026, ₩66.08tn of long-term investments moved from Level 3 to Level 2 after a change in valuation method.^S2 The amount matches the bond marked to the KIOXIA share price, so we treat it as the bond's 30 June value. The prospectus also describes the bond as convertible into about 15.0% of SPC 2 equity; the valuation note is the basis we use, because it reconciles to the reported values below.

KIOXIA's issued shares come from its own release; the 31 December count was not disclosed there, so the 31 March count is used (*).^S55 Prices and exchange rates are daily closes from the stored series.^S56 All three reported values reconcile within 6%. The 30 June value implies a discount of 1.25%, which we carry forward. SPC 2 holds about 235m KIOXIA shares after the split. At ¥13,375 on 16 October the bond is worth ₩26.1tn, ₩40.0tn below 30 June. Deferred tax on that fall at the 27.5% statutory rate would be ₩11.0tn; we release only ₩7.8tn, the reported deferred tax liability, because the filings do not show that a net deferred tax asset would be recoverable. Net effect: ₩-44,087 per share.

What stays open. The notes say SK hynix's KIOXIA interest is limited to a certain percentage for a specified period, and SK hynix cannot appoint directors.^S2 The conversion terms of the bond are not disclosed. This is an irreducible item: it moves value through the discount, shown below, not through the share count.

The ₩40.0tn buyback runs to 19 November.^S5 The policy returns over 50% of 2025–2027 free cash flow.^S6 H1 2026 finance expenses include ₩4.0tn of non-cash losses on exchangeable-bond derivatives, offset in equity.^S11 We leave cash and the share count unchanged. If the whole programme executes at its implied average of about ₩1.66m per share, above the central value, value per remaining share falls to ₩1,261,829.^S5

Capital spending was ₩8.3tn in 2023, ₩15.9tn in 2024 and ₩27.5tn in 2025.^S1 H1 2026 spending was ₩18.3tn.^S2 Coverage of the Q2 call puts 2026 spending in the high-₩40tn range; the exact figure is Unverified, and we use ₩47tn.^S46 Unrecognised capital commitments rose from ₩6.7tn to ₩61.5tn in six months.^S2 The board approved Y2 for ₩35.2tn, M17 for ₩19.1tn and P&T7 for ₩7.1tn.^S7S8S9

Capital spending averaged 26% of revenue in 2023–2025.^S1 The central and high cases use 25% after 2027; the low case uses 22%.

The company does not split maintenance from growth spending. Depreciation is our proxy, and it is labelled as such.

The incremental return is mostly price, not volume: the trackers report modest bit growth and large price rises.^S28 The H1 2026 annualised return of 109% is a peak figure, not a norm. No purchase-obligation stress is run, so no loss rate is used; the commitments are capital spending, which the cases carry directly.

SK hynix sold 177.9m ADSs at $149.00 on 9 July 2026, for $26.51bn gross.^S1 Ten ADSs represent one common share. The offering issued 17.79m new shares, which is 2.4% of the 728.9m shares outstanding after it.^S1S4 There was no over-allotment option. The prospectus says the company and certain affiliates may agree a 90-day lock-up from its date.^S1 By our count it ended about 7 October 2026. Whether any locked holder has sold since is Unverified. Three cornerstone investors indicated up to $7bn.^S1 The prospectus does not describe a lock-up for them.

Holders

SK square must hold at least 20% under Korean holding-company law, and this capped the offering size.^S1 On outstanding shares its stake is 20.04% after the offering and 20.73% after the buyback cancels 24.07m shares. The legal test uses issued shares including treasury shares, which gives about 20.00% and 20.68%, and headroom of about 24.1m shares on issued shares; the formula below uses outstanding shares and gives more.^S1S5

The premium is structural

The Korea Securities Depository was reported to cap conversion into ADSs at 2.5% of shares, and the offering used the cap.^S40 New ADSs can be created only after others are cancelled. The prospectus warns that a holder who withdraws shares may not be allowed to deposit them again.^S1S42

Rows one and two are as reported; row one uses the prospectus conversion rate of ₩1,538.05 (noon rate, 2 July). At the 9 July Seoul opening base rate of ₩1,509.90 in the KRX filing, the premium was 2.9%.^S1S4S41 Rows three and four are calculated by us from closes on the same calendar day.^S22S34S49S56 The Seoul close comes about 15 hours before the US close, so a same-day premium also contains one US session of news.

If the conversion limit were lifted and the premium closed through the ADS price, the ADS would move by that amount with no change in the business.

The Seoul shares fell for seven sessions to ₩1,262,000 on 15 October, down 11.06% on the day, and closed at ₩1,272,000 on 16 October.^S22S23S56 That is 57.4% below the record close of ₩2,987,000.^S24 One source dates the high to 19 June intraday, another to 25 June at the close.^S24S25 A press report gives a record close of ₩2,919,000 on 22 June, which gives −56.4% at the 16 October close. Both press sources agree on ₩2,987,000; the exchange record itself was not opened, so the level is Unverified within about 2%. Foreign investors sold a net ₩1.11tn of the shares on 15 October.^S22 Over five sessions the shares fell 26%, Samsung 22% and Micron about 11%.^S43

The two-year raw KOSPI beta explains the whole fall; the residual is small and positive. Against Samsung the shares fell 9% more. In the US over 2–15 October the ADS fell 25%, Micron 13% and the semiconductor index 12%.^S56 SK hynix is a large part of KOSPI, so the market return partly contains its own. The test still shows a sector and market move, not news specific to the company.

What did not change. No SK hynix filing since 19 August changes guidance, contracts or capacity plans.^S10S12 Two September 6-Ks answered press reports on a Solidigm pre-IPO raise and a Chongqing packaging stake sale; nothing was decided. Trackers still expect DRAM contract prices to rise 10–15% in Q4 2026 and HBM prices to rise in 2027.^S29S30 The Q3 report on 28 October is the next hard data.^S10

Three events in the week before the sell-off involve AI companies raising public money. We separate what they say about HBM demand from what they say about valuation multiples.

For SK hynix the demand channel works with a lag. Agreements and HBM allocation cover most of 2027.^S30S46 The multiple channel works at once, through the discount rate and the terminal margin the market will pay for. Our cases change the margin after 2027, not the 2026–27 demand.

These inputs were closed from primary sources before any review round. Each is computed in the model from stored files.

Tax

From fiscal years starting on 1 January 2026 the top national corporate rate is 25%, and local income tax adds 10% of it.^S60S61S62 The top statutory rate is therefore 27.5%, against 26.4% in 2025.

The company attributes the low 2024 and 2025 rates primarily to tax credits on capital spending.^S1 H1 2026 uses management's estimate of the full-year effective rate.^S2 We use 23.0% for H2 2026 to 2028. From 2029 and in the terminal value we use 27.5% less a sustainable credit of 4.5%, the smallest gap of the three periods: 23.0%. The smallest gap is the conservative choice, because credits scale with capital spending and the 2026 profit base is the highest in the series. If credits end, the statutory rate applies; that row is in section 10.

Discount rate

The KTB yield on 16 October was 4.20%; the 15 October close was 4.193%.^S57S33 A Korean government bond is not default-free, and the country premium already prices Korea risk, so we deduct the rating-based default spread from the same country file (0.39%, Aa2).^S59 Without the deduction the central value is ₩1,224,860. The equity risk premium is the implied premium for the S&P 500 on 1 October 2026, 4.20%; the July 2026 country file uses the same mature-market figure.^S58S59 Korea's country premium is 0.61% in the July 2026 file, against 0.64% in January.^S59

Beta is a regression of weekly returns. Against KOSPI we use won returns. Against MSCI ACWI we use dollar returns, with the share price converted at the weekly won rate. Blume adjustment: 0.67 × raw + 0.33.^S56

The regression measures the beta of the equity, which holds net cash. Net cash (cash and financial assets plus ADS proceeds, less borrowings, leases and tax payable) has a beta near zero and is added separately, so the operating cash flows carry the higher operating beta. KIOXIA is left in the risky part. The table's r and values use the operating beta for each regression. With the unadjusted equity beta the central value would be ₩1,358,263.

We use the five-year ACWI beta. The premium is a mature-market premium, so the index should be global. Five years cover the 2022–23 downturn and the HBM upturn; two years cover only the upturn. SK hynix is a large part of KOSPI, which pulls the local beta towards its own volatility. The choice is judgement, and the four rows show its cost: the KOSPI two-year beta gives a value 14% lower.

Working capital is not released when revenue falls; in past downturns inventory rose.

All cases share H2 2026. Revenue is twice Q2, converted from the Q2 average of ₩1,501.93 per dollar to ₩1,332, with a 6% Q4 contract-price uplift: factor 0.940.^S51S52S29 Won costs stay at the Q2 level, so the H2 margin falls below Q2. Cash flows are discounted from 16 October: H2 2026 at 0.25 years, each later year at its mid-point (year minus 2026, less 0.29), and the terminal value at 8.71 years, the same mid-year timing. Revenue grows 6% a year from 2030 to 2035 in the low and central cases and 8% in the high case, then 3% in the terminal value. The cases differ from 2027. Cycle repeats copies the 2019 and 2023 revenue falls of about one third and returns to the ten-year mean margin. Plateau holds 2027, then falls, and settles at 32%. New trend line treats HBM and long-term agreements as a higher floor, at 42%.

Terminal margin above the cycle mean. The class rule allows a terminal margin above the ten-year mean of 24.7% only with evidence of a structural shift. Our evidence: HBM is sold under long-term agreements with about ten customers, with price bands on part of supply;^S45S46 SK hynix held 56.4% of Q1 2026 HBM revenue;^S1 DRAM supply stays short of demand in 2026 and 2027 per one tracker.^S63 Against it: CXMT's denser DRAM, and ₩61.5tn of committed capacity.^S26S2 The evidence is medium strength. At the mean the central value is ₩1,075,461; at the revenue-weighted mean of 29.7% it is ₩1,212,248.

The ADS values convert the per-share value at the spot rate and carry no premium.

The matrix holds 2027 at +10% revenue and a 65% margin. It then sets the margin from 2029 (rows), with 2028 at the midpoint between 65% and the row margin, and the 2029 revenue as a multiple of 2025 revenue (columns). Values are per share in thousands of won; bold cells are at or above the ₩1,272,000 close.

At the ten-year mean margin of about 25%, no column reaches the price. At a 35% margin the price needs 2029 revenue of about 2.1 times 2025, or ₩207tn; at 40%, about 1.8 times, or ₩175tn (interpolated). The 2026 estimate is ₩281tn; no year before 2026 exceeded ₩97tn.

For memory the class route values book equity on the return it can sustain over a cycle. A business that earns ROE on book and grows at g is worth this multiple of book:

At mid-cycle operating ROE the justified value is ₩858,920 per share; at the central terminal ROE it is ₩1,412,545. The price needs a sustained ROE of 28%. That is below the 2024 and 2025 figures and above the mid-cycle figure. The cross-check agrees with the cash-flow model: the price is fair only if the margin after 2028 stays well above the old cycle mean.

The test uses operating profit for the mid-cycle row, so financial income is excluded. 2023 ROE is not shown: the prospectus gives 2023 year-end equity but not 2022.

The class route requires a panel of cycle indicators. Each row gives the latest reading and its direction for the margin after 2027.

Five rows point to tight supply into 2027. Two point to a supply response after it. The panel supports the shared H2 2026 and 2027 inputs. It does not settle the margin after 2028, which is where the cases differ.

The review ran in two phases. Phase A closed every input that a source can settle, before any reviewer saw the report. Phase B gave independent reviewers the report and the source list without the conclusion, the case names or our open questions. One reviewer looked only for omissions that could change the conclusion. Others argued the strongest positive and negative case and rechecked numbers against sources. A new resolvable fact found in phase B sent the work back to phase A.

Phase A: facts closed from sources

Phase B: judgement rounds

Status: provisional. The four-round cap was reached. Rounds 1–3 each found a resolvable fact and returned the work to phase A; round 4 changed nothing, but one unchanged round does not meet the two-round rule.

This chapter holds context that is not evidence. It does not move the cases.

• A Seoul column argues that memory oversupply starts in 2029. It is opinion.^S44
• A Korean industry figure argues that more capacity lets Korean makers win back price-sensitive share from Chinese suppliers.^S50
• US bonds of a large GPU cloud provider traded near 77 cents on 14 October, according to a market report.^S48 Credit stress at one buyer is not a demand fact for HBM.
• Chief executives of two large chip companies named financing costs as a constraint on AI spending.^S47
• US export-control licences for the Chinese fabs and chip tariffs are listed as risks in the prospectus; we did not size them, and their effect is Unverified.^S1
• SK hynix is reviewing more production bases, including Japan; nothing is decided.^S12

Level 1 is a regulator or official dataset, 2 audited or reviewed accounts, 3 company releases, 4 industry trackers, 5 independent press, 6 platform signals. Conclusions rest on levels 1–3; levels 4–6 set context and are graded in the text.

A1. Annual forecasts, ₩tn

A2. Valuation outputs

A3. Base inputs and provenance