SK Hynix

Refresh delta (since 2026-07-07)

What changed — the material moves

1. Q2 2026 printed an all-time record AND MISSED. Revenue ₩79,318,746M (₩79.32T), +50.9% QoQ / +256.8% YoY; operating profit ₩60,542,608M (₩60.54T), +61.0% QoQ / +557.2% YoY; operating margin 76.33% [primary: 6-K 2026-07-29, acc 0001193125-26-321989]. That breaks the 72% the prior dossier called "the most mean-reverting number in hardware" — upward. But the Korean street consensus was revenue ₩83.40T / OP ₩64.68T (LSEG SmartEstimate OP ₩64T), so the print missed OP by ~6.4% and revenue by ~4.9%. Cause given: HBM4 shipments below plan, revenue recognition pushed into H2 — management's own phrasing was that some high-value products were "pushed back into the second half".
2. Two-thirds of the headline net income is a non-repeatable Kioxia gain. Pre-tax profit ₩122,708,355M exceeded revenue [primary: 6-K 2026-07-29]. Net non-operating profit was ₩62.2T, of which ₩63.27T is gains on the sale and valuation of investment assets — the Kioxia stake, held since 2018, sale closed June 2026. Reported net income ₩93.92T is therefore not an earnings-power number. Stripping it at the Q2 effective tax rate of 23.46% leaves clean net ≈ ₩45.5T. Any screen quoting SK Hynix's trailing P/E right now is quoting a contaminated denominator.
3. The ADR sold the news — the prior seed's own falsifier fired. SKHY listed Jul 10 at $149.00, peaked at $193.92 (Jul 14, +30%), bottomed $126.79 (Jul 29, the print) and sits at $135.90 (Aug 10) — below its IPO price. The Korean line fell harder: ₩1,420,000 (Aug 10) vs ₩2,164,000 on Jul 7 = −34%, and roughly −52% from a ~₩2.9–3.0M July peak. Market capMarket capitalisationThe share price multiplied by the number of shares. What the market says the equity is worth. ₩1,034.99T ≈ $730B, down from ~$1.12T on Jul 7. Context that cuts the other way: the stock is still 5.8× its 52-week low (₩245,000). This is a parabola giving back half, not a value stock breaking down.
4. A ~35% ADR premium opened, and arbitrage is structurally blocked. 10 ADS = 1 ordinary share. Parity per ADS = ₩1,420,000 ÷ ₩1,417.60 ÷ 10 = $100.17, versus SKHY at $135.58 → +35.3% premium, corroborated at ~33% with a 2.5% conversion cap described as an "iron curtain". Which line you own is now a first-order decision, not a formality — a US buyer's effective forward multiple is ~4.65x, not 3.44x.
5. The "slowing HBM expansion" debate is over — decisively on the expansion side. FY2026 capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. guided to the "high ₩40 trillion range" vs FY25 actual ₩27,519B [primary: 424B4 for the FY25 base] — clearing the ≥₩40T threshold the 2026-07-20 delta note pre-registered as killing the claim. Then on 2026-08-07 the board approved ₩35,224,600M (Y2, Yongin Fab 2 phases 1–6) + ₩19,100,000M (M17, Cheongju) = ₩54.32T of NEW committed capex [primary: 6-Ks acc...-339024 and...-339011], on top of ₩7,093,100M for P&T7 on Jul 22. M15X mass production was pulled forward — the single near-term HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips. supply variable, accelerated not stretched. The delta note's watch item — does announcement convert to board commitment? — resolved yes, in three weeks.
6. Contract structure disclosed for the first time: LTAs with ~10 customers. SK Hynix "finalized Long-Term Agreements (LTAs) with around 10 customers, including key strategic partners", and said "our major customers are still requesting more memory supply" with "tight supply-demand conditions expected to persist for a considerable period". This partly retires the 2026-07-20 caveat that "sold out" appeared nowhere in the primaries — there is now a first-party contract disclosure. It is still not the phrase "sold out," and no volumes or prices are given.
7. A new competitor listed: CXMT. China's ChangXin Memory priced an $8.6B IPO (largest mainland listing since 2010), closed +466% on debut 2026-07-27, and became China's most valuable A-share at ~$480B — now the world's #4 memory maker. Critically, its prospectus earmarks nothing for HBM — proceeds go to conventional DRAM lines and process upgrades. The threat is to the conventional-DRAM rent that carried Q2 margin, not (yet) to HBM.
8. Balance sheet transformed. Cash ₩88T (+₩33.6T QoQ), total debt ₩18.6T (−₩0.7T), net cash ₩69.4T. Management: financial flexibility "significantly strengthened."
9. The dilution bear point is much weaker than it looked. Against 17,790,000 new shares issued for the ADR, the company cancelled 15,300,000 treasury shares YTD [primary: 6-K 2026-08-07, acc...-339002]. Net share change ≈ +2.5M on 728.87M, ~+0.34%. Share count is now primary-derived: the Q2 dividend 6-K gives ₩273,324,801,750 ÷ ₩375/share = 728,866,138 shares [primary: 6-K 2026-08-07, acc...-339031] — which reconciles exactly to the ₩1,034.99T market cap. The prior dossier's 708.3M vendor figure was ~2.9% light.
10. Two live corporate-action rumors, both formally "nothing determined." Solidigm pre-IPO raise ~₩5T [primary: 6-K 2026-08-06] and a Chongqing packaging-plant stake sale ~₩4T [primary: 6-K 2026-08-10 — filed today]. Both carry a one-month disclosure clock. Read together with the Intel-Ohio denial (Jul 22), the pattern is portfolio reshaping around the China footprint and subsidiary-level funding, not core-fab retreat.

What held — the structural thesis stands, and strengthened

Every load-bearing element of the prior thesis is intact or better. SK Hynix remains #1 in HBM revenue share (56.4%, IDC, Q1'26) and #2 in DRAM incl. HBM (29.1%) [primary: 424B4]; HBM4 entered mass production first, with yields "nearing the levels of HBM3E, which is already in the maturity stage," full ramp H2'26 and HBM4E volume 2027; margins made a new record, not a lower high; the balance sheet went to ₩69.4T net cash; and the capacity program converted from announcement to board-approved commitment. "Rising — durable ✅" per rent-is-moving-thesis.md survives this print. What changed is the price, not the position.

Phase A — Understand the business

Company Overview (HBM4 in production, capacity committed)

World's #1 HBM supplier by revenue share (56.4%, IDC Q1'26) and #2 in DRAM incl. HBM (29.1%) [primary: 424B4]. DRAM 77.3% of Q1'26 revenue, NAND 22.0% [primary: 424B4; now in our figures]. Crown jewel of SK Group (SK Square ~20.07%). New this pass: HBM4 moved from qualification to mass shipment in Q2, with full ramp guided for H2'26 and HBM4E volume production 2027; on NAND, 321-layer product is already the largest share of output, targeted to ~50% of domestic capacity by year-end. The company is no longer planning the next capacity wave — ₩61.4T across P&T7, Y2 and M17 is board-approved as of Aug 7 [primary: three 6-Ks]. The prior dossier's "voluntary DDR5-over-HBM4 throttle" narrative did not survive the call: management described managing "sales mix between HBM and conventional DRAM based on customer demand and medium-to-long-term product strategy" — mix management, not a throttle, and it declined to quantify the split.

Supply Chain (carried — unchanged; see the previous dossier)

Double chokepoint: ASML EUV + TSMC 12nm HBM4 logic base die ("One Team"), itself the binding downstream constraint on NVIDIA. Effective HBM ceiling = stacking yield × CoWoSCoWoSTSMC’s method of packing a processor and its memory onto one carrier so they sit close together. Supply of it has been a hard limit on how many AI chips can be built... One update worth noting inside a carried lens: the 424B4 commits ₩11.9T to EUV scanners for December 2027 delivery — the physical option on the 2028+ ramp is bought.

Competitive Advantages (HBM4 first to production, but a new entrant below)

The lead is execution-and-yield, and Q2 sharpened both sides of it. For: SK Hynix reached HBM4 mass shipment first, with yields described as approaching mature HBM3E levels — the hardest thing to copy in this business — and converted ~10 customers to long-term agreements, which is a structural change in bargaining position, not a price event. Against, and new: the same quarter missed on HBM4 shipment volume, proving the ramp is not frictionless; and CXMT arrived with $8.6B and a ~$480B market cap. The correct reading of CXMT is narrow but real — no HBM in its prospectus, so it does not touch the HBM rent near-term, but it is aimed squarely at conventional DRAM, which is where the +198% YoY ASP forecast and a large share of Q2's record margin actually came from [primary: 424B4 cites Gartner; web: Tom's Hardware]. The 2026 margin is more exposed to Chinese conventional-DRAM supply than the HBM-centric framing implies. That is the single most under-priced structural change this pass.

Segments (carried — see the previous dossier)

Q1'26 DRAM 77.3% / NAND 22.0% now seeded to our figures[primary: 424B4]. Q2'26 segment split was NOT disclosed in any SEC filing — n/a, not estimated. HBM as a % of revenue remains undisclosed in every primary source — the #1 value driver is still modeled, never reported.

Phase B — Measure performance

Earnings Result (Q2'26 actuals, primary)

[primary: 6-K 2026-07-29, acc 0001193125-26-321989 — preliminary, unaudited K-IFRS, "may be subject to change"]; margins. H1'26 YTD: revenue ₩131.90T, OP ₩98.15T, net ₩134.27T. All three quarters are now in our figures.

Versus consensus — a record that missed. Street: revenue ₩83.40T / OP ₩64.68T; LSEG SmartEstimate OP ₩64T. Actual missed OP by ~6.4% and revenue by ~4.9%. Guidance (Q3'26): DRAM bit shipments ~+10% QoQ, NAND bits up a low single-digit %; FY26 capex "high ₩40T range".

Balance-sheet flags — three, and the first is the important one.

• (a) The headline net income is ~52% non-operating. ₩63.27T of investment-asset gains (Kioxia) sit in other non-operating income. Clean net ≈ ₩45.5T. Every trailing multiple on this name is currently wrong by roughly 2×.
• (b) Net cash ₩69.4T (cash ₩88T, debt ₩18.6T) — the ₩39.89T raise plus the Kioxia proceeds landed simultaneously. This is the strongest balance sheet in the company's history, entering the largest capex program in its history.
• (c) Inventory, receivables, PP&E, total assets and equity were NOT disclosed in any SEC filing — the Korean quarterly report goes to DART, which is not machine-ingested. The inventory-build check the prior dossier wanted remains impossible to run. That is a real gap, not a clean bill: an ASP-driven revenue print with unverifiable inventory is exactly where a memory cycle turns first.

Market reaction: SKHY −8.98% to $130.17 on the print, then to $126.79 the next session; a +17.5% bounce to $149.00 on Jul 30 on Amazon AI-demand projections through 2028; $135.90 on Aug 10.

Earnings Calls — sentiment shift vs the prior 3–4 calls **

The tone moved from "sold out" to "contracted," and from swagger to defensiveness. Across 1Q26 → Computex → 2Q26 the through-line was demand exceeds capacity for ~3 years. This call replaced the unsourced "sold out" framing with something more durable and more auditable — LTAs with ~10 customers — while noticeably hedging the language: "tight supply-demand conditions expected to persist for a considerable period" is weaker than "structurally short to 2030," and management declined to quantify HBM-vs-conventional wafer allocation. Three shifts worth naming:

1. Defensive posture on oversupply. The CFO addressed AI-demand scepticism directly: "We view these developments not as a sign of AI investment slowdown, but more as a transition towards higher utilization of the AI infrastructure that has already been built at scale". Management is now arguing with the bear case on the call — it was not doing that in April.
2. The capex message inverted. April's frame was "heavy capex with a balance-sheet guardrail (net cash >₩100T)." July's is acceleration — high-₩40T guidance, M15X pulled forward, ₩54.3T board-approved eight days later. The guardrail language receded as the net-cash position hit ₩69.4T.
3. The DDR5-throttle story was quietly dropped. What the prior dossier logged (from commentary sites) as a deliberate throttle, management described as ordinary mix management. Treat the "voluntary throttle" as unconfirmed by the company and retire it from the thesis — it was single-source-soft and did not survive first-party contact.

Comps & Valuation — and the two ACTIVE MarketCalls **

Same-source (stockanalysis.com), as of 2026-08-10:

Three corrections the headline table needs, and they matter:

• The trailing multiple is contaminated. Strip the Kioxia gain from TTM net income: ₩162.08T − ₩63.27T = ₩98.81T → trailing P/E ≈ 10.5x, not 6.23x. Still well below MU's 19.80x, but the gap is half what the screen shows.
• A cleaner operating read. EVEnterprise valueWhat it would cost to buy the whole company: its market value plus its debt, minus the cash you would get with it. Often very different from market cap. = ₩1,034.99T − ₩69.4T net cash = ₩965.6T. Against Q2 operating profit annualised (₩60.54T × 4 = ₩242.2T) that is EV/EBIT ≈ 4.0x; against H1 annualised (₩196.3T), 4.9x. Four times peak operating earnings, net of a fortress balance sheet. That is the entire bull case in one number — and the entire bear case is the word "peak."
• The forward P/E of 3.44x is not available to a US buyer. Through the ADR at a ~35% premium it is ~4.65x.

MarketCall assessment (review finding only — no row touched):

Call A — "HBM Suppliers Are the Picks-and-Shovels of AI: SK Hynix Leads" (MEDIUM) → SUPPORTED on substance, WEAKENED on framing. The leadership premise is more established than when written: first to HBM4 mass production, yields approaching HBM3E maturity, 56.4% HBM revenue share, and ~10 customers now on long-term agreements. But "picks-and-shovels" carries an implied lower-volatility, toll-road reading, and the quarter falsified that implication twice — a −6.4% OP miss caused by the picks-and-shovels product itself slipping (HBM4 shipment timing), and a −52% peak-to-trough drawdown. The call is right about who leads and wrong about how it will feel to hold. No premise break.

Call B — "SK Hynix: the largest HBM rent, valued like a Korean cyclical" (MEDIUM) → SUPPORTED, and roughly twice as true. Both halves moved in the call's favour simultaneously. The rent got bigger: 76.33% operating margin (a record), ₩60.5T quarterly OP, ₩69.4T net cash, LTAs signed. The valuation got more cyclical: forward P/E 3.44x vs 6.06x on Jul 7, EV/EBIT ~4.0x on annualised Q2. The spread the call names has approximately doubled in five weeks. The honest risk to the call is that it is not a mispricing at all — the market may simply be forecasting 2027–28 correctly, and the ₩54.3T of new capacity landing 2029–31 plus CXMT is the mechanism by which it would be right. The call's edge rests entirely on durability, and durability is the one thing this print did not settle.

Stock-Price Catalysts (±5% moves since Jul 7)

The prior dossier's "leveraged bet on one demand narrative" pattern held, violently, in both directions.

. The sharpened pattern: the ADR listing marked the top almost to the day. The instrument now has two prices (a 35% ADR premium with a 2.5% conversion cap blocking arbitrage), which adds a second, non-fundamental source of volatility for US holders. Next catalysts: Q3'26 print (late Oct); HBM4 full-ramp evidence in H2; the Solidigm (₩5T) and Chongqing (₩4T) decisions, both on one-month disclosure clocks from Aug 6 / Aug 10.

Phase C — Judge people & books

Management & Governance (the dilution story reversed)

CEO Kwak Noh-jung; CFO Kim Woohyun; IR head Seonghwan Park signs every 6-K. The governance read improved materially, against expectation.

• Dilution largely neutralised. 17.79M new shares issued for the ADR vs 15.30M treasury shares cancelled YTD → net ~+0.34% [primary: 6-K 2026-08-07]. The prior dossier's "chaebol is diluting ~$29B into a fearful tape" is substantially overstated as a share-count matter. (It remains true as a market-supply matter — fresh tradeable float is what the July drawdown ran into.)
• Quarterly dividend maintained at ₩375/share, record date Aug 31, total ₩273.3B [primary: 6-K 2026-08-07] — token yield (~0.03%), but a maintained cadence through a capex surge.
• Board process visible and clean. All four August resolutions record 6 of 6 independent directors present [primary].
• Straight dealing with rumours. Three prompt clarifications in three weeks — Intel-Ohio (Jul 22, denied), Solidigm raise (Aug 6), Chongqing stake sale (Aug 10) — each with a one-month follow-up commitment. This is better disclosure hygiene than the chaebol discount assumes.
• Open flag: a ₩7.09T P&T7 approval (Jul 22) that the filing calls "an increase," unreconciled against the ₩19.0T P&T7 total in the 424B4 use-of-proceeds table. Flagged, not merged — carried forward from the pre-print file, still unresolved.

Forensic Red Flags **

1. Earnings quality — now the top flag. ₩63.27T of the ₩93.92T net profit is investment-asset gains. Pre-tax profit exceeding revenue is a screening artifact that will flatter every value filter on this name for four quarters. Operating profit ₩60.54T is the only line to underwrite.
2. Unverifiable working capital. Inventory, receivables and PP&E were not disclosed to the SEC and DART is not ingested. In a memory upcycle driven by ASP, inventory is the leading indicator of the turn, and it is dark to this dossier.
3. Peak-cycle accounting, at a new peak. 76.33% operating margin. The 2023 comparison remains the anchor: this company printed operating losses three years ago.
4. Capex vs durability — materially larger this pass. ₩61.4T board-approved in 17 days, projects completing 2031, against a demand book contracted with ~10 customers on undisclosed terms. The 424B4's own phasing puts ₩9.3T in 2028, ₩8.3T 2029, ₩7.8T 2030 — spending lands well after the visibility does.
5. Customer concentration. Two largest customers 14.8% and 12.4% of Q1'26 revenue; largest 23.9% of FY2025 [primary: 424B4].
6. China exposure in motion. The Chongqing packaging stake sale under review (~₩4T) is the second China-footprint signal in a month.
7. Regulatory: a putative antitrust class action by indirect purchasers of conventional DRAM alleging supply restriction via "capacity shifts toward HBM production" [primary: 424B4, Legal Proceedings — allegation, not a finding]. No independent regulatory verification run this pass — regulatory/ remains absent; SEC EFTS enforcement coverage does not apply to a foreign private issuer, and DART was not scanned. Label: not an audited clean bill.

Phase D — Project & stress-test

Forward Projection (EPS, 3 years) **

MODEL INTEGRITY — read before using any number here. our model rebuilt on the 3 new quarters, but our model reports "No computed values" — 51 formulas, zero cached results. Nothing in the workbook is citable. It also flags: opening balance sheet partly sourced (2/10 lines); gross margin seeded at a 50% default (actual is far higher — the seed is a placeholder, not a finding); revenue growth seeded at 8% because there are fewer than 8 quarters; blank share price. I therefore cite NO model output. Everything below is with arithmetic shown. Fixing this needs the DART balance sheet, which is the highest-value ingest task outstanding.

Calendar years, KRW. Shares 728.87M [primary-derived: dividend 6-K]. EPS stated ex the Kioxia one-off so the years are comparable; FY2026 reported EPS will be ~₩66,000 higher.

Read. At ₩1,420,000 the market pays ~5.9× base-2026 clean EPS and ~6.2× base-2027 — but only ~4.5× bull-2027 and ~29.6× bear-2028. The market is not pricing the base case; it is pricing something close to the bear path arriving early. That is the mispricing the active calls assert. The honest counter is that FY2026 is nearly locked (H1 banked, Q3 guided) and the market knows that — so a 3.44x forward multiple is a statement about 2028, not 2026, and 2028 is precisely where ₩61.4T of new capacity and a $480B Chinese entrant collide. The bet is not "is 2026 good." It is "does the 2028 trough clear ₩100T of net income."

No our model create executed — unattended pass, per standing rule.

Bull vs Bear **

Bull. You are paying ~4× peak operating earnings, net of ₩69.4T of cash, for the #1 supplier of the single binding input to AI compute — a company that just posted a 76.33% operating margin, moved HBM4 into mass production ahead of both rivals at maturing yields, locked ~10 customers into long-term agreements, and committed ₩61.4T to the next capacity wave from a fortress balance sheet. The stock is 52% below its July peak on a 6.4% miss caused by shipment timing, not demand — management's own account is that the product moved to H2, and customers "are still requesting more memory supply." Bear. Three things are true at once and they compound. (1) The reported earnings are not the earnings — ₩63.3T of Kioxia gains inflate every trailing screen, and the clean trailing multiple is 10.5x, not 6.2x. (2) The margin that produced the record is partly a conventional-DRAM rent (+198% YoY ASP forecast) that a newly-capitalised $480B Chinese entrant is aiming directly at — CXMT needs no HBM capability to break that leg. (3) The company is committing ₩61.4T to capacity completing 2031 on a demand book whose volumes and prices are undisclosed, having just demonstrated it can miss its own shipment plan by enough to cost 6.4% of operating profit. A 3.44x forward multiple is not a market error; it is a market forecast of 2028. Pre-mortem (2028). HBM4 supply from three qualified vendors plus HBM4E arrives together in 2027; conventional-DRAM ASPs break as CXMT's IPO-funded lines come up; the LTAs turn out to fix volume but not price; the Kioxia gain rolls out of TTM and the "cheap" screen disappears; margins go 76% → 35%; and the ₩54.3T approved in a single August board meeting is still being poured into a glut. The ADR premium collapses on top, so US holders take the de-rating twice. Contrarian view. The market is reading a timing miss as a demand miss and a one-off gain as a quality problem, and pricing 2028 into a stock whose 2026 is already banked. The genuinely new information in the quarter — LTAs with ~10 customers — is the first evidence the contract structure is becoming infrastructure-like, and it arrived on the day the stock bottomed. The variant perception is that the July drawdown was a float event (₩39.89T of fresh supply meeting a CXMT-triggered sector de-rating), not an earnings event.

Devil's Advocate (short-seller) **

Look at what actually happened, not what was announced. A company that is genuinely supply-constrained does not miss its shipment number. SK Hynix missed operating profit by 6.4% because it could not ship the HBM4 it planned to ship — in the quarter it chose to call a triumph. Then it reported a pre-tax profit larger than its revenue, because a 2018 Kioxia punt paid off, and let every screen in the world print a 6.2x trailing P/E on earnings it cannot repeat. Meanwhile the actual driver of the record margin is conventional DRAM at a +198% ASP forecast — a rent with no moat, which is why a Chinese company with no HBM programme just raised $8.6B, closed up 466%, and became China's most valuable listed company aiming straight at it. Management's answer to demand scepticism is a rhetorical construction — "not a slowdown, a transition towards higher utilization" — which is what you say when utilisation is the thing being questioned. And the board's response to a 35% drawdown was to approve ₩54.3T of new fabs in one sitting, completing 2031, with the working-capital line that would show the cycle turning undisclosed to the SEC. The ADR is the tell: the largest listing in history priced Jul 10, topped Jul 14, and trades below issue — while a 35% premium over the Korean line, walled off by a 2.5% conversion cap, means the US price isn't even a real price. You are being offered peak-cycle economics, contaminated earnings, an invisible balance sheet, and a fresh Chinese competitor, at four times a number that has never held for two consecutive years in this industry's history.

Management Questions (carried + 4 added)

Carried 13 from the previous dossier. Added this pass: 14. The LTAs with ~10 customers — do they fix volume, price, or both, and over what tenor? (Highest info value on the page: it is the difference between infrastructure and a cyclical.) 15. Q2 HBM4 shipments missed plan — was the constraint stacking yield, CoWoS allocation, or customer pull-in timing, and is the H2 ramp de-risked against the same cause? 16. With CXMT capitalised at ~$480B and targeting conventional DRAM, what share of the Q2 operating margin came from conventional DRAM ASP, and what does that leg look like at 2024 ASPs? 17. ₩61.4T board-approved in 17 days for projects completing 2029–2031 — what demand evidence underwrites the 2031 tranche, and what is the trigger to defer it?