company:SK hynix Inc. (Nasdaq ADS: SKHY; KRX: 000660)
sources used:424B4 prospectus (10 Jul 2026, data to Q1 2026); Q2 2026 earnings call (29 Jul 2026); 6-Ks 15 Jul–4 Sep 2026, incl. preliminary Q2 results (29 Jul), H1 2026 semi-annual business report (18 Aug), ₩40tn share cancellation (19 Aug); Q1 2023 call for trough behaviour.
latest period:Q2 2026 (quarter to 30 Jun 2026); material events to 4 Sep 2026.
missing or stale data:No 20-F, annual report or earnings slides in sources. HBM revenue and margin not disclosed. LTA coverage, price floors and deposit sizes not quantified. Q2 2026 figures provisional (unreviewed). No market price or valuation data used.

SK hynix — Bull & Bear Memo

1. Business in one line

SK hynix turns wafers into DRAM and NAND bits and sells them at industry-set prices; profit is price × bits − cost per bit. DRAM was ₩97.6tn (74%) and NAND ₩33.5tn (25%) of ₩131.9tn H1 2026 revenue H1 report. The differentiator is HBM — stacked DRAM for AI accelerators — where it holds 56.4% revenue share, against 29.1% of total DRAM and 18.5% of NAND (Q1 2026, IDC) 424B4.

CyclicalA cyclical with a fast-grower franchise (HBM) inside it. The type matters: Q2 2026 revenue rose 51% QoQ while DRAM bits grew only high-single-digit and ASP ~30% Q2 call. Price is doing the work, so the bear case lives in the cycle — and the bull case must show the cycle has been partly contracted away.

2. Bull case — Peter Lynch pitch

The simple reason this stock could work is that AI turned the scarcest memory product into one that SK hynix makes more reliably at volume than anyone else — and the company is using record cash to lock customers into multi-year contracts and shrink its share count, so the next trough sits higher than any before it.

A1HBM leadership is a manufacturing moat, not a timing lead
What must happen
Majority HBM share and a price premium hold through HBM4 (ramping H2 2026) and HBM4E (volume from 2027).
Why this company
HBM is won on yield, quality and delivering volume on the customer's schedule, not on a spec sheet. SK hynix has supplied every generation since HBM2E; HBM4 entered mass production in Q2 with yield and quality "nearing the levels of HBM3E", and HBM4E samples are delivered Q2 call. A customer's switching cost is a qualification cycle plus system-failure risk.
Evidence
56.4% HBM share vs 29.1% of DRAM overall 424B4. HBM carried a >5× per-GB premium to conventional DRAM in 2025 (Gartner) 424B4.
What to monitor
2027 HBM price and volume settlements (under negotiation); share on major accelerator platforms.
A2AI demand absorbs supply through at least 2027
What must happen
Hyperscaler AI spend holds and efficiency gains widen usage instead of cutting memory per task, keeping supply tight "for a considerable period" Q2 call.
Why this company
Its mix is aimed at all three AI tiers: HBM, server DRAM (SOCAMM2 on 1c nm) and eSSD — eSSD revenue doubled QoQ in Q2 and Solidigm's 30TB+ drives more than tripled Q2 call. HBM also eats more wafer per bit, which tightens conventional DRAM for everyone.
Evidence
Enterprise share of the memory market 26.5% (2020) → 43.1% (2025) → 51.9% forecast (2027) (Gartner) 424B4. Customers "still requesting more memory supply" Q2 call.
What to monitor
Whether bits take over from price: Q3 guide is DRAM bits ~+10% QoQ, NAND low-single-digit Q2 call.
A3Long-term agreements raise the trough
What must happen
The ~10 LTAs (≈5-year terms, volume commitments, deposits, price mechanisms) act as floors, so the next downturn compresses margins instead of producing a 2023-style ₩9.1tn loss 424B4.
Why this company
Only a supplier holding scarce HBM capacity has the leverage to demand deposits and multi-year volume commitments at the top of a cycle.
Evidence
Management statement only. It "cannot say how much of our total sales will be covered by LTAs" Q2 call.
What to monitor
Disclosed LTA coverage; deposits and contract liabilities on the balance sheet.
A4Capacity comes in stages and surplus cash comes back
What must happen
Fabs are equipped against confirmed demand, and excess cash is returned rather than spent on capacity.
Why this company
It cut capex by more than 50% in 2023 Q1'23 call. It now holds ₩69.4tn net cash at 7% debt/equity Q2 call; H1 2026 operating cash flow was ₩91.7tn against ₩18.3tn PP&E capex H1 report.
Evidence
₩40.0tn open-market purchase and cancellation of 24.07m shares — 3.3% of 730.5m issued — by 19 Nov 2026 6-K 19 Aug.
What to monitor
Capex relative to operating cash flow; buyback completion; phasing of Yongin.
Why the market might be missing it: the market treats memory as a pure cycle where every peak reverts to the prior trough. If HBM qualification lock-in plus deposit-backed LTAs make even part of revenue contractual, trough earnings — the number that sets a cyclical's value — step up. That shift is invisible in reported numbers until the first downturn tests it.

3. Bear case — Munger invert

The most likely way I lose money is that the industry — SK hynix included — commits capacity at 76% operating margins, and that capital lands in 2027–2030 just as price, which did nearly all the work in 2026, turns.

B1Peak-margin capex becomes permanent overcapacityPermanent

Attacks A4 Capacity comes in stages and surplus cash comes back

How it could fail
Management targets doubling wafer capacity within five years, with ~₩600tn planned for Yongin (four fabs; first cleanroom of the fourth by 2033) 424B4, plus ~₩100tn at Cheongju and ~₩400tn for a new Southwestern cluster (disclosed 29 Jun 2026) H1 report. 2026 capex is guided to the high-₩40tn range from ₩8.3tn in 2023 Q2 call; 424B4 — while its own 2026 demand forecast is DRAM bits +mid-20%, NAND +high-teens.
What would confirm failure
Capex held or raised after the first negative ASP quarter; Yongin phases 2–6 (₩21.6tn approved Feb 2026) equipped without LTA-backed volume.
Damage to economics
Q2 D&A was just ₩4.0tn Q2 call; depreciation multiplies as fabs come online, while ASPs structurally decline over time 424B4. The cost base rises just as pricing weakens, cutting margins and returns on capital.
Permanent or fixable
Permanent — Sunk capital cannot be un-spent; only halting phases early limits it.
B2HBM becomes a three-way price fightPermanent

Attacks A1 HBM leadership is a manufacturing moat, not a timing lead

How it could fail
Competitors close the yield/quality gap at HBM4 and customers dual-source; the Q2 call fielded a direct question on rivals' "rapid progress". Management conceded rising conventional DRAM prices "may also have some influence" on HBM pricing Q2 call — the premium is negotiated, not fixed.
What would confirm failure
2027 HBM pricing converges toward conventional DRAM, or share lost on a major accelerator platform.
Damage to economics
The one differentiated profit pool shrinks toward commodity returns. Worse, HBM uses more wafer per bit: capacity freed from lost HBM volume floods conventional DRAM at the same time — a double hit.
Permanent or fixable
Permanent — A lost qualification is lost for a product generation.
B3Price mean-reverts before volume takes overFixable

Attacks A2 AI demand absorbs supply through at least 2027

How it could fail
2026 growth is price: DRAM ASP rose mid-60% QoQ in Q1 on flat bits 424B4 and ~30% in Q2 Q2 call. Gartner has the memory market at US$216bn → US$633bn (+192.7%) in 2026 424B4 on ~20% more bits. PC and mobile already show "temporary sales adjustments" from shortage — demand destruction at high prices — and CSP spending is discretionary.
What would confirm failure
A quarter of negative DRAM ASP; bits growing while revenue does not; customer inventory build.
Damage to economics
Operating leverage reverses. Precedent: 2023, a ₩9.1tn net loss and capex cut >50% 424B4; Q1'23 call.
Permanent or fixable
Fixable — Cyclical alone — but combined with B1 it becomes permanent.
B4LTAs turn out to be collars, not floorsFixable

Attacks A3 Long-term agreements raise the trough

How it could fail
Price mechanisms are "designed to address price volatility" — that cuts both ways. Counterparties are concentrated and powerful: the largest customer was 23.9% of 2025 revenue; the top two were 14.8% and 12.4% in Q1 2026 424B4. When shortage ends, they hold the leverage.
What would confirm failure
Deposits stay immaterial — contract liabilities were ₩0.5tn at 30 Jun 2026 against ₩79.3tn quarterly revenue H1 report — or LTA prices track spot down.
Damage to economics
The trough-raising argument fails and earnings keep full cyclical beta; the rerating case disappears, though the business survives.
Permanent or fixable
Fixable — Hurts the multiple, not the franchise.
Also on the list: Headline profit flatters: Q2 net profit of ₩93.9tn included ₩63.3tn of investment gains, largely the Kioxia stake sold in June 2026 Q2 call; 424B4. China fabs (Wuxi, Dalian) run on an annual US licence, granted for 2026 only 424B4.

4. Signals to monitor

Each card links to the assumption (green) or risk (red) it tracks.

DRAM ASP change, QoQ
~+30% (Q2 2026)
Direction that mattersFirst negative print while capex holds
Q2 callB3B1
DRAM bit shipments, QoQ
High-single-digit ↑ (Q2); ~+10% guided Q3
Direction that mattersBits overtake price as the growth driver
Q2 callA2
2027 HBM price vs conventional DRAM
Under negotiation
Direction that mattersPremium held (A1) vs convergence (B2)
Q2 callA1B2
Capex vs operating cash flow
2026 guide high-₩40tn; H1 capex ₩18.3tn vs OCF ₩91.7tn
Direction that mattersCapex ratio rising into falling ASPs
Q2 call; H1 reportA4B1
LTA coverage and deposits
~10 customers; coverage undisclosed; contract liabilities ₩0.5tn
Direction that mattersDisclosed coverage with material deposit build
Q2 call; H1 reportA3B4
Issued share count
730.5m; 24.07m to be cancelled by 19 Nov 2026
Direction that mattersCompletion, then a standing return framework
6-K 19 AugA4
Customer concentration
Largest 23.9% (2025); top two 14.8% / 12.4% (Q1 2026)
Direction that mattersRising share of top customer
424B4B4
Capacity phase approvals
Yongin ph. 2–6 ₩21.6tn approved; ₩600tn / ₩400tn plans
Direction that mattersApprovals not tied to LTA volume
424B4; H1 reportB1

5. External challenge notes

Not run. Write "external challenge" to add outside-source challenges here, kept separate from company-source evidence.

6. Bottom line

  1. SK hynix could work because it owns most of the one memory product AI cannot do without, and it is turning that scarcity into multi-year contracts and a ₩40tn share cancellation while holding ₩69.4tn of net cash.
  2. HBM4 and HBM4E must keep their share and price premium, and the LTAs must hold price when the shortage ends.
  3. The thesis breaks if the industry's peak-margin build-out — here, a plan to double wafer capacity in five years — lands in 2027–2030 just as ASPs, which drove nearly all of 2026's growth, roll over.
  4. A negative DRAM ASP quarter with capex unchanged, or 2027 HBM pricing converging toward conventional DRAM, would change my mind; disclosed LTA coverage backed by material deposits would strengthen it.

Sources

Prospectus 1
  • 424B4 424B4 prospectus, 10 Jul 2026 (financials to Q1 2026)
    sources/Other_Key_Documents/2026-07-10_424B4_Prospectus.html
Earnings calls 2
  • Q2 call Q2 2026 earnings call, 29 Jul 2026 (Seoul)
    sources/Transcripts/FY2026-Q2_earnings_call_transcript.txt
  • Q1'23 call Q1 2023 earnings call, 26 Apr 2023
    sources/Transcripts/FY2023-Q1_earnings_call_transcript.txt
6-K filings 3
  • 6-K 29 Jul Preliminary Q2 2026 results
    sources/Material_Events/2026-07-29_6-K.html
  • H1 report Semi-annual business report, H1 2026 (6-K 18 Aug 2026)
    sources/Material_Events/2026-08-18_6-K.html
  • 6-K 19 Aug Treasury share cancellation, ₩40.0tn
    sources/Material_Events/2026-08-19_6-K.html