SK hynix Inc. SKHY / KRX:000660

1Business snapshot

SK hynix designs and manufactures DRAM and NAND flash memory. Revenue is a price × bits equation: the company converts wafers into bits and sells them into a global market where prices are set by industry supply and demand, not by the company. What has changed the character of that equation is HBM — stacked DRAM connected by TSV, sold to AI accelerator makers. HBM carried a price premium of more than five times conventional DRAM per gigabyte in 2025 Gartner, via 424B4, and SK hynix holds 56.4% of that market by revenue IDC, via 424B4.

The product mix in Q1 2026 was DRAM ₩40.7tn (77.3%), NAND ₩11.6tn (22.0%), other ₩0.3tn 424B4. Fabs are in Icheon and Cheongju (Korea) and Wuxi and Dalian (China), with back-end assembly and test in Icheon, Cheongju and Chongqing. Solidigm, the former Intel NAND business acquired for US$8.8bn total consideration, supplies high-capacity enterprise SSDs. Sales are heavily US-facing: 64.7% of Q1 2026 revenue booked through US sales subsidiaries, 24.3% through China 424B4.

SK square holds 20.50% and is legally required to hold at least 20% under Korea's Monopoly Regulation and Fair Trade Act; the National Pension Service holds 8.06%. A Nasdaq ADS listing closed 14 July 2026, raising ~US$26.2bn net — the largest US IPO by a foreign company 424B4; Q2 2026 call.

2Core debate

Is the current AI memory boom a structural shift in the industry's economics, or the biggest cyclical peak in its history?

The structural case rests on a tiered AI memory architecture — HBM beside the GPU, server DRAM beside the CPU, eSSD underneath — where each tier scales with the others, and on enterprise demand rising from 26.5% of the memory market in 2020 to a forecast 51.9% in 2027 Gartner, via 424B4. The cyclical case rests on the arithmetic of the last two quarters: DRAM ASP rose mid-60% QoQ in Q1 2026 and ~30% in Q2, NAND ASP mid-70% then mid-50%, while bit shipments were roughly flat to modestly up 424B4; Q2 2026 call. Nearly all the revenue growth is price. Price mean-reverts.

Everything in this file — margins, capex, the ₩40tn buyback, the valuation — resolves to which of these is right.

3Base thesis

SK hynix entered the AI cycle holding the one memory product that is genuinely hard to make at scale, and it has used the resulting cash to buy optionality rather than leverage. HBM leadership is not a spec advantage but a manufacturing one — yield, quality and the ability to deliver volume on a customer's schedule — and management argues this is "accumulated competitiveness… that cannot be replicated in a short period" Q2 2026 call. HBM4 entered mass production in Q2 2026 with yields near mature HBM3E levels, and HBM4E samples have shipped for 2027 volume production. Around the franchise, the company has begun converting spot exposure into contract exposure: long-term agreements concluded with roughly ten customers, terms around five years, incorporating volume commitments, customer deposits and price mechanisms designed to dampen volatility Q2 2026 call. The balance sheet — ₩69.4tn net cash, 7% debt-to-equity — funds ₩50tn of announced Korean capacity projects plus ₩11.9tn of EUV scanners without dilution or leverage 424B4; Q2 2026 call.

Key assumptions

1HBM revenue share holds above roughly 45% through the HBM4/HBM4E generations — i.e. the moat is manufacturing, not timing.

2LTA coverage meaningfully raises the trough, so the next downturn is a margin compression rather than a loss year like 2023.

3Hyperscaler AI capex grows or holds through 2027–28; efficiency gains broaden adoption rather than reduce memory content.

4Industry capacity arrives in phases tied to confirmed demand, not as a synchronised 2027–2030 wave.

5Annual BIS licensing for the China fabs continues to be granted.

4Main bear case

This is the cycle, not a new regime — and the peak is visible in the numbers. Operating margin of 76% in a commodity manufacturing business is not a durable state; it is the signature of a shortage. Management's own demand forecast is DRAM bits +mid-20% and NAND +high-teens Q2 2026 call, while Gartner has the memory market going from US$216bn in 2025 to US$633bn in 2026 424B4 — a 193% revenue increase on ~20% more bits. The gap is price, and price is what the entire industry is now spending to capture: SK hynix capex from ₩8.3tn (2023) to ₩27.5tn (2025) to a high-₩40tn range in 2026 424B4; Q2 2026 call, with Samsung, Micron and CXMT investing against the same signal. Capacity ordered in 2026 lands in 2027–2030 into whatever demand exists then.

Second-order problems compound it. A profit-sharing programme pays employees 10% of operating profit, so the margin is partly self-taxing 424B4. A putative antitrust class action filed 25 June 2026 alleges the industry restricted conventional DRAM supply by shifting capacity to HBM 424B4 — an allegation whose factual predicate is exactly the mix shift the bull case celebrates. And headline profit flatters the operating story: Q2 2026 net profit of ₩93.9tn included ₩63.3tn of gains on the sale and revaluation of investments Q2 2026 call, principally the exited Kioxia stake, which does not recur.

How it breaks: supply catches demand — through M15X, Yongin, and competitors' additions — ASPs fall faster than unit costs, LTA price mechanisms prove to be collars rather than floors, and depreciation from a ₩40tn+ capex year lands on a falling revenue line. The precedent is recent: in 2023 this company reported a ₩9.1tn net loss and cut capex by more than 50% 424B4; Q1 2023 call.

5Business quality

High, but conditionally so. The industry structure is genuinely good: three players hold more than 90% of DRAM revenue, five hold more than 90% of NAND, and entry requires capability plus capital that only a handful of firms possess — global semiconductor capital spending in 2026 is projected at US$237bn, with the top 20 spenders accounting for 87.6% 424B4. Within that structure, HBM is the most defensible position SK hynix has ever held, because the customer's switching cost is a qualification cycle and a system-level failure risk, not a price comparison.

But the majority of bits remain price-taking, and management is explicit that memory ASPs decline over time irrespective of the cycle, so profitability depends on cutting per-bit cost faster than per-bit price falls 424B4. Quality here means a better-than-peer position inside a structurally cyclical business — not escape from it. Whether LTAs change that is the central unproven claim (§9).

6Key value drivers

1HBM share × price × volume. The single largest swing factor. 2027 HBM pricing is under negotiation; management notes rising conventional DRAM prices "may also have some influence" on HBM pricing discussions Q2 2026 call.

2Conventional DRAM ASP. The largest revenue block and the most cyclical.

3eSSD / Solidigm. Q2 eSSD revenue doubled QoQ; Solidigm 30TB+ drives more than tripled Q2 2026 call. This is the leverage on AI inference and KV-cache offload.

4Capex efficiency and timing. ₩50tn of Korean projects plus Indiana; returns depend on landing capacity into demand rather than into a glut.

5Capital returns. Newly material after August 2026 (§9).

7Financial profile

Revenue: ₩32.8tn (2023) → ₩66.2tn (2024) → ₩97.1tn (2025) → ₩131.9tn in H1 2026 alone 424B4; 6-K 29 Jul 2026. Profit swung from a ₩9.1tn loss in 2023 to ₩42.9tn in 2025 to ₩134.3tn in H1 2026. Operating margin ran 48.6% in 2025, 71.5% in Q1 2026 and 76% in Q2 2026 — the last an all-time high 424B4; Q2 2026 call. Q2 EBITDA was ₩64.6tn on an 81% margin.

The balance sheet is the strongest part of the story. Cash and short-term investments of ₩88tn against ₩18.6tn interest-bearing debt gave net cash of ₩69.4tn at end-Q2 2026, with debt-to-equity at 7% Q2 2026 call. Operating cash flow was ₩53.4tn in 2025 against ₩27.5tn capex. R&D runs ~₩2.5tn per quarter 424B4.

Two quality caveats. First, a large share of recent pre-tax profit is non-operating investment gains (₩63.3tn in Q2 alone). Second, an exchangeable-bond derivative produced a ₩4.0tn accounting loss in H1 2026 with no cash outflow, offset against treasury share disposal gains 6-K 14 Aug 2026.

8Valuation

At the Q2 2026 run rate the business earns ~₩242tn of annualised operating profit against ₩50tn of multi-year committed capex. The relevant question is not the multiple on peak earnings but what mid-cycle earnings look like — and sources contain no mid-cycle estimate, so any answer is judgment, not fact.

9Management & capital allocation

Nohjung Kwak has been CEO and representative director since March 2022, through both the 2023 loss year and the AI upcycle. The board has nine members with an independent majority and an independent chair 424B4.

The capital allocation record is better than the industry norm. The Intel NAND acquisition (US$8.8bn across 2021 and 2025) bought the eSSD franchise that is now doubling revenue quarterly. The 2018 Kioxia investment — ₩3.9tn deployed, carried at ₩20.2tn as of March 2026 — was fully exited in June 2026 424B4, a well-timed monetisation into the strongest market in the company's history. Capex was cut by more than 50% in 2023 when conditions demanded it Q1 2023 call, which is the discipline that matters most in this industry.

Shareholder returns were, until recently, immaterial: the FY2025–27 policy sets ₩1,500 per share annually in ₩375 quarterly instalments, and even with a topped-up FY2025 annual dividend of ₩1,875 the full-year payout was ₩2.1tn against ₩42.9tn of profit — under 5% 424B4. That changed on 19 August 2026, when the board approved acquiring and cancelling 24.07m shares for approximately ₩40.0tn via open-market purchase between 20 August and 19 November 2026 6-K 19 Aug 2026 — a genuine shift in policy scale, and the first time capital returns have been sized to the cash generation.

Claims to verify over time. Three management assertions carry thesis weight and are not yet demonstrable: (i) that LTA deposits and price mechanisms will actually protect earnings in a downturn; (ii) that capacity will be added "in stages… considering customer demand visibility" rather than on schedule regardless; (iii) that HBM manufacturing leadership is non-replicable. Evidence for each is tracked in §11–12.

10Risks ranked by damage

1. ASP mean reversion / cycle turn

Mechanism: prices, not bits, produced the earnings; when supply catches demand, ASPs fall faster than per-bit costs and operating leverage runs in reverse. Evidence today: DRAM ASP +mid-60% then +30% QoQ on flat-to-modest bit growth 424B4; Q2 2026 call; management's own bit demand forecast is mid-20%/high-teens. What would confirm it: a quarter of negative ASP change, or bit shipments growing while revenue does not. Impact: 2023 showed the full swing — a ₩9.1tn net loss.

2. Peak-cycle capex creating structural overcapacity

Mechanism: ₩50tn of Korean projects plus Indiana and M17, decided at 76% margins, deliver into 2027–2030 alongside competitors' additions; depreciation is permanent, prices are not. Evidence: 2026 capex guided to the high-₩40tn range from ₩8.3tn in 2023 Q2 2026 call; 424B4. What would confirm it: capex maintained after ASPs roll over. Impact: fixed-cost base sized for a market that does not arrive.

3. AI infrastructure demand air-pocket

Mechanism: hyperscaler capex is discretionary and concentrated; a digestion phase, an ROI reassessment, or architectural efficiency gains that cut memory content per node would hit HBM and server DRAM simultaneously. The prospectus names this explicitly, including customer inventory accumulation ahead of true end demand 424B4. What would confirm it: order deferrals or LTA renegotiation requests. Impact: the demand side of the thesis inverts.

4. Competitive erosion in HBM

Mechanism: Samsung and Micron close the yield/quality gap at HBM4, turning a 56% share franchise into a three-way price fight. Evidence today: none adverse — HBM4 in mass production with near-HBM3E yields Q2 2026 call — but the analyst question was asked directly on the Q2 call. What would confirm it: share loss on a major accelerator platform, or HBM pricing decoupling downward from its DRAM premium. Impact: permanent loss of the one differentiated position.

5. China operations and export controls

Mechanism: the Wuxi (DRAM) and Dalian (NAND) fabs need US-origin equipment. VEU status was revoked effective 31 December 2025 and replaced by an annual approval mechanism; a 2026 licence was granted in December 2025 424B4. What would confirm it: a delayed or narrowed 2027 licence. Impact: a material share of installed capacity becomes un-upgradeable.

6. Customer concentration

Mechanism: the largest customer was 23.9% of 2025 revenue; the two largest were 14.8% and 12.4% in Q1 2026 424B4. Pricing and volume are agreed transaction by transaction. What would confirm it: an LTA lapsing, or a customer qualifying a second HBM source at scale.

7. Antitrust and allocation scrutiny

Mechanism: the June 2026 US class action alleges coordinated supply restriction via the shift to HBM; the prospectus separately warns that allocation decisions invite political and regulatory attention 424B4. Impact: damages, and constraints on pricing or allocation freedom. Note the Chinese SAMR condition from the Intel deal already limits NAND price increases in China through December 2026.

8. Earnings quality / non-operating dependence

Mechanism: ₩63.3tn of Q2 profit was investment gains Q2 2026 call; the Kioxia position is now sold, so the comparison base is inflated and will not repeat. Impact: optical, but it distorts every trailing multiple.

9. Tariffs and trade policy — Medium · Low · Low. Semiconductors are currently exempt from the US Section 122 surcharge, but a 100% semiconductor tariff has been threatened twice 424B4.

10. IP litigation — Low–Medium · Low · Medium. Two MonolithIC ITC complaints (Feb and May 2026) seek exclusion orders; first target date 30 August 2027 424B4.

11Metrics to monitor

Next quarter's guided markers: DRAM bits +~10% QoQ, NAND bits +low-single-digit % Q2 2026 call.

12Thesis tripwires

13Open questions

1What share of volume and revenue do the ~10 concluded LTAs actually cover, and what is the price floor mechanism? Management declined to quantify Q2 2026 call.

2What is HBM as a percentage of DRAM revenue and of operating profit? Not disclosed in any source here.

3What does mid-cycle margin look like after ₩40tn+ annual capex converts into depreciation?

4Is the ADS/KRX price discrepancy (§8) a data error or a real dislocation?

5What are the Solidigm pre-IPO raise, Chongqing packaging stake sale, and possible Japan facility — all confirmed as "under review, nothing determined" 6-Ks 6/10/21 Aug 2026? Each would change the capital structure.

6Coverage gap: no annual report, 20-F, or audited quarterly report is in sources/. The 424B4 prospectus is the only comprehensive financial document, and it stops at Q1 2026.

14Update log

15Source map