Business Overview

SK hynix Inc. (Nasdaq: SKHY; KRX: 000660)

6 September 2026

Evidence base: SK hynix Form 424B4 prospectus dated 10 July 2026 and Form F-1 dated 24 June 2026; eleven Forms 6-K furnished between 15 July and 4 September 2026, including the semi-annual business report for the six months ended 30 June 2026 (KPMG Samjong review report dated 14 August 2026). Industry structure, competitor economics and export-control status are drawn from independent sources named inline: Counterpoint Research, TrendForce, Micron’s own SEC filings, Samsung’s own results release, and export-control legal commentary.

This document is an explanation of how the business works and what drives its economics. It is not a valuation and not a recommendation.

1. Executive Snapshot

Sources: 424B4 prospectus (10 July 2026); semi-annual business report for H1 2026 (6-K, 18 August 2026); Gartner via company filings.

2. What the Company Does

Every computer has a bottleneck between the processor that does the arithmetic and the place the data sits while it waits. SK hynix sells the place the data sits. Its two products solve two versions of the same problem: DRAM holds data the processor needs in the next few nanoseconds and forgets it when the power goes off, and NAND flash holds data the machine needs to keep. In an AI data centre both have become scarce at once, because a graphics processor that can perform a trillion operations a second is worthless if memory cannot feed it fast enough — the constraint the industry calls the memory wall.

The unit of economics is one bit. Memory is priced per bit, contracted per bit, and manufactured with a cost per bit that falls as circuit line-widths shrink and as more layers are stacked into one chip. Trace a single bit: it begins as part of a 300mm silicon wafer bought from suppliers in Japan, Korea, Germany and the United States — wafers are roughly 10% of cost of sales. It is patterned into a die in a fab at Icheon or Wuxi (DRAM) or Cheongju or Dalian (NAND), then sent to a back-end plant at Icheon, Cheongju or Chongqing to be tested, stacked and packaged into a module, an HBM stack or a solid-state drive. It is sold, almost always in US dollars, to a hyperscaler, a server maker or a handset manufacturer, at a volume and price agreed with that customer monthly or quarterly. There is no order backlog: the company states plainly that supply volumes and prices are set by mutual agreement on a monthly and quarterly basis and that it has none.

That last fact is the single most important structural feature of the business. A company with no backlog and no long-dated contracted price has no cushion between the spot market and its own income statement. When the price of a bit rises, revenue rises the same quarter; when it falls, it falls the same quarter. This is why the company earned an operating margin of 76.3% in the second quarter of 2026 and lost ₩9,138bn in 2023 on the same assets.

The revenue mix has narrowed towards DRAM. DRAM was 63.4% of revenue in 2023, 67.6% in 2024, 77.1% in 2025 and 77.3% in the first quarter of 2026; NAND fell from 29.5% to 22.0% over the same period, and everything else — foundry through the wholly-owned SK hynix system ic and SK keyfoundry subsidiaries, lease income and small domestic subsidiaries — collapsed from 7.2% to 0.7%. Part of that is arithmetic: DRAM prices rose faster. Part is deliberate. In March 2025 the company decided to transition its CMOS image sensor division into the AI memory business, and revenue from other products fell 29.0% to ₩1,552bn in 2025 as CIS sales declined. The company has been shedding the parts of its portfolio that do not earn AI-cycle economics and pointing the engineers at the parts that do.

Reporting follows that concentration. Because semiconductor revenue exceeds 90% of the total, the company presents a single segment, which means investors cannot see DRAM profitability separately from NAND, or HBM separately from commodity DRAM. That is the most consequential disclosure gap in the business and is picked up in Section 7.

3. Industry, Competitive Position and Moat

Memory exists because processing and storage are separate problems with separate economics, and because no computing device can be built without both. What is actually sold is a commodity with a specification: a bit of DDR5 from one supplier is functionally interchangeable with a bit of DDR5 from another once it has passed the customer’s qualification. That interchangeability is why the industry has no pricing power at the product level and enormous pricing power at the industry level: when total supply is short, every supplier’s price rises together, and when it is long, every supplier’s price falls together.

The profit pool sits with the integrated device manufacturers who own both the process technology and the fabs. Three of them — Samsung, SK hynix and Micron — hold roughly 90% of DRAM revenue between them, with CXMT of China growing from a small base at 716% year-on-year in the second quarter of 2026 (Counterpoint Research). NAND is more fragmented: Samsung 28%, SK hynix 19%, Micron 15%, Kioxia 14%, YMTC 14% and SanDisk 11% in the second quarter of 2026 (Counterpoint). The relevant market is global, not regional, because bits ship anywhere at negligible freight cost — so market share figures mean what they appear to mean.

The barrier that binds is not capital, which several state-backed entrants have. It is the pairing of process yield with customer qualification. A new HBM supplier must persuade an accelerator designer to requalify an entire memory subsystem, and the company’s own account describes deep customisation and integration as creating substantial switching costs. Outside evidence partially supports this and partially contradicts it. Samsung took four quarters to move from 15% to 33% of HBM revenue while SK hynix fell from 64% to 50% (Counterpoint). Qualification is real — it took four quarters, not one — but it gates a programme, not a customer relationship, and it has now been passed.

Source: Counterpoint Research, Global DRAM and HBM Market Share (quarterly). SK hynix’s own filings cite IDC at 56.4% HBM share for Q1 2026, consistent in level with the series above.

The same test applied to conventional DRAM is less flattering. The company reports IDC data showing its DRAM share at 29.9% in 2023, 33.4% in 2024, 34.8% in 2025 and 29.1% in the first quarter of 2026. Counterpoint’s independent series puts it at 26% in the second quarter of 2026, against Samsung at 39% and Micron at 25%, down from 39% a year earlier. The two providers use different bases and are not directly comparable in level, but they agree on direction: SK hynix has been losing relative share through the upcycle while its own revenue set records. The mechanism is straightforward and is disclosed in the company’s own numbers — its fabs have run at 100% utilisation throughout the first half of 2026, so it has had no incremental bits to sell into a market growing 57% quarter-on-quarter.

The most important reconciliation is on margin. SK hynix earned a 76.3% operating margin in the second quarter of 2026. In the same quarter Samsung’s Device Solutions division earned 70.0% (₩89.2tn on ₩127.5tn of revenue), and Micron reported an 80.4% GAAP operating margin on US$41,456m of revenue in its quarter ended 28 May 2026, guiding to an approximately 86% gross margin in the next. Three companies with different technology positions, different product mixes and different capital bases are all earning between 70% and 85%. That is not evidence of anyone’s moat. It is the price of a bit, arriving at three income statements at once.

What kind of company wins in this industry over a full cycle is therefore not the one with the highest margin at the peak. It is the one that enters the downturn with the lowest cost per bit, a balance sheet that can fund capex through a loss-making year, and a product mix weighted towards the segment that stays scarce longest. SK hynix is credibly two of those three: it is in net cash with ₩88.0tn of cash and equivalents against ₩18.6tn of borrowings at 30 June 2026, and its mix is weighted towards HBM and server DRAM, the segments TrendForce expects to stay short into 2027. Whether it has the lowest cost per bit is not disclosed by anyone and cannot be established from these sources.

4. Growth Engine

Revenue grew 256.8% year-on-year in the second quarter of 2026, to ₩79,318,746m. Almost none of that is volume, and none of it is acquired.

Take acquisitions first, because they are quickly dismissed. The Intel NAND business has been consolidated since December 2021; the final US$2.2bn instalment paid in March 2025 was a payment, not a new consolidation. SK keyfoundry was bought for ₩576bn in August 2022. Neither changed the reporting perimeter in 2025 or 2026. Reported growth and organic growth are the same number, which is unusual and worth stating plainly: nothing in the headline rate is bought.

What the headline rate is made of is price. The company discloses bit volume and average selling price only as quarter-on-quarter percentage bands, and those bands tell the story unambiguously.

Sources: 424B4 prospectus (quarters to Q1 2026); semi-annual business report, 6-K of 18 August 2026 (Q2 2026). The company discloses bands, not absolute figures.

Over the six quarters from the first quarter of 2025 to the second quarter of 2026, DRAM bit volume rose by something in the order of a third — one decline, one large increase, then three small ones and a flat quarter. Over the same six quarters DRAM average selling price rose from flat to low-single-digit to mid-single-digit to mid-20% to mid-60% to mid-30%, compounding to a multiple. Revenue rose more than fourfold. The volume contribution is real but secondary; price is the engine, and the reason price is the engine is that the company has no spare capacity with which to chase volume.

Ranked drivers of revenue growth

• DRAM average selling price (cyclical). The dominant driver by a wide margin. Prices are set by industry supply and demand; server DRAM contract prices were forecast to rise a further 13–18% quarter-on-quarter in the third quarter of 2026 (TrendForce, 9 July 2026). This lever reverses as readily as it rises.
• Mix shift into HBM and server DRAM (structural). HBM stacks many DRAM dies into one package and sells at a large premium per bit. The company held 50–58% of HBM revenue through the period and cites a 56.4% IDC share for the first quarter of 2026. Mix is what makes SK hynix’s ASP move differently from the commodity index.
• NAND price and the eSSD mix (cyclical, with a structural core). NAND average selling price rose in the mid-50% range in the second quarter of 2026 with bit volume up in the mid-10% range, helped by a weak first-quarter base and expanded enterprise SSD sales. The structural part — AI inference offloading key-value cache from memory to SSD — is real; the price part is not.
• Capacity additions (management-driven). The M15X extension fab in Cheongju opened its cleanroom in October 2025 and began wafer input in the first quarter of 2026. This is the only lever management directly controls, and it takes years: the Yongin phase 1 cleanroom opens in the first quarter of 2027 and the Indiana packaging plant in the second half of 2028.
• Won weakness (temporary). The company reports in Won and sells in dollars. The average rate moved from ₩1,363.4 per US$1.00 in 2024 to ₩1,421.4 in 2025, and the company names this as a contributor to the 2025 increase in DRAM and NAND revenue in Won terms.

The geography of demand has been re-pointed by the same forces. Measured by the location of the selling entity, the United States was 47.0% of revenue in 2023 and 64.7% in the first quarter of 2026, while China fell from 30.9% to 24.3%. That is the AI data centre build-out expressed as an invoice address, and it carries a concentration cost: the two largest customers were 14.8% and 12.4% of first-quarter 2026 revenue, and the single largest was 23.9% of 2025 revenue.

5. Margin, Cash and Capital Allocation

The margin mechanics of a memory fab are the simplest in industrial manufacturing and the most violent. Almost the entire cost base is fixed within a quarter: depreciation of ₩13,890bn in 2025 on plant already built, the salaries of a workforce whose average tenure in the semiconductor segment is 12.1 years, and the electricity and purified water to keep clean rooms running. Raw materials are a minor line — total raw material and supply purchases were ₩10,970bn in the first half of 2026, of which wafers were ₩999bn, against ₩131,895bn of revenue. Because the marginal cost of an additional bit is small and the fixed cost of the fab is enormous, every Won of price increase falls almost intact to operating profit.

The 2025 accounts show this precisely. Revenue rose 46.8% to ₩97,147bn while cost of sales rose 11.9% to ₩38,456bn, and gross margin went from 48.1% to 60.4%. In the first quarter of 2026 revenue rose 198.1% year-on-year while cost of sales rose 44.6%, and gross margin reached 79.3%. The operating leverage works with equal force downwards: in 2023, with the same fabs and the same people, gross margin was minus 1.6% and the operating loss was ₩7,730bn.

One cost behaves unusually and deserves naming. Salaries inside cost of sales rise sharply in good years because of a profit-sharing incentive programme, which the company cites as a main reason for the increase in cost of sales in both 2025 and the first quarter of 2026. It is a genuine variable cost that self-corrects in a downturn, and it means the reported margin at the peak is already net of a large bonus accrual.

Operating profit for FY2023–FY2025 is computed as gross profit less selling and administrative expenses less research and development expenses, the line items disclosed in the 424B4 income statement, which does not present an operating profit subtotal. The method reproduces the company’s own reported Q1 2026 operating profit of ₩37,610bn exactly. FY2023–25 figures are IFRS as issued by the IASB (424B4); H1 2026 figures are K-IFRS (6-K of 18 August 2026) — the company states it has made no attempt to quantify the differences, so the last column is not strictly comparable with the first three.

Why profit for the period is not the earnings power

The final row of that table needs its own explanation, because taken at face value it is misleading. Profit for the first half of 2026 was ₩134,269bn against operating profit of ₩98,153bn. The ₩36tn gap, and more, is net finance income of ₩76,251bn, of which ₩63,179bn is a gain on the valuation of financial instruments and ₩13,980bn is dividend income. The Level 3 fair value note isolates it: long-term investment assets opened the year at ₩14,547bn, took a valuation gain of ₩62,944bn, disposed of ₩10,656bn, and had ₩66,077bn transferred out to Level 2 as the valuation methodology changed. Long-term investment assets on the balance sheet stood at ₩85,310bn at 30 June 2026 against ₩14,547bn six months earlier.

These are the company’s Kioxia-linked holdings (inferred, from the amounts and the disclosed portfolio). SK hynix invested ₩2,637bn for a limited partnership interest in a Bain-led vehicle in 2018 and ₩1,279bn in a convertible bond of a second vehicle convertible into roughly 15.0% of it; carrying values at 31 March 2026 were ₩6,616bn and ₩13,609bn, and the first vehicle sold its remaining Kioxia equity in June 2026. Kioxia shares rose roughly 2,220% in the year to May 2026, at which point Korean press valued the SK hynix interest at above ₩60tn against a carrying value of about ₩14tn at the end of 2025. The gain is real, it is very large, and it is non-cash, non-repeating, and unrelated to selling memory. Operating profit of ₩98,153bn is the number that describes the memory business; a reversal in the Kioxia holding would run back through profit or loss.

A second, smaller distortion runs the other way. Finance expenses include a ₩3,979bn loss on derivatives from exchangeable bonds issued in April 2023, arising because the company’s own share price rose and the exchange rights were exercised. The company states there is no cash outflow and that the loss is offset by gains on disposal of treasury shares recognised in equity, so the net effect on total equity is minimal.

Cash and the balance sheet

Cash conversion is strong but not one-for-one, because price increases inflate receivables before they are collected. Operating cash flow of ₩91,743bn in the first half of 2026 includes ₩13,984bn of dividends received, while trade receivables rose from ₩18,199bn to ₩47,821bn and inventories from ₩14,289bn to ₩17,986bn. The balance sheet has been transformed: total equity of ₩262,693bn at 30 June 2026 against ₩120,667bn six months earlier, borrowings down to ₩18,587bn from ₩22,248bn, and cash and equivalents of ₩87,958bn — a net cash position, with the liabilities-to-equity ratio at 32.8% against 87.5% at the end of 2023.

Where the cash has gone, ranked

• Capital expenditure — ₩70,119bn from 2023 to June 2026 (₩8,325bn, ₩15,946bn, ₩27,519bn and ₩18,329bn), with 2026 spending planned to rise considerably above 2025. Committed projects as at 31 March 2026 were Yongin fab 1 phase 1 at ₩9,412bn and phases 2 to 6 at ₩21,608bn (construction only), the P&T7 advanced packaging plant in Cheongju at ₩19,000bn including equipment, and an Indiana packaging plant at ₩5,900bn, the last supported by up to US$458m of CHIPS Act subsidies and US$570m of loans.
• Share cancellation — an estimated ₩40,004bn, announced 19 August 2026. The board resolved to buy treasury shares in the open market between 20 August and 19 November 2026 and cancel them, sized at 24,070,000 shares against 730,492,365 issued using the prior day’s close of ₩1,662,000. This post-dates every reported figure above and, at roughly a third of first-half operating profit, is the largest single change in capital allocation in the period covered by these documents.
• **Acquisitions — US$8.8bn for the Intel NAND business** (US$6.6bn in December 2021 and US$2.2bn in March 2025) and ₩576bn for SK keyfoundry in August 2022.
• Dividends — a fixed ₩1,500 per share per year under the policy set in November 2024 for 2025 to 2027, paid in four ₩375 instalments; the second-quarter 2026 dividend was ₩273bn. The company said on 7 August 2026 that it is reviewing additional shareholder returns and expects to announce them within the third quarter.
• Debt reduction — borrowings from ₩29,469bn at end-2023 to ₩18,587bn at 30 June 2026, alongside ₩26.2bn of net proceeds raised in the July 2026 Nasdaq listing of 17,790,000 new shares, equal to 2.50% of the share count, at US$149.00 per American Depositary Share.

The ranking says something specific about how this management thinks. A fixed dividend of ₩1,500 per share was set in November 2024 and has not been raised despite a fivefold increase in profitability, while the variable return has come through a one-off cancellation announced only once the cash had arrived. Capital expenditure has been allowed to triple. The order of priorities is capacity first, balance sheet second, variable returns third, and fixed commitments last — which is the allocation of a company that expects to need cash in a downturn.

6. Cyclicality, Constraints and What to Monitor

This business sits at or very near the top of its cycle, and the evidence is unambiguous on all three of the axes that matter. On price, DRAM average selling price has risen quarter-on-quarter for six consecutive quarters. On volume, the fabs ran at 100% utilisation for the whole of the first half of 2026 — available working hours of 136,255,098 exactly equalled actual working hours — which is the definition of a ceiling. On margin, the second-quarter operating margin of 76.3% is an all-time record, set against a trough three years earlier at minus 23.6%. A record margin at a peak and a record margin at a trough are opposite facts, and this is the first kind.

The binding constraint follows directly from the utilisation figure: there is no volume lever left. Any further revenue growth in 2026 must come from price or mix, because the next meaningful capacity does not arrive until the Yongin phase 1 cleanroom in the first quarter of 2027, and clean rooms are not production — equipment installation and qualification follow. TrendForce expects new industry capacity to contribute materially only from the second half of 2027 and substantially from 2028, partly because HBM consumes far more wafer input per bit than conventional DRAM, so wafer starts translate into fewer sellable bits than they used to.

That same supply arithmetic is why the downside case is asymmetric between the two products. TrendForce puts the DRAM sufficiency ratio at minus 1% to minus 2% for 2026 with the gap widening in 2027, so the DRAM squeeze is expected to persist. NAND is the opposite: accelerated migration to higher-layer products and new fabs ramping are expected to turn NAND sufficiency positive in 2027, shifting to a balanced market while consumer demand stays subdued. NAND was 22.0% of first-quarter 2026 revenue and its average selling price rose in the mid-50% range in the second quarter. If the NAND ASP band flips negative before DRAM does, roughly a fifth of revenue starts deflating while the other four fifths are still inflating — and given the fixed cost base, a fifth of revenue is a much larger fraction of profit.

A second mechanism caps the upside on price even while supply stays short. Several large US cloud providers hold multi-year agreements that restrict suppliers from raising prices to them, so increases fall disproportionately on customers without such agreements (TrendForce, 9 July 2026). Counterpoint attributes part of Samsung’s second-quarter strength to rivals being locked into earlier agreements struck at lower negotiated levels. With the United States at 64.7% of revenue by selling-entity location, SK hynix is more exposed to that dynamic than its share of the market alone would suggest.

Leading indicators, and where they are published

• The quarter-on-quarter ASP and bit-volume bands for DRAM and NAND — SK hynix quarterly and semi-annual business reports, furnished on Form 6-K. This is the single most informative disclosure the company makes.
• HBM revenue share against Samsung and Micron — Counterpoint Research quarterly market share releases; IDC via the company’s own filings.
• Server DRAM and eSSD contract prices, and the DRAM and NAND sufficiency ratios — TrendForce press releases.
• Micron’s quarterly results, which report a fiscal quarter ending roughly a month ahead of SK hynix’s and read on the same cycle — SEC filings and micron.com investor relations.
• Samsung Device Solutions divisional revenue and operating profit — Samsung quarterly results releases.
• The annual US export licence covering foreign-owned fabs in China — Bureau of Industry and Security actions and company disclosure.

7. Risks, Unknowns and Questions for Deeper Work

The risks below are those that cyclicality alone does not capture. Cycle position, the utilisation ceiling and the NAND-versus-DRAM supply split are covered in Section 6 and are not repeated here.

• China manufacturing runs on a licence that is renewed annually at another government’s discretion. The validated end-user authorisations that let foreign-owned fabs in China receive US-controlled equipment without per-shipment licences expired on 31 December 2025 and were replaced by annual export licences. Samsung and SK hynix received 2026 licences, but the framework is described as supporting operation and maintenance rather than expansion, and each renewal is a fresh decision. The Wuxi DRAM fab and the Dalian NAND fab sit inside that framework. A refusal or a narrowing of scope would not shut them immediately; it would freeze their technology node while competitors migrate, which converts a producing asset into a depreciating one.
• A regulatory condition caps NAND pricing in a quarter of the revenue base. China’s market regulator conditioned its approval of the Intel NAND acquisition on maintaining a reasonable pricing policy and production level in the Chinese enterprise SSD market for five years from December 2021. The company states in its own prospectus that it expects this to limit its ability to significantly raise NAND prices in China in 2026. China was 24.3% of first-quarter 2026 revenue. The condition expires in December 2026, and any waiver beyond that is at the regulator’s discretion.
• Customer concentration converts a commercial negotiation into a political one. The largest customer was 23.9% of 2025 revenue and the two largest were 14.8% and 12.4% of first-quarter 2026 revenue. The company’s own risk disclosure goes further than concentration: it states that having to allocate finite supply among customers may damage those relationships and attract heightened political, legal and regulatory scrutiny. A supplier who must choose which hyperscaler gets the bits is exposed in a way a supplier with spare capacity is not.
• Depreciation from ₩70tn of capital spending arrives after the price peak, not with it. Depreciation begins when a plant is completed, and the largest tranche of recent spending — Yongin, P&T7, Indiana — is still pre-production. KPMG identified the appropriateness of the commencement timing of depreciation for machinery and equipment as a critical audit matter for the 2025 financial statements, which is a direct signal that the timing involves judgement and is material. If average selling prices normalise as this depreciation lands, margin falls faster than price, because the denominator of cost per bit has grown while the numerator of revenue per bit is shrinking. This risk compounds with the utilisation constraint: the capacity that creates the depreciation is the same capacity that arrives too late to sell into the current price.
• Three announced but undetermined transactions could each move the perimeter. The company has confirmed to the Korea Exchange that it is reviewing, without having decided, a pre-IPO capital raising of approximately ₩5tn at its Solidigm subsidiary, a sale of an ownership stake in its Chongqing packaging plant reported at approximately ₩4tn, and the establishment of a semiconductor facility in Japan. In each case the company stated that no matters have been determined. All three post-date the reported financial statements.

What these sources could not answer

• Absolute bit shipments and dollars per gigabyte. The company discloses only quarter-on-quarter percentage bands, so an outsider cannot build a unit-economics model, cannot separate mix from price within the ASP move, and cannot compare cost per bit with Samsung or Micron.
• Profitability by product. With a single reported segment, DRAM margin cannot be separated from NAND margin, and HBM revenue and margin are not disclosed at all. The most important product in the business is invisible in the financial statements.
• The split of capital expenditure between maintenance and growth, and any Won-denominated capital expenditure figure for 2026 — the guidance is that spending will rise considerably above 2025.
• Second-quarter 2026 gross margin and cost of sales. Korean disclosure rules permit the management discussion to be omitted from semi-annual reports, and the company omitted it, so the most recent quarter has a revenue and operating profit line but no cost detail.
• Earnings call transcripts and investor day materials, which are not filed on EDGAR. Management’s own framing of long-term targets, and the tone of analyst pushback in question and answer, are absent from this evidence base entirely.
• The identity of the two largest customers, and the terms of any long-term agreements governing price with them.

Before forming a view, the questions that would need resolving are: what SK hynix’s cost per bit is relative to Samsung’s and Micron’s, since that is what determines who survives the next trough intact; what proportion of revenue is covered by price-capped long-term agreements; how much of the HBM share loss to Samsung is programme timing versus a durable second-source decision by the accelerator makers; and what the depreciation schedule looks like as Yongin, P&T7 and Indiana come into service.

8. Investor Takeaways

• What this business really is: a two-product bit factory — DRAM 77%, NAND 22% — with no backlog and no contracted price, whose earnings are made by an industry price it does not set.
• The core economic engine: fixed-cost fabs at 100% utilisation. In 2025 cost of sales rose 11.9% while revenue rose 46.8%; that gap is the whole business, and it runs in both directions.
• The main growth lever: DRAM average selling price, and the HBM and server mix that lifts it above the commodity index. Volume has been flat to slightly up for four quarters.
• What could break the story: HBM ceasing to be scarce — Samsung moved from 15% to 33% of HBM revenue in four quarters — or NAND supply normalising in 2027, either of them landing at the same time as depreciation on ₩70tn of new capacity.
• What to monitor: the quarterly ASP and bit-volume bands in the company’s own 6-K filings, HBM share against Samsung, and the annual US export licence covering Wuxi and Dalian.