SK hynix

IThe business

SK hynix makes memory. DRAM is the fast memory a processor works from. NAND is the flash that keeps data when the power is off. The company designs both and builds them in its own fabs at Icheon and Cheongju in Korea and at Wuxi and Dalian in China, then packages them in Korea and at Chongqing.⁴ It also owns two small foundry businesses. Almost everything else on this page follows from owning those factories: when prices rise, nearly all the extra revenue is profit, and when prices fall the factories still have to be paid for.

The company’s history explains its habits. It began as Hyundai Electronics in 1983, went through a debt-for-equity swap with its creditors after the DRAM glut of the late 1990s and early 2000s, and took the SK name in 2012.⁴¹⁰ Today SK square owns 146.1M shares, about a fifth, and the law requires it to stay at 20% or more.⁴ The shares trade on the Korea Exchange as 000660. In July 2026 the company added American depositary shares on Nasdaq as SKHY, each worth one tenth of a common share, and raised ₩39.9T.⁴¹³ This page values the Korean share in won, and shows the ADS beside it, because the two prices have come far apart.

In 2025 revenue was ₩97.1T, up 47%. DRAM was ₩74.9T, NAND ₩20.7T and other products ₩1.6T.⁴ Then came the steepest rise in its history. The first quarter of 2026 was ₩52.6T, the second ₩79.3T and the third ₩104.6T. In the nine months operating profit was ₩181.5T, nearly four times all of 2025 C (calculated).²

Look at where that came from. In the third quarter DRAM prices rose in the high 20s percent and bits about 10%. NAND prices rose in the mid 20s percent and bits fell slightly.⁶ A bit is the unit of memory sold. Revenue rose by ₩25.3T from the second quarter, while all costs including staff bonuses rose by about ₩2.5T C (calculated).² So roughly nine won in every ten of extra revenue fell through to operating profit. This is the most important fact about the company today. It works in reverse too.

Why memory became scarce

HBM stacks DRAM dies on top of each other and sits next to an AI processor. It needs more wafers and more back-end work for each bit than ordinary DRAM, so every HBM bit takes capacity away from the rest of the market.⁴ AI servers also need ordinary server DRAM beside the CPU, and enterprise SSDs below that. The prospectus calls this a tiered memory architecture, and it means AI demand now reaches all three layers at once.⁴

New supply takes years. The first cleanroom at Yongin, a new site south of Seoul, broke ground in February 2025 and opens in the first quarter of 2027. That is the shell only. The company put the shell of its first Yongin fab at ₩31.0T, not counting equipment, finishing by the end of 2030.⁴ Meanwhile finer nodes add fewer bits each generation. Management expects achievable demand growth in 2027 to be the mid 20s percent for DRAM and the high teens for NAND, about the same as this year, and says customers want far more.⁶ For HBM it says requirements are more than double what it can supply today.⁶

Those are company claims. Samsung says much the same: it expects AI demand in 2027 to far exceed industry supply.⁹ Hearing this from every seller is reassuring, and it is also what one would hear near a peak. Consensus, at least, assumes one. Nasdaq’s estimates per ADS run $25.83 for 2026, $36.03 for 2027, $42.54 for 2028 and then $31.43 for 2029, though only two analysts go that far.¹²

What it sells, and to whom

In the third quarter server DDR5 shipments rose over 40% from the second, and HBM4, the newest generation, began shipping in volume. Management expects HBM4 to pass HBM3E in both bits and revenue in the fourth quarter, and it expects to run M15X, a new extension at Cheongju, at full capacity by year end. Fourth-quarter bits are guided up in the mid 20s percent for DRAM and mid single digits for NAND.⁶

NAND comes in two parts: the Korean and Chinese plants that were always SK hynix, and Solidigm, the American business built from Intel’s NAND and SSD arm. SK hynix paid $8.8B for it in two stages, in December 2021 and March 2025.⁴ IDC put the company second in NAND with 18.5% in the first quarter of 2026.⁴ More than half of Korean NAND capacity is due to run on 321-layer technology in the fourth quarter.³

Customers are few and large. The biggest was 23.9% of revenue in 2025. In the first quarter of 2026 the top two were 14.8% and 12.4%, so concentration eased as ordinary DRAM prices rose.⁴ The company doesn’t name them in that note. In June 2026 it did announce a technology partnership with NVIDIA that includes memory supply.⁴ By the location of the selling subsidiary, 64.7% of first-quarter revenue was the United States and 24.3% China.⁴

Who pays for all this

A simple check. WSTS’s spring forecast put 2026 memory sales above $800B in a chip market of about $1.51T.¹⁷ Gartner, quoted in the prospectus, had DRAM alone going from $143B in 2025 to $401B in 2027, and NAND from $68B to $341B.⁴ Sales like that pass through the budgets of a small number of cloud companies and, behind them, the buyers of phones and PCs. Management itself expects PC and phone shipments to fall because of memory costs.⁶ I can’t tell how long the world will pay this. I can say that a bill this size is now somebody’s first target for savings. On the call an analyst asked about customers lowering HBM stack heights, which is that process starting.⁶

IIThe moat

A durable business keeps others from taking its profits. For memory the record is the test, and the record is mixed. From 2016 to 2025 SK hynix’s operating margin averaged 29.7%, weighted by revenue, and swung from 51.5% in 2018 to −23.6% in 2023, when it lost ₩7.7T of operating profit on ₩32.8T of revenue C (calculated).⁸⁴ A firm with true pricing power doesn’t sell below cost, and in 2023 gross profit was negative.

What exists is a high wall around a small group. TrendForce put Samsung at 39.4% of DRAM revenue in the second quarter of 2026, SK hynix at 24.9% and Micron at 23.3%, or 87.6% between them C (calculated).⁷ Getting in takes years of process work and tens of billions of won. Three firms is why prices can stay high. It doesn’t stop all three from building at once, which is how downturns here begin.

Evidence it is widening

The long-term agreement is the new fact. SK hynix has signed 14 customers, including large cloud companies, typically for five years, for more than half of 2027 and 2028 capacity. Management says it plans to keep that share similar over time, and that with volumes still under discussion, DRAM and HBM through 2029 are effectively sold out.⁶ That last sentence includes talks not yet signed, so I read it as a statement of demand and not of contract.

The second fact is that HBM has made memory less interchangeable. A customer can’t swap a stack for a rival’s without qualifying it for the platform, and customers now want custom HBM, with some processor functions built into the base die.⁶⁴ That moves memory a little closer to a designed product. It has not yet been tested in a period when customers have more than enough.

Evidence it is narrowing

Compare these contracts with Micron’s. Micron describes take-or-pay agreements with floors and ceilings on most of the contracted revenue.¹⁸ SK hynix says its agreements reflect market conditions within an agreed price range, with deposits as safeguards.⁶ It has published neither the floor nor the deposit total, and contract liabilities on the balance sheet were only ₩0.5T in June.¹⁰ So I can’t say how much of the 50% would hold in a real fall. Management says the structure secures stable profitability. I believe that is the aim. It is an unproven claim.

China is the other threat. CXMT’s share of DRAM revenue reached 9.5% in the second quarter, and press reports have it moving toward 420,000 wafers a month by 2027.⁷¹⁹ Every month of high prices funds a challenger that doesn’t answer to shareholders.

My judgement: the industry’s structure is better than it was, and SK hynix’s own position inside it is the best of the three in the product that matters most. Whether that adds up to something that lasts through a glut, nobody can say yet. One downturn with these contracts in place will teach more than any argument.

IIIThe competitive advantage

The table compares each maker’s latest quarter. The quarters end a month apart and prices were rising, so later reporters have an edge.

Same product, same quarter, same margin. Micron’s shares cost more than twice as much for each won of operating profit. Part of the gap is the old discount on Korean shares. Part is that Micron builds in America. I can’t split it, but I note that the same SK hynix profit costs 4.2 times in New York. Investors who can only buy in dollars are paying nearly Micron’s multiple already.

Where SK hynix wins

HBM first. IDC gives it 63.2% of HBM revenue in 2025 and 56.4% in the first quarter of 2026.⁵⁴ Management calls itself first supplier to its key customers, says HBM4 yields beat plan so the target went up, and says it has shipped HBM4E samples.⁶ NAND second, in a narrow way: Solidigm gives it a strong position in high-capacity enterprise SSDs, the products now seeing the best demand.⁴⁶ Cash third. Net cash of ₩98.7T at 30 September is more than the company’s entire 2025 revenue, and it arrived in about a year.³

Where it loses

Here is something the headline share misses. On IDC’s count SK hynix was first in DRAM in 2025 with 34.8% of revenue, and second in the first quarter of 2026 with 29.1%; TrendForce had it at 24.9% in the second quarter.⁵⁴⁷ The counts differ, but the direction is the same. Why? TrendForce’s answer: HBM was the largest part of its bit shipments among the three, so its prices rose less when ordinary DRAM surged.⁷ HBM is sold on annual contracts. Ordinary DRAM is not. So leadership in HBM became a partial lock-out from the steepest part of this rise, and management said on the call it expects the profit gap between HBM and conventional DRAM to narrow in 2027.⁶ That sentence tells you ordinary DRAM has been the better business lately. RBC estimated that HBM prices for 2027 were set 80% to 100% higher, which I can only cite secondhand.²⁶

Scale is the second weakness. Samsung expects to spend ₩76.8T on its chip division in 2026, and has scaled up HBM4.⁹ It can carry more for longer. Geography is the third. Its DRAM is made in Korea and at Wuxi in China, the Chinese plants now run on a yearly US licence, and its only American site, the Indiana packaging plant, starts in the second half of 2028.⁴ Governments and customers planning for trouble in Asia have fewer SK hynix options than Micron options. None of these positions is permanent. They have to be earned node by node.

IVManagement

Kwak Noh-Jung, 61, has been chief executive since March 2022. He holds a Ph.D. from Korea University and came up through manufacturing and technology, including the Cheongju fab.⁴ He took the company through 2023 and has run it through the best period in its history. Above him sits Chey Tae-won, 65, chairman of the SK Group, whose title at SK hynix is Chairman, with responsibility for long-term vision.⁴ He is not on the board. The chief financial officer, Kim Woo-Hyun, came from SK Telecom and SK Broadband, which is a common path in a group like this.⁴

How they are paid, and what that pays them to do

Start with who gets the money. SK hynix’s staff receive a profit share built on 10% of operating profit under Korean accounting.⁴ That is a very large thing. On ₩181.5T of nine-month operating profit, 10% is ₩18.1T C (calculated). Employee benefit expense in the first half was ₩13.8T against ₩5.5T a year earlier.¹⁰ The profit share has no published ceiling that I can find in the filings. Staff are therefore a partner with a 10% claim on the top of the cycle, and they carry no share of a loss year. That pays people to stay, which matters in a business where a few thousand engineers decide the yield. It also means that in a boom an owner’s margin is permanently lower than it looks.

The chief executive’s pay went from ₩4.2B in 2025 to ₩26.1B in the first half of 2026, of which ₩20.5B was bonus.¹⁰⁴ So the top of the company is paid on the same wave. He also has long-dated stock: at 30 June, undelivered rights over 35,045 shares from stock appreciation rights, 39,042 shares from performance units, and options on 5,199.¹⁰ What that pays him to do is produce a strong year and a strong share price. I found no measure in the filings that pays for return on capital across a full cycle, and the filings I read don’t give the terms of the performance units. For a company about to spend more than ever on new fabs, that is the gap I would watch.

How much of their own money is in

Not much. Kwak held 14,312 shares at 30 June, about ₩20B at the close, or $15M C (calculated).¹⁰ All of it came as grants or option exercises in the filings I saw, apart from one open-market purchase in May 2022.¹⁰ Counting everything undelivered at face, his exposure is about ₩133B C (calculated). That is real money to him and far smaller than the sums Micron’s chief executive holds on my Micron page. Cha held 6,834 shares. The independent directors take small stock grants.¹⁰

The people who matter most own through a chain. Chey held 17.90% of SK Inc., which held 32.14% of SK square, which held 20.50% of SK hynix. Multiply it out and he has about 1.2% of the economics of a company he is deemed to control C (calculated).⁴ That is how Korean groups work. It means related dealings deserve attention. SK hynix bought ₩4.7T of assets from SK ecoplant, a group construction company, in 2025, up from ₩1.1T in 2024, and sold ₩2.7T of NAND through ESSENCORE, another group company.⁴ The board approves these and I have no evidence they are unfair. A 1% economic owner with control has different interests from a 99%, and an outside owner has to price that.

What they did with capital over a full cycle

This is the fairest test. From 2016 to 2025 SK hynix earned ₩101.4T of net income and generated ₩183.4T of operating cash, and spent ₩146.1T of that on plant and intangibles, 34% of revenue C (calculated).^14815164 It paid ₩9.0T of dividends and bought ₩1.7T of its own shares. Most of what a decade of memory earned went back into memory.

Put the last two equity moves side by side, because few people have. In July the company sold 17.79M shares for ₩39.9T. In the next three months it bought back 23.69M for ₩40.0T.¹³²⁸ Net, it spent about the same money and ended with 5.9M fewer shares, and the shares will be cancelled in November.²⁹ On that narrow test, management sold high and bought lower. It also told American buyers the money was for capital spending and then spent an equal sum on stock. Money is fungible and the cash pile can do both. A careful holder of the ADS would still notice.

The stated return plan is about ₩100T for 2026: at least ₩30T of dividends and at least ₩60T of buybacks, including most of the Kioxia gains, or about a tenth of the market value.³⁰ The third-quarter dividend is ₩20,000 a share, half of it special, with a record date of 30 November.³¹ A further ₩20T of buybacks runs to January. About half of that is to fund equity pay for executives and staff, so only ₩10T is a return to owners.³² On the call the company said its shares were undervalued. It said that while the price it had paid in the prior programme was above the market.⁶²⁸

Two open items are worth stating. A press report says Solidigm may build a US NAND plant for about ₩60T, and the company says nothing is decided. It is also reviewing a site in Japan and a possible sale of a stake in its Chongqing plant.³³³⁴³⁵ With the Solidigm IPO dropped, any such plant would be paid for by SK hynix’s owners.

VFinancials

Ten years of SK hynix contain two full cycles and the start of a third. Revenue rose 75% in 2017, fell 33% in 2019, recovered, fell 27% in 2023, then rose 102% and 47% C (calculated).¹⁴⁸⁴ Any one year is close to useless as a guide. Averages over a cycle are better. One accounting note: 2016 to 2022 come from Korean-language K-IFRS filings, with restatements taken from the latest filing, and 2023 to 2025 from the audited IFRS statements in the prospectus.⁴ The two bases agree on operating profit where they overlap.

Ignore net income this year

Net income tells you nothing in 2026. In the second quarter it was ₩93.9T, more than the ₩60.5T of operating profit, because the company booked about ₩66T of finance gains, mostly on Kioxia. In the third quarter it was ₩39.8T, less than half the ₩83.3T of operating profit, because of a ₩22.9T valuation loss on investments and an ₩11.6T foreign exchange loss.¹⁰²⁶ Neither has anything to do with making memory. I work from operating profit, tax and capital spending, and I value the investments separately.

What an owner could really take out

Reported earnings flatter this business in any normal year. Over 2016 to 2025, ₩101.4T of net income became ₩36.7T of owner free cash flow, about 36 cents on the won, and that figure was negative in four years out of ten C (calculated). Stock pay barely matters so far, ₩0.4T in 2025, and lease payments of ₩0.6T a year sit outside operating cash flow, so owner cash is about that much lower again.⁴ Depreciation does. It runs well below what the fabs need.

What does it take to stand still? Depreciation and amortization will be about ₩17T in 2026 E (estimate). Capital spending was ₩18.3T on plant in the first half, and it was held back by the lack of cleanroom space.¹⁰ Over the ten years spending took 34% of revenue C (calculated). The industry needs about 20% more bits a year, so holding a place means growing with it, and new nodes give less each time. I put the cost of standing still at about ₩55T a year today, with a range of ₩40T to ₩75T, and I can’t narrow that honestly E (estimate). On that basis, the first nine months of 2026 gave owners about ₩110T: operating profit taxed at 23%, plus depreciation, less three quarters of ₩55T. That is about 64% of the ₩174.0T of reported net income, with a range of about ₩100T to ₩125T E (estimate).²

One open question sits inside that number. The new ₩10T of shares bought for pay: does it fund part of the 10% profit share already in my costs, or is it on top? The filing says it is for equity-based pay including bonuses, and that the amount may change.³² It doesn’t say. If it funds the profit share, it is a change of form and costs nothing extra. I assume that. If it were an extra ₩10T a year, my base would fall by about ₩100,000. If it were ₩10T every quarter, by about ₩400,000 E (estimate). I would want this answered before trusting the base.

The balance sheet, and what I count

The company reports net cash of ₩98.7T at 30 September: ₩114.5T of cash less ₩15.8T of debt.³ Its cash figure includes short-term financial instruments and short-term investment assets, as the June balance sheet shows: ₩26.8T, ₩22.4T and ₩38.7T add up to the ₩88.0T that the third-quarter rise of ₩26.5T starts from.¹⁰³ The third-quarter balance sheet is not out yet, so I use June detail and say where I do. Not all that cash is free.

The investments need a word. Long-term investment assets jumped to ₩85.3T at June from ₩14.5T at December, and ₩82.0T of it had a market-based value.¹⁰ The company took a ₩22.9T valuation loss in the third quarter as Kioxia fell from ¥29,893 to ¥18,095.⁶³⁹ Kioxia closed at ¥13,470 on 30 October, down another 26%.³⁹ So ₩62T or so of book value at September may be nearer ₩46T to ₩53T today. Tax on the gains and a discount for a stake I can’t see inside take that to my ₩36T. The notes don’t break the holding down, and part of it is limited in how and when it can be held, so this is an estimate.¹⁰ Each ₩10T is worth about ₩14,000 a share. That is 1% of the price, and I wouldn’t lose sleep over the error.

Customer deposits are the other unknown. Micron discloses $32B of customer commitments. SK hynix says it has deposits and gives no number, and contract liabilities were ₩0.5T in June, before most of the agreements.¹⁰⁶¹⁸ Other non-current liabilities did rise to ₩2.9T from ₩1.3T.¹⁰ If deposits are large at September, my count is a little too high. The third-quarter report, due by mid-November, will settle several of these lines.

Currency

Revenue is mostly in dollars, and many costs are in won. The noon buying rate was ₩1,538 in early July. The ECB’s rates imply ₩1,335 on 30 October, the dollar buying 13% fewer won in under four months.⁴²⁵ That cost revenue in won, cost ₩11.6T in the third quarter on dollar balances, and helps explain why results in won look less steep than results in dollars.⁶ I hold ₩1,335 flat. At ₩1,200 throughout my base would be about ₩1,763,000 E (estimate). Nobody can forecast this, and for a Korean exporter this size it deserves a line of its own.

Quality of earnings

Receivables reached ₩47.8T at June, up from ₩18.2T in December, on a much larger quarter; customers pay in about 55 days at that run rate C (calculated).¹⁰ Inventories were ₩18.0T, up 26%, while revenue tripled.¹⁰ The ADS raised reported equity, and tax of 22.7% over nine months is well above the 14.9% of 2025, when old losses and credits helped C (calculated).²⁴ I use 23%. Related sales through ESSENCORE rose to ₩1.55T in the first quarter alone, and I would like that line to stay in proportion.⁴

VIValuation

First, how this goes badly wrong

Before any numbers, I asked what would most likely ruin this investment, and worked back. Three things can.

• The new fabs all arrive together. Yongin’s first cleanroom opens in early 2027 and SK hynix says shells at Yongin and M17 will speed up, with tools added to match demand.⁴⁶ Samsung’s chip capital spending is ₩76.8T this year. Micron’s Idaho, Tongluo and Hiroshima sites arrive in 2027 and 2028. CXMT is adding capacity.⁹¹⁹ Each choice is rational alone. Together they are how memory downturns start. Asked about this on the call, management said more fabs don’t mean more bits, and that it will phase cleanrooms. Every maker says that at this point in the cycle.⁶
• Part of today’s demand is fear of running out. The prospectus itself warns that current demand may reflect inventory building and that real end demand may be lower than orders suggest.⁴ Buyers who are short order double. If AI spending pauses, those orders go away just as new supply comes on. Nearly half of 2027 and 2028 capacity has no long-term agreement, and the half that does has a price band I can’t see.⁶
• One customer changes its mind. The largest buyer was 23.9% of 2025 revenue.⁴ A shift in how AI systems use memory, or Samsung winning a platform, or lower stacks, would land first on HBM, the product that makes SK hynix different. The prospectus warns that technology may reduce memory needed for AI.⁴

None of these needs a surprise. They set my bear and the reason I use a through-cycle margin at all. North Korea sits outside the model. The prospectus lists tension there as a risk to the shares, nearly all the leading capacity is in Korea, and no spreadsheet handles that.⁴

What the price implies

At ₩1,418,000 the market values SK hynix at ₩1,000T, about $749B, on 705.2M shares outstanding after the buyback.¹³¹ Book equity is about ₩302T by my count, June’s ₩262.7T plus the ADS proceeds and third-quarter profit less the buyback, so the price is about 3.3 times book E (estimate). The price is 2.9 times the Nasdaq consensus for 2027 and 3.4 times 2029’s, on two analysts C (calculated).¹²²⁵ Low multiples at a peak are how cyclical stocks usually look, so that alone proves nothing. The market is plainly saying these earnings won’t last. The share has fallen 51% from ₩2,919,000 on 22 June, and it fell 58% at the low.¹ That has happened with estimates still going up.¹²

A growth rate from a peak year means nothing here, so I ask the question another way. Take my base cash through 2029. That is ₩900T, or 90% of today’s market value, in a little over three years. Discounted it is ₩719T. Add what I count outside the business and the price has only ₩203T left to pay for. That needs ₩25T of owner cash a year from 2030, growing 3%, for good C (calculated). In 2025 the company generated ₩24.4T. So if my next three years are right, the price asks the business never to be much better than it was last year.

The “if” carries everything, and it is why the market may be right. Most of that ₩900T hasn’t been earned. It rests on selling prices that are several times what they were a year ago. If 2028 is half my number, the cushion goes. A market that has seen 2023 is entitled to pay little for such cash.

Put in margin terms, the price needs operating margin to settle near 27% on my base costs and volumes, with the next low near 24%. The ten-year average was 29.7%. So at this price the market assumes the old world comes back, on my base costs, after the boom. At the ADS price, in New York, the same test gives 44%.

What I can’t judge

I can’t tell you the normal margin of a memory maker after this, and nobody can. No downturn has come since these agreements began, and no earlier cycle had AI in it. I also can’t judge how long the HBM lead lasts, how the price bands work, or what a Taiwan or North Korea event would do. Here is the range, and read my ₩2,067,000 with it in mind.

The discount rate

I use 12.0%, the same as on my Micron page, in won. The Korean ten-year government bond yielded 4.01% at the close, and Damodaran’s implied equity risk premium was 4.20% on 1 October, with 0.64% for Korea.⁴⁰⁴¹⁴² My betas against KOSPI, on month-end returns, are 1.62 over ten years, 1.78 over five and 1.90 over three C (calculated).¹⁴³ The ten-year figure gives 11.9% and the five-year 12.6%. I take 12.0% because the downturn is already in the cash flows. It would be easy to argue for a point more on governance. A point more takes the base to about ₩1.94M, still above the price.

Three cases

I keep the model simple, and I use the same shape of cycle as on my Micron page so the two can be compared. I grow bits, take cost down per bit, roll depreciation from capital spending, and set one number a year: revenue for 2027 and 2028, operating margin after that. Costs are three pieces: cash costs of about ₩34T in 2026, depreciation at 21% of opening plant, and the staff profit share at 10% of operating profit. The fourth quarter of 2026 is my estimate: ₩130T of revenue and ₩105T of operating profit, which puts 2026 near ₩366T, a little over the ₩361T Naver consensus.⁴⁴² Tax is 23%, working capital 15% of added revenue, and each year’s cash arrives at year end. I hold the won at ₩1,335.

Revenue falls 65% from peak to trough in my base, when the worst fall in the past decade was 33% from 2018 to 2019 C (calculated). The peak is much further above cost than any before, so to reach even a 42% margin price has to fall a long way. Revenue of ₩239T in 2031 is still 2.5 times 2025’s.

Most of my bear is cash that comes soon: 71% of its value is cash due through 2029. So I also ran a stress case with no weight: 2027 revenue of ₩400T, 2028 of ₩220T, margins down to 5% and the first agreements renegotiated. It gives ₩764,000, about 46% below the price. I think it is unlikely. It is close to what happened in 2023, scaled up.

How good, and how good a price

These are two separate questions. The business: a very good position in a brutal industry, with group control that an owner has to discount, and the best product it has ever had. It isn’t a great business in the sense of steady returns on capital, since 36 won in 100 of ten years’ profit reached owners. The price: on my numbers, this looks like a merely good business at an unusually low price for what it is earning, the reverse of the usual AI stock. That is a statement about my numbers, and a low price on peak earnings is the most common trap in this sector.

Other uses of the same won

The Korean ten-year pays 4.01%. My base here returns 22.6% a year at this price and my bear 5.5%, against 12.0% required E (estimate).⁴⁰ The closest comparison on this site is Micron: on that page the base returns 7.1% and the bear −3.4%, with the price 26% above base. Same industry, same cycle, same method. TSMC, with far steadier prices, is 9.4% and 5.3%. NVIDIA is 9.7% and 5.5%. Those pages use their own prices and rates. If the cycle is what I fear, TSMC keeps its earnings and SK hynix doesn’t, so these returns don’t carry the same risk.

The nearest alternative of all is the same share in another market. The ADS closed at $149.21, or $1,492.10 a share. The Korean close converts to about $1,062, so the ADS costs 41% more for the same claim, down from nothing near that at listing: about 3% at pricing and about 16% after the first day.^1112545 Why doesn’t it close? New ADSs need the company’s consent, the issue was capped at 2.5% because SK square must stay at 20%, and shares withdrawn may not be re-deposited.⁴ So money flows out of ADSs freely and in only with permission. The premium measures how much American funds want this exposure in dollars. In my base the ADS is worth about $155, about 4% above the close, and it returns 12.9% a year; on my bear it returns -1.3%. An American buyer takes two bets, on memory and on the premium. If it fell to zero with the Korean price unchanged, the ADS would lose 29%. This page’s values are for the Korean share. I don’t assume the gap closes, and I can’t judge when it might.

What past lows looked like

Five falls of 30% or more in ten years: 39% in 2018 and 2019, 34% in 2020, 49% from 2021 to 2022, 37% in 2024, and 58% to the low of ₩1,223,000 this October C (calculated).¹ So the deepest fall in a decade has already happened, in four months, with the reported numbers rising. In 2018 the high close came in May, in a year that went on to set a profit record that stood until 2025.¹⁸⁴ The market has a habit of being early.

Book value is the old floor. At the 2019 and 2022 lows the shares sold near 0.8 to 0.9 times book C (calculated).¹⁸¹⁶ It sells near 3.3 times today. If my base is right and nothing were paid out, book would reach about ₩1.63M a share by the end of 2028 E (estimate). That is above today’s price. I read that as a measure of how much the market doubts the next two years. It is no forecast of a floor.

The margin of safety

• Price₩1,418,000Close, Q4 2026 [1]
• Base value₩2,067,000My estimate E (estimate)
• Margin of safety below₩1,479,000set from my bear case C (calculated)

I set the margin of safety from the bear case, as on my Micron page, because one cycle drives this business. Below ₩1,479,000, 28% under my base value, my ₩1,109,000 bear value is at least three quarters of the price paid. The price is 4% under that line. I don’t often write that, so I went looking for what was wrong with it. Three things. The bear’s value is 71% cash yet to be earned. At ₩1,200 to the dollar the bear falls to about ₩927,000, and the line with it. And the stress case sits well below the price. So the rule is met, narrowly, and the safety it describes depends on 2027 arriving much as priced. That is the one thing a cycle can take away. The agreed 2027 HBM prices and the signed volumes are real support for that year, and they are a year.

Where this could be in 5 and 10 years

Each value is my model’s value at that date: the owner cash that comes after it, discounted at 12.0% back to that date, plus the ₩78T I count outside the business, held flat. It is per today’s share count and in that year’s won. Cash paid out is everything before October 2036, with nothing reinvested.

For the base to be right at 5 years, 2027 and 2028 have to bring revenue of ₩560T and ₩690T, and margin has to bottom at 42% in 2031, not lower. An owner will have been paid ₩1.47M a share by then, 71% of today’s base value, so the value left is lower, ₩1.54M, even if I am right. For year 10 the base needs revenue near ₩321T and a settled margin of 45%, which is ₩96T of owner cash in 2036. The numbers that drive it are that margin, capital spending of ₩70T to ₩90T a year, and the 12.0% rate. Of the year-10 value, 86% is what comes after 2036, so it leans on a far-off guess.

The bear and bull differ on how long the boom lasts. In the bear, revenue falls to ₩127T by 2030 and margin settles at 30%. In the bull, revenue stays between ₩715T and ₩824T and margin fades only to 63%. The most likely way my base is wrong is that settled margin. At 35% instead of 45%, the year-10 value is ₩1.24M, not ₩1.77M C (calculated). The yearly return differs from the one above because this one stops in 2036 and values what is left at 12.0%.

These are estimates of value, not forecasts of the share price. The price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three things cut in the company’s favour and don’t make the headlines. First, the HBM contracts that held revenue back in 2026 are about to reset. HBM is priced a year at a time, which is why SK hynix lost DRAM share while ordinary chips soared.⁷⁵⁴ Talks for 2027 are finished, and management expects the profit gap with ordinary DRAM to narrow.⁶ RBC’s estimate, which I can only cite secondhand, is that 2027 HBM prices were set 80% to 100% higher.²⁶ Second, the share count. In about three months the company sold 17.79M new shares near ₩2.24M and bought 23.69M at about ₩1.69M, for the same money.¹³²⁸ Third, the ADS premium looks like froth but has a legal cause: SK square’s 20% floor capped new ADSs at 2.5% of the company, and more need the company’s consent.⁴ Demand for this share in dollars can’t easily be met.

The bull case

In plain words: memory has become the limit on AI, SK hynix owns the best position in the scarcest memory, and the price treats that as a one-time gift. Kwak told holders the business environment is changing, and that long-term agreements are giving the company a stronger base for sustained profit.⁶ Management says 2027 to 2029 DRAM and HBM are effectively sold out, counting volumes under discussion, that supply will trail demand through 2030, and that HBM profitability will stay far above past peaks.⁶³

Three investors put a number on it at the listing. The prospectus names Baillie Gifford, Coatue and Situational Awareness as cornerstones who indicated interest in up to $7B of ADSs at $149.⁴ That interest was not binding, and it doesn’t tell me what any of them bought. No 13F yet shows SKHY: the first will be for 30 September, due in mid-November. On the sell side, Rosenblatt raised its target to $400 from $320 and RBC kept $200 after the results, as reported secondhand via The Fly; the article does not say per what, and I take it to be per ADS.²⁶

Against reality

Today’s numbers support the first half. If SK hynix does what it says through 2028, my base has it generating ₩900T by the end of 2029, 90% of the market value E (estimate). If they deliver 2027 and 2028 and margin settles anywhere above about 27%, the price is possible.

The bull case needs much more, and multiplying shows how the odds shrink. Say tightness lasts through 2028, 70%. New fabs don’t cause a glut in 2029 or 2030, 50%. AI spending keeps rising to 2030, 60%. Price bands hold when the market turns, 60%. SK hynix keeps its HBM lead, 70%. Those are my guesses and they aren’t independent. Multiplied, they give 9% E (estimate). Allowing for overlap, perhaps 20% to 25%, the weight I give the bull. Each step looks likely. All five together do not.

What today’s price needs is shorter. 2027 arrives near my number, 85%. 2028 doesn’t collapse, 65%. Margin afterwards beats the old average, 70%. That multiplies to 39%, and with overlap perhaps 50% to 55% E (estimate). So it is about an even bet on my own guesses, which is a fair description of why a stock can sit at three times earnings.

What a 10x would take. Ten times ₩1,418,000 is ₩14,180,000, about ₩9,999T or $7.5T of market value. NVIDIA, the largest company I cover, is about $5T.⁴⁶ At 20 times earnings that needs about ₩500T of net income every year, about a fifth above my base peak, as a normal year, at about seven times today’s multiple on Street 2027. At a 40% net margin, far above the 24% of the past decade, that is ₩1,250T of revenue, or $936B, more than the whole memory market WSTS forecast for 2026.¹⁷ With a quarter of DRAM, the world would need to spend trillions of dollars a year on memory. By 2036 that means 26% a year for ten years from about $749B C (calculated). The stock has risen about 35 times in ten years, from ₩41,000, but that began under ₩30T of value.¹ The best run I know from this size is NVIDIA’s: about five times from its $40.11 close of May 2023, near $1T, in about three and a half years.⁴⁶ SK hynix would need twice that, in memory. The multiple helps, since a stock at three times earnings has more room than most. Tenfold from here is still close to impossible, and I wouldn’t build a case on it. Getting back to June’s high would be about double.

Where I come down

The 2027 pricing and the size of the near-term cash moved me most: they are why my base sits above the price. The talk of sold-out years to 2029 did not, because unsigned volume in a shortage is a soft thing. A first downturn in which the bands hold and margin stays above 50% would raise my values most.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• The cycle itself. Cathie Wood said in Hong Kong on 13 October, as reported by Barchart, “The trend is up, but the cycles are vicious,” and warned that Chinese capacity would make swings larger.⁴⁷ She was speaking of memory makers in general. The share has fallen 30% from a high five times in ten years C (calculated).¹
• Korea. Shares in Korean groups have long traded at a discount. Recent changes in the Commercial Act may narrow it, and that is something I can’t put a number on.⁴ The prospectus also lists North Korea and the health of the Korean economy as risks.⁴
• China. Revenue through Chinese subsidiaries was 24.3% in the first quarter.⁴ The US revoked the validated end-user status of the Chinese fabs at the end of 2025 and replaced it with a yearly licence, granted for 2026. A refusal would hit Wuxi and Dalian.⁴ China’s regulator also limits NAND pricing there through 2026, under the terms of the Intel deal.⁴
• Currency. A stronger won cuts revenue in won and hit the third quarter with an ₩11.6T loss.⁶
• Legal. An indirect-purchaser antitrust class action over DRAM supply was filed in California in June 2026. It says three makers conspired to restrict supply. Two ITC investigations by MonolithIC are running, with a first target date of 30 August 2027. A bench trial over the Indiana site’s zoning is set for December 2026.⁴ At these margins an antitrust regulator somewhere will take an interest, and memory makers have been fined before.
• Staff. Three unions hold collective agreements, and the 10% share is now large enough to attract attention from all sides, including Samsung’s workforce.⁴ I have no evidence of trouble. Wage talks happen every year.
• Solidigm and new sites. A ₩60T American NAND plant and a Japanese site are unconfirmed. Either would be a large call at the top of a cycle.³³³⁴
• The ADS itself. Holders of the ADS carry the premium, which has ranged widely, as well as the business. See Valuation.⁴⁵

Geopolitical risk, and what it means for the value

SK hynix is exposed in four places: the fabs it runs in China, the US rules on what can be shipped there, the tariffs on what it sells, and the peninsula it sits on. Taiwan matters differently. It owns no leading plant there, but the AI accelerators its memory sits beside are made and packaged there. Each scenario below reruns my base, ₩2,067,000, with one change.

Wuxi makes about 40% of its DRAM and Dalian about 30% of its NAND.⁴⁹ The US ended their validated end-user status on 31 December 2025 and replaced it with a yearly licence; 2026’s was granted.⁴ Commerce decides each year, and I found no announcement for 2027. If it refuses, the fabs keep running but fall behind, and I assume a fifth of owner cash is gone by 2031. If they are lost outright, I take away 35%. HBM sold to China already needs a licence under the December 2024 rule, and sales to Huawei stopped in 2020, so China’s 24.3% of revenue is ordinary memory.⁵⁰⁴

Tariffs are the White House’s call. The reciprocal tariffs exempted chips and were struck down in February 2026; a 100% chip tariff has been threatened twice, with exemptions for firms that build in the US.⁴ Korea’s deal promises terms no worse than other major countries’.⁵¹ The Indiana plant starts in the second half of 2028, so a US-made exemption would favour Micron’s Idaho fab first. Samsung stands under the same Korean deal.

Seoul’s rules cut both ways. Since March 2026, bought-back shares must be cancelled within a year, which helps owners and is not in my value.⁵² The corporate rate is 25% nationally, plus 2.5% in local tax, or 27.5% in all, against the 23% I use.⁵³ The won is a currency risk more than a political one; my page already shows ₩1,763,000 at ₩1,200.

In the price: the close is below my value in every case except the loss of China, so the market already allows for a good deal. I can’t separate how much of that is geopolitics and how much is the cycle. Taiwan and the peninsula are not in my base, and a long war would be worse than a year of lost cash, which this model understates.