Samsung Electronics

IThe business

Samsung Electronics is four companies under one roof. Device Solutions makes chips: memory (DRAM and NAND), the chips inside phones that Samsung designs itself (System LSI), and a foundry that makes chips for other companies. Device eXperience makes Galaxy phones, TVs and home appliances, and sells network equipment. Samsung Display makes screens. Harman makes car electronics and audio.⁵ The company is listed on the Korea Exchange as 005930, and it has a second, non-voting class of share, 005935, which pays a slightly higher dividend under its terms.⁵ This page values the ordinary share in won, and treats the preferred as the same claim at a lower price, which I come back to.

In 2025, before the shortage, phones and appliances were the biggest part. Revenue was ₩333.6T. Device eXperience sold ₩188.0T and earned ₩12.9T. Device Solutions sold ₩130.1T and earned ₩24.9T. Samsung Display sold ₩29.8T and earned ₩4.1T, and Harman ₩15.8T and ₩1.5T. By product, smartphones were ₩126.5T, memory ₩104.1T, TVs and monitors ₩30.9T and display panels ₩29.8T.⁵ Then the chip division became the company. In the first quarter of 2026 it earned ₩53.7T, in the second ₩89.2T and in the third ₩106.5T. The company as a whole earned ₩107.4T in the third quarter, which Samsung says makes it the first technology company to pass ₩100T of operating profit in a quarter.¹⁴⁸³ In the nine months operating profit was ₩254.1T, nearly six times all of 2025 C (calculated).

Look at where that came from. Memory revenue was ₩104.1T in the whole of 2025. It was ₩74.8T in the first quarter of 2026, ₩120.8T in the second and ₩140.7T in the third.⁵⁸³ Costs did not move with it: gross margin went from 39.4% in 2025 to 69.3% in the third quarter, because a chip costs about the same to make whatever it sells for C (calculated).⁵³ So nearly every extra won of price became profit. That is the fact that matters most today, and it cuts both ways.

Why memory became scarce

Memory has the same problem the page on SK hynix describes. AI servers use stacked high-bandwidth memory (HBM) next to the processor, ordinary server DRAM next to the central chip, and fast flash drives beyond that. Each HBM chip takes more wafers than an ordinary one, so HBM growth takes capacity away from every other kind of memory. New factories take years, and finer manufacturing steps add fewer chips each generation. Samsung says that in 2027 AI-driven demand will far exceed what the industry can supply, and that it will sell more HBM, server DRAM and drives while securing multi-year contracts.²³ The June interim note says it has entered multi-year supply agreements with several global customers, and gives no volume or price.⁸ Contract liabilities, which include prepayments, rose to ₩23.6T from ₩14.0T in six months.⁸

These are company claims, and every maker is saying them. Consensus assumes it: Naver’s 2026 operating profit estimate is ₩382.8T, against ₩254.1T reported for nine months, so analysts expect about ₩129T in the fourth quarter.¹⁷³ My own fourth quarter is lower, ₩116T, because I do not assume the rate of increase from the second to the third quarter repeats E (estimate). For 2027 the spread is very wide. Goldman Sachs put it at ₩438T and SK Securities at ₩493T in spring, before the latest results.¹⁸ Even two houses disagree by ₩55T, which tells you how little anyone knows.

What it sells, and to whom

The chip division is a bundle. Memory is by far the largest part. Samsung does not report System LSI or foundry results separately. The one outside estimate I found, from Kiwoom Securities, has the two together losing ₩6.74T in 2025 and ₩3.92T in 2026, with the loss narrowing as HBM4 base dies, which Samsung makes on its own foundry process, add foundry revenue.¹⁹ The foundry has about 5.9% of the world market by revenue against TSMC’s 72.5%.²⁰ Samsung says it has won 2-nanometer orders from large cloud customers and will open its first Taylor, Texas fab in 2027.²

Device eXperience is a different business with a different cost problem. Phones, TVs and appliances use memory, and memory prices have gone up. IDC says third-quarter smartphone shipments fell 22% from a year earlier to 255.3M, and that memory costs, up about 350% in a year, were the main reason.²¹ Samsung kept its place as the largest maker, with 57.8M phones and 22.6% of the market, and still lost money on phones and networks in the second and third quarters.²¹³ The division earned ₩3.0T in the first quarter and lost ₩0.8T in each of the next two C (calculated).¹⁴⁸³ Hold that thought: one part of the company is paying higher prices for what another part sells. I come back to it, because it is the nearest thing to a hedge in the numbers.

Customers are not named. Samsung Display sells screens to phone makers, and Harman sells to carmakers. The memory and foundry customers are cloud and chip companies, and the filing says only that supply agreements exist.⁸ The regional split is in the annual notes, and I did not need it for the argument.

What the plant costs

Samsung expects to spend ₩84.4T on plant in 2026, of which ₩76.8T goes to chips and ₩3.9T to the display business, up from ₩47.5T of cash spending on plant in 2025.²⁵ Contracted spending not yet paid was ₩30.5T at 30 June.⁸ The third quarter alone was ₩29.0T. Depreciation was ₩11.9T in that quarter, so the company is spending more than twice what it depreciates C (calculated).²³ Spending at that rate is how this industry grows, and it is also how it breaks.

IIThe moat

Protection shows up as profits that rivals cannot take. For memory the record is the test, and it is mixed. From 2016 to 2025 Samsung’s operating margin averaged 14.6%, weighted by revenue, and swung from 24.2% in 2018 to 2.5% in 2023.^1541695 The company never lost money in those years, which is the first thing that separates it from its rivals. In 2023 the chip division lost ₩14.9T on ₩66.6T of revenue. The phone, TV, display and Harman businesses earned ₩21.1T that year, and the company as a whole earned ₩6.6T C (calculated).⁹ SK hynix, which has only memory, lost ₩7.7T in the same year.

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%; Counterpoint, counting differently, had 38%, 25% and 24%, with the Chinese maker CXMT at 10%.⁶⁷ A newcomer needs years of process work and tens of trillions of won. That keeps prices firm, but it leaves three firms free to add capacity in the same year, and that is how the last busts began.

Evidence it is widening

The first fact is that Samsung came back. Counterpoint’s count has its share of HBM revenue at 33% in the second quarter, against 15% in the second quarter of 2025, while SK hynix fell to 50% from 64%.⁷ TrendForce says Samsung’s early HBM4 shipments gave it the strongest bit growth of the three makers in the same quarter.⁶ The second is the company’s own claim that it is signing multi-year agreements, which, if they hold, would make revenue less dependent on the spot price.³ The third is the foundry: HBM4 base dies mean the chip businesses now feed each other, and Samsung raised prices on new 4-nanometer orders in July, according to the Kiwoom note.¹⁹

Evidence it is narrowing

China is the main threat. CXMT held 10% of DRAM revenue on Counterpoint’s count in the second quarter, and every month of high prices pays for its next factory.⁷ Samsung also makes memory in China, and the United States ended the special status that let it ship chipmaking tools there freely at the end of 2025. It has an annual licence for 2026, and nothing I found says what happens after.²² Second, the contracts. A multi-year agreement with no disclosed floor is a promise from a buyer who is paying a very high price. I cannot say how much of it would hold in a real fall. Third, the workforce now owns a claim on profit that no earlier cycle had, which I take up under Management.

My view: the structure is firmer than it was and Samsung is the most diversified of the three. Whether that survives a glut cannot be known until there is one with these agreements in force.

IIIThe competitive advantage

The table compares the latest quarter of Samsung and SK hynix, the other large memory maker I cover. Samsung’s chip division includes the foundry and System LSI, which lost money, so its margin understates its memory margin.

The two are close. Samsung costs 3.3 times its annualized third-quarter operating profit and SK hynix costs 3.0 times, on my SK hynix page. Samsung’s profit is bigger and its margin is lower, because it carries a foundry that loses money and a phone business under cost pressure. After my adjustments Samsung has ₩201T of cash and holdings outside its chip business that I count, and SK hynix about ₩78T on its page. If you own either, you own the same bet on the memory price, and the two are not independent.

Where Samsung wins

Scale first. It is the largest DRAM maker, with 98% of this year’s operating profit coming from its chip division, and it can afford to build ahead. Breadth second: it has a foundry for HBM base dies and sells its own phones, so some of what it pays more for in one division it earns in another. Cash third. Net cash of ₩197.6T at 30 September is 59% of the company’s entire revenue in 2025, and about twice SK hynix’s.³⁵ And HBM, where it is coming back: 33% of revenue on Counterpoint’s count, from 15%.⁷

Where it loses

HBM share is still behind. SK hynix had 50% on the same count to Samsung’s 33%, and Samsung is the smaller supplier of the product the AI buyers want most.⁷ The second weakness is the foundry: 5.9% of the market against TSMC’s 72.5%, still losing money, and dependent on a Taylor, Texas fab that Samsung says starts operating in 2027.²⁰¹⁹² The third is phones. The division that sells the largest number of phones in the world lost money in two quarters in a row at the moment its sister division made a record. None of these gaps is fixed; each moves with the next product generation.

IVManagement

Lee Jae-yong is the chairman and the person the group answers to. He has not drawn a salary since 2017 and is not in the pay disclosure.²³ The two executives who run the businesses are Jun Young-hyun, who leads Device Solutions, and Roh Tae-moon, who leads Device eXperience.²³ Samsung does not have a holding company. Control runs through a chain of cross-holdings: Samsung Life Insurance is the largest holder of Samsung Electronics with 8.51% of the shares, and the largest-shareholder group, which includes affiliates, family members and executives, holds 19.7% of the voting shares.¹⁰ Press reports put Samsung C&T above Samsung Life in the chain; I did not open a filing for the exact percentages and leave them out.

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

Start with the staff, because that is where the new money goes. The average employee earned ₩158M in 2025, up 21.5%.²³ Then came the strike threat. The union wanted 15% of operating profit as bonus and no cap. A deal reported on 21 May, which suspended a planned 18-day strike, set 10.5% of the chip division’s profit aside as special bonuses, paid in stock, and Reuters’ report mentions no cap.¹¹ Members ratified it on 27 May by 73.7%, and a chip employee can receive up to ₩600M in stock if Samsung reaches ₩300T of operating profit this year, with payouts over at least 10 years.²⁵ JPMorgan estimated all performance pay for 2026 at about 12% of operating profit.¹¹ A small group of individual shareholders said they would sue if it were ratified; I found no later report.¹¹

That is a very large thing. If 10.5% applies to the chip division’s nine-month operating profit of ₩249.4T, it is ₩26.2T C (calculated). The June notes say the shares will be granted on the final 2026 result and that the detailed terms are under discussion.⁸ Wages and salaries were ₩25.6T in the second quarter against ₩8.5T a year earlier, which is where I see it in the accounts, and so operating profit is already after it.⁸ Staff therefore hold a claim on about a tenth of chip profit in a boom and share none of a loss. That keeps engineers, who decide yields, from leaving. For an owner it means the boom margin is lower than the gross numbers suggest, for as long as the formula runs. SK hynix has a similar arrangement, 10% of operating profit, and I use the same approach on its page.

The top is paid differently. Executives get salary and a bonus, and a three-year plan: a Performance Stock Units programme granted on 14 October 2025 with a target of 36M shares, vesting in October 2028 if the share price rises by a set minimum from its level at grant, ₩91,600. The expense was ₩0.17T in 2025 and ₩0.77T in the first half of 2026.⁵⁸ The shares are now worth about 2.5 times the base. What that pays them to do is lift the share price, which in a boom is mostly the memory price. Nothing I read pays for return on capital over a whole cycle, and with ₩84.4T of plant coming that is what I would want added.

How much of their own money is in

The chairman’s stake is the main one: 1.75% after the gift, about ₩23T to ₩26T at the close depending on whether the percentage is of ordinary shares or of all shares E (estimate).¹⁰ The group above him, which has control, holds 19.7% of the voting shares.¹⁰ So a vote is cast with a fifth of the votes by people who own a small share of the economics. That is how Korean groups work, and it means related dealings and the use of the cash pile deserve attention. For the two division heads I could not find a holdings figure in anything I opened, so I leave it out. The gift tax on the transfer from his mother is estimated at about ₩1.56T, to be fixed early next year, and the article does not say how it will be paid.¹⁰ The family finished paying about ₩12T of inheritance tax on Lee Kun-hee’s estate in April, after Hong Ra-hee’s block sale.²⁴ So the old overhang, which weighed on the shares for five years, is gone, and a smaller one has begun.

What they did with capital over a full cycle

The ten years to 2025 give the fairest test. Samsung Samsung earned ₩341.6T of net income and generated ₩617.0T of operating cash, and spent ₩437.8T of that on plant and intangibles, 16.7% of revenue C (calculated).^1541695 It paid ₩100.4T of dividends and bought back ₩26.9T of its own shares. That left ₩41T more cash and short-term deposits than it started with, ₩125.8T at the end of 2025.

Now the 2026 plan. On 21 August the board said total shareholder returns for the year would be ₩90T to ₩110T, which The Elec called the largest by a Korean company, about five times the 2020 record of ₩20.3T.²⁹ The pieces so far are the third-quarter dividend, ₩4,601 on every ordinary and preferred share, ₩30T in all, and a ₩15T buyback of 56.86M shares from 24 August to 6 October at an average of ₩263,823, now worth ₩12.8T at the close.²²⁹²⁸ The ₩15T is held for staff pay: performance incentives, special management bonuses and other uses, The Elec reports.²⁹²⁸ It is not a return to owners. The real distribution so far is the dividend and about ₩5T of shares cancelled in March, the ₩35T the company counts.³⁸ The remaining ₩55T to ₩75T will be set at a board meeting in late January and will include a “substantial” buyback and cancellation programme, with the share of preferred shares in it under review.³ The policy for 2024 to 2026 is to return 50% of free cash flow with a regular dividend of ₩9.8T a year.²⁷ On the nine months’ ₩184T of free cash flow that is about ₩92T C (calculated).³⁸

A new law bears on the staff buyback. The Commercial Act amendment passed on 25 February requires companies to cancel treasury shares within a year unless shareholders approve a retention plan, and employee pay is an allowed use, with 18 months for shares already held.³⁰ Samsung holds 136.7M ordinary shares, 2.34% of the class, for that purpose.²⁸ If the staff shares are delivered in the planned way they do not dilute owners, because they were bought. That is why I count them out of the share count and deduct the cost once, in the operating profit.

VFinancials

Ten years of Samsung contain two full memory cycles and the start of a third. Revenue rose 18.7% in 2017, fell 14.3% in 2023 and has risen 16.2% and 10.9% since C (calculated).¹⁵⁹⁵ Operating profit ranged from ₩6.6T in 2023 to ₩58.9T in 2018, so a single year is a poor guide. Cycle averages say more. A note on per-share figures: 2016 and 2017 per-share figures are shown after the 50-for-1 split of May 2018, which I applied myself, 157,967 won to 3,159 and 299,868 to 5,997.¹⁵³¹

Cash flow runs ahead of profit this year

Reported operating cash flow is a poor guide in 2026, for a large reason and a small one. The small one: finance income and expense swung from +₩4.7T in the second quarter to −₩9.5T in the third, so net income of ₩75.1T in the third quarter ran ₩32.3T below operating profit.³ The large one is tax. The company earned ₩251.8T before tax in nine months and paid tax of only ₩6.1T in the first half, while it booked ₩34.4T of tax expense in the same period.⁸³ Income tax payable was ₩24.4T at 30 September against ₩7.0T at the end of 2025.³⁵ Operating cash flow of ₩244.0T for nine months is therefore too high by the tax not yet paid. I work from operating profit, tax at the full rate, and capital spending.

What an owner could really take out

Over 2016 to 2025, ₩341.6T of net income became ₩179.2T of owner free cash flow, about 52 won on every 100 of profit C (calculated). It was negative in one year, 2023, by ₩16.4T. Stock pay was small until now: the performance units cost ₩0.17T in 2025 and the share grants to staff about ₩0.5T.⁵ In 2026 it is large, and it is in operating profit.

What does it take to stand still? Depreciation and amortization were ₩46.9T in 2025 and I estimate ₩49T in 2026 E (estimate). Plant spending in 2025 was ₩47.5T. Over the ten years spending took 16.7% of revenue C (calculated). Demand for bits grows about a fifth a year, so keeping a place means spending to grow. My figure for the cost of merely keeping up is ₩55T a year today, between ₩45T and ₩70T E (estimate). On that basis the first nine months of 2026 gave owners about ₩190T: operating profit taxed at 24%, plus depreciation and amortization, less three quarters of ₩55T. That is 98% of the ₩193.6T of net income to owners, so on this measure the reported figure is about right. Net income is a fair guide for Samsung in a way it was not for SK hynix in 2026, because Samsung has no large investment gain to strip out. What it is not is cash in hand: the growth capital on top of standing still, ₩29T this year, is spent first.

The balance sheet, and what I count

The company reports net cash of ₩197.6T at 30 September: ₩214.8T of cash, short-term deposits and short-term assets less ₩17.2T of debt.³ That is not all of the liquid money. In the first half it moved ₩32.4T into long-term financial instruments, which do not appear in the cash line and which I find in other non-current assets, up ₩32.7T since December, and sold ₩1.1T.⁸ It is also not all free. The third-quarter dividend of ₩30T has a record date of 30 September, as the first two quarters of 2026 did, so the ₩225,000 price is without it.²⁸

The holdings need a table of their own, because they are where Samsung looks most like a holding company. At 30 October prices the stakes in listed companies are worth ₩72.7T, 5.1% of the market value.⁸³²²⁶ I take 25% off for the tax on a sale and for stakes that Samsung will not sell, since the Electro-Mechanics, Biologics, SDI and SDS holdings are core to the group.

One of them has already told you how much to trust such marks. Corning closed at $123.31 on 30 October against $255.43 on 30 June, so ₩15.6T disappeared from the holdings in four months. The 3Q deck shows investments of all kinds down ₩7.6T in the third quarter, net of purchases.²⁶³ I do not know what Samsung will do with it, and nobody has told me.

Customer prepayments are the other unknown. Contract liabilities were ₩23.6T at June against ₩14.0T at December, and they include advances received.⁸ Samsung does not say which part is prepayment for memory. If it is large, my count is a little too high, because that cash is owed back in product. The third-quarter statements, due in November, will settle several of these lines.

Currency

Memory is priced in dollars and the costs of the Korean plants are in won. The ECB’s rates imply ₩1,335 to the dollar on 30 October.³³ The won has strengthened this year: my SK hynix page has it at ₩1,538 in early July. I hold ₩1,335 flat. At ₩1,200 throughout, my base would be about ₩307,000 and my bear ₩172,000 E (estimate). No one forecasts currencies well, but an exporter this large owes readers the number.

Quality of earnings

Receivables were ₩106.1T at 30 September against ₩50.5T a year earlier, on revenue that has more than doubled; at the third-quarter run rate customers pay in about 49 days, against 54 a year ago C (calculated).³ Inventories were ₩72.0T, almost unchanged from June, while revenue rose 14%.³ Research and development is ₩14.4T a quarter, 7.4% of revenue.³ The tax rate of 23.8% in the third quarter is what I use.³ Gross margin of 69.3% in the quarter is about three times its ten-year level of 39.1%, so none of it is a normal level.

VIValuation

First, how this goes badly wrong

I start from the ways an owner loses money here, then price it. There are three.

• The new fabs all arrive together. Samsung plans ₩84.4T of plant spending this year, and a ₩30.5T commitment was still unpaid at June.²⁸ SK hynix and Micron are building too, as my SK hynix page describes, and CXMT held 10% of DRAM by Counterpoint’s count.⁷ Each build makes sense alone; all at once, they are what turns a shortage into a glut. Every maker says new fabs do not mean proportionally more bits.
• Part of today’s demand is fear of running out. Smartphone shipments fell 22% in the third quarter because memory cost more, and IDC puts memory costs up about 350% in a year.²¹ A buyer afraid of running short orders twice. If AI spending pauses, the extra orders vanish just as new supply arrives, and phone and PC makers, already cutting volumes, are the first to step back.
• Costs that rise with the boom do not fall with it. The staff bonus runs for at least ten years, the plant budget is up 78% in a year, and the company has just promised ₩90T to ₩110T to owners for one year.²⁵²²⁹ If profit halves, the bonus falls with it, which is the design, but plant already built and promises already made do not.

None of the three needs a surprise, and together they shape my bear and my use of a through-cycle margin. North Korea sits outside the model, as do the US licences that let Samsung ship tools to its Chinese plants; the licence is annual and nothing I found says what happens after 2026.²²

What the price implies

At ₩225,000 the market values Samsung at ₩1,424T, about $1,067B: 5,710M ordinary shares at the common price and 802M preferred shares at ₩174,100, after the treasury shares.¹¹³²⁸ Book equity to owners is about ₩600T, so the price is about 2.4 times book E (estimate). The price is 4.9 times the Naver consensus earnings per share for 2026, ₩45,983, and about 3.3 times the annualized operating profit of the third quarter C (calculated).¹²³ A low multiple at a profit peak is normal for cyclical shares and proves nothing alone. The share fell 43% from its high close of ₩362,500 on 18 June to ₩207,000 on 30 July and is now 9% above that, and it is still up 88% from the last close of 2025, ₩119,900.¹ The market is saying these earnings will not last.

Growth from a peak year says little, so I turn the question round. Start with my base cash through 2029. That is ₩1,136T, or 80% of today’s market value, in a little over three years. Discounted it is ₩922T. Add what I count outside the business and the price has ₩342T left to pay for. That needs ₩37T of owner cash a year from 2030, growing 3%, for good C (calculated). In 2018, the best year before this one, owners could take ₩36.5T. If my next three years are right, the price asks only that the business later match its best year before 2026.

Everything hangs on that “if”, and it is the market’s best argument. Most of that ₩1,136T has not been earned. It rests on selling prices that are several times what they were a year ago. If 2028 is half my number, most of the cushion goes. After 2023 the market has reason to discount cash like this.

Put in margin terms, the price equals my base value if memory operating margin is 18 points lower than my base in every year, which is a boom margin of about 58% in 2028 instead of 76% and a settled margin near 23% instead of 41% C (calculated). For comparison, the company’s operating margin averaged 14.6% over ten years, and the chip division lost money in 2023.

What I can’t judge

Nobody knows what a normal memory margin will be after this. The new multi-year agreements have never met a downturn, and no earlier cycle had AI in it. I also cannot judge how fast Samsung wins back HBM share, whether the foundry ever earns its cost of capital, or what a Taiwan or North Korea event would do. Here is the range, and read my ₩331,000 with it in mind.

The discount rate

I use 11.0%, 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.³⁴³⁵³⁶ My betas for Samsung against KOSPI, on month-end returns, are 1.20 over ten years, 1.23 over five and 1.28 over three C (calculated).³⁷ Those give 9.8%, 10.0% and 10.2%. SK hynix has betas near 1.6 to 1.9 on its page and I use 12.0% there. KOSPI is now heavily Samsung and SK hynix, so these betas partly measure the stock against itself. I take 11.0% because about nine tenths of the profit comes from the same memory cycle that earns 12.0% on that page, and a point less because Samsung has phones, displays and ₩201T of cash behind it. At 12.0% my base would be about ₩308,000, still above the price.

Three cases

I keep the model simple, and I give memory the same cycle as on my SK hynix page, because the two share one price. Samsung’s memory revenue is 1.3 times SK hynix’s in each case, between the ratio of 1.35 in the third quarter and 1.07 for 2025, and its memory margin is 4 points lower, which is the gap between SK hynix’s 79.7% and my estimate of Samsung’s memory margin, about 76%, once the foundry and System LSI losses are taken out of the chip division E (estimate).³ Everything else is a separate line: the foundry and System LSI, which lose ₩2.5T in 2027 and break even in 2029, and the phone, TV, display and Harman businesses, which I put at ₩5T in 2027 and ₩20T by 2031 against ₩18.5T in 2025 E (estimate).⁵ Tax is 24%, working capital 15% of added revenue, and each year’s cash arrives at year end. The fourth quarter of 2026 is my estimate: ₩210T of revenue and ₩116T of operating profit, below the ₩129T that the Naver consensus implies.¹⁷³ I hold the won at ₩1,335.

My 2027 base operating profit is ₩560T, above the two spring forecasts I found, Goldman’s ₩438T and SK Securities’ ₩493T, and my bear’s ₩479T is close to the second.¹⁸ I would not defend ₩560T as likelier than ₩493T. It follows from using the same cycle as my SK hynix base, which sits above its consensus too, and it is the number that most needs your scepticism.

Memory revenue falls 65% from peak to trough in my base, when the worst fall in the past decade was 36% from 2022 to 2023 for the memory line, ₩68.5T to ₩44.1T C (calculated).⁹ Peak prices sit far above cost, so reaching even a 41% margin means a long fall in price. Memory revenue of ₩311T in 2031 is still 3.0 times 2025’s.

Most of my bear value is near-term cash: 61% of its value is cash due through 2029. So I also ran a stress case with no weight: memory revenue of about ₩520T in 2027 and ₩290T in 2028, margins down to 5% and the new agreements renegotiated. It gives ₩132,000, about 41% below the price. I think it unlikely; it is 2023 on a larger scale.

How good, and how good a price

Quality and price are separate questions. The business: the largest of three in an industry that has made and lost fortunes, with a second set of businesses that cushion it, ₩201T of cash and holdings that I count, and the best product it has had in years. It is not a great business in the sense of steady returns, since 52 won in 100 of ten years’ profit reached owners as free cash. The price: on my numbers, this looks like a good business at an unusually low price for what it is earning, as with SK hynix. That describes my numbers, and cheapness on peak earnings is the usual trap in this sector.

Other uses of the same won

The Korean ten-year pays 4.01%. My base here returns 18.5% a year at this price and my bear 8.4%, against 11.0% required E (estimate).³⁴ The closest comparison on this site is SK hynix: on that page the price is 31% below the base value, the base returns 22.6% a year, and the bear 5.5%. Samsung’s base is 18.5%, four points less, and its bear is 8.4%, because it has cash, the phone hedge and a foundry option behind the memory profit. The cycle, the method and the discount to value are nearly the same, so holding both doubles one bet and does not spread it.

The nearest alternative of all is the same company at a lower price. The preferred share closed at ₩174,100, 23% below the ordinary share, with the same third-quarter dividend per share and no vote.¹³² Over the last ten years the discount averaged 17.1%, reached 37.5% on 27 May and was 4.8% in September 2021 C (calculated).¹¹³ At the preferred price my base returns 27.9% a year and my bear 12.2%. Samsung said on 29 October that it is “actively considering” changing the share of preferred stock in its buyback and cancellation programme, which would put cash behind the discount.² Why does the discount exist? The preferred has no vote and trades about a quarter of the ordinary share’s volume, and foreign investors hold about 75% of it against 46% of the ordinary.³⁸ I do not assume the gap closes, and I cannot judge when it might. If it went back to its ten-year average, a preferred holder would gain about 7% on top of everything else; the ordinary holder would not.

What past lows looked like

Four falls of 30% or more in ten years on the ordinary share: 35% from November 2017 to January 2019, 32% in early 2020, 45% from January 2021 to November 2024, and 43% from 18 June to 30 July this year C (calculated).¹ So a 43% fall has already happened this year, in six weeks, with the reported numbers rising. Markets turn early: the 2017 high came a year before the 2018 profit record, which stood until this year.

Book value has been a rough floor before. At the 2019 low the share sold near 1.1 times book and at the 2024 low near 0.9 times E (estimate).¹⁴⁵ It sells near 2.4 times today. If my base is right and nothing were paid out, book would reach about ₩246K a share by the end of 2028 E (estimate). That is above today’s price. That shows how much the market doubts the next two years; it does not forecast a floor.

The margin of safety

• Price₩225,000Close, Q4 2026 [1]
• Base value₩331,000My estimate E (estimate)
• Margin of safety below₩247,000set from my bear case C (calculated)

I set the margin of safety from the bear case, as on my SK hynix page, because one cycle drives this business. Below ₩247,000, 25% under my base value, my ₩185,000 bear value is at least three quarters of the price paid. The price is 9% under that line. I rarely get to write that, so I tested it. It has three weak points. The bear’s value is 61% cash yet to be earned. At ₩1,200 to the dollar the bear falls to about ₩172,000, and the line with it. And the stress case sits well below the price. So the test is passed narrowly, and the protection depends on 2027 arriving about as priced, which is exactly what a cycle can remove.

Where this could be in 5 and 10 years

The table gives my model’s value per share at two later dates. Value in a year is the present value at that date of the owner cash that comes after it, in that year’s won, on today’s share count. It leaves out the ₩30.9K a share of cash and holdings I count today, which I treat as received now, so the return starts from a price net of it.

The base holds at five years if the shortage lasts through 2028 and memory margin then settles near 41%. By October 2031 the cycle has already paid ₩203K a share, and what is left is worth ₩237K. By October 2036 owners have been paid ₩295K, more than today’s price, and the ₩275K left rests on 2036 owner cash of ₩132T growing 3% E (estimate). Three numbers drive it. Memory margin: five points is about ₩29K of base value. Capital spending: ₩15T a year more is about ₩9K. The discount rate: a point is about ₩23K. The bear differs by a bust that never fully recovers, with a memory margin near 26% and owner cash of ₩53T by 2036. The bull is a plateau, with margins near 59% and revenue held near its peak.

The most likely way the base is wrong is a deeper or longer trough in 2029 to 2031 than the cycle I borrowed from my SK hynix page, because the bonus formula and plant already built do not shrink with profit. An error there shows up in the year-10 value, not in the early cash.

The 21.5% here sells at the October 2036 value taken at 11%; the 18.5% in the cases table takes that value at the solved rate, so it is lower. These are estimates of value, not forecasts of the share price, and the market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three facts cut in the company’s favour and stay out of the headlines. First, the company hedges itself. In 2023 the chip division lost ₩14.9T while the phone, TV, display and Harman businesses earned ₩21.1T, so the company still made ₩6.6T.⁹ Today those businesses pay more for memory, but a bust in memory helps them, and my base has them earning ₩20T by 2031 E (estimate). SK hynix has no second leg. Second, money sits outside the cash line. In the first half Samsung moved ₩32.4T into long-term deposits that net cash does not count, and its listed holdings are worth ₩73T at today’s prices, against ₩14.4T carried for its associates.⁸³²²⁶ Third, the preferred share paid the same dividend per share this quarter and trades 23% below the ordinary share, against a ten-year average of 17.1%.²¹¹³

The bull case

Memory has become the limit on AI, Samsung is the biggest maker with the most money to build, and it is catching up in the one product where it fell behind. The company says AI infrastructure spending in 2027 will reach a historic level, that AI-driven demand will far exceed industry supply, that it has scaled up HBM4 and that it is securing multi-year contracts.²³ Counterpoint has its HBM share more than doubling in a year.⁷ Goldman Sachs put 2027 operating profit at ₩438T and 2028 at ₩495T on 8 May, and SK Securities put 2027 at ₩493T.¹⁸ Naver’s mean target price is ₩490,227, 118% above the close.¹² I found no verifiable statement about Samsung from a named long-term investor, so I attribute a view to no one.

Against reality

If Samsung does what it says through 2028, my base has it generating ₩736T of owner cash in 2027 and 2028, 52% of the market value, and ₩1,136T by the end of 2029, 80%. The price is possible if memory prices stay high through 2028, Samsung keeps its share of the DRAM profit, and the new fabs do not turn the shortage into a glut before that cash arrives.

The bull needs four things at once: tightness through 2028, 70%; no glut in 2029 or 2030, 50%; AI spending still rising in 2030, 60%; a foundry that earns its keep, 40%. Together that is 8%. The price needs less: 2027 near my number, 80%; 2028 not a collapse, 65%; a later margin above my bear’s, 75%. That is 39%, which is why the price is not obviously wrong even at 32% under my base value. These odds are my guesses.

What a 10x would take. Multiplying ₩225,000 by ten gives ₩2,250,000 a share, about ₩14,652T or $11.0T of market value. NVIDIA, the largest company I cover, is about $5.0T at $204.64, and it rose 5.1 times in three and a half years from a value near $1T, from $40.11 on 30 May 2023.³⁹ At a profit multiple of 12 and a net margin of 40%, a tenfold move needs revenue of about ₩3,052T a year, 3.9 times the third quarter’s annualized revenue, held for years, after a rise that has already lifted quarterly operating profit nearly nine times in a year. Over ten years it is about 26% a year from here. I do not think it is plausible.

Where I come down

The near-term cash and the money outside the chip business moved me most: they are why my base sits above the price. A downturn that leaves the memory margin above 50%, a foundry that earns its keep and a buyback that cancels shares would raise my values. A price decline the multi-year agreements do not blunt would lower them. I keep the base where it is, and the bear is the figure to hold it against.

VIIIRisks, and what would change my view

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

• Cost ratchets. A ten-year bonus formula, a capital budget up 78% in a year and a one-year promise of ₩90T to ₩110T are all easier to start than to reverse.²⁵²²⁹
• Control and governance. The group controls 19.7% of the votes with a small economic stake, and the chairman faces a gift tax of about ₩1.56T next year that the reports do not say how he will pay.¹⁰ The family’s old inheritance-tax sales are over, and I cannot say what will replace them.²⁴
• Politics and trade. The licence to ship tools to Samsung’s Chinese memory plants is annual.²² I found no settled tariff terms on Korean chips. The foundry depends on a US fab and US customers.
• The foundry. One outside estimate has it losing ₩3.9T in 2026 and Samsung’s share of the market is 5.9%.¹⁹²⁰ A foundry that never earns its cost of capital is a drag, which is why my bear has it losing ₩3T a year throughout.
• Disclosure. Samsung does not report memory profit apart from the foundry and System LSI, so part of what I call memory margin is my estimate.³ The third-quarter statements are not yet reviewed and may change.³
• The won. A stronger won cuts the won value of dollar prices. At ₩1,200 my base is ₩307,000.

Geopolitical risk, and what it means for the value

Samsung’s plants are mostly at home. Of ₩244.8T in long-lived assets at the end of 2025, ₩182.1T were in Korea, ₩33.0T in America and ₩9.9T in China.⁵ C (calculated) Revenue is the reverse: ₩287.0T of ₩333.6T was booked outside Korea. So the exposure is to rules that decide what Samsung can build and where it can sell.

China and the tool licence. The United States ended the special status that let Samsung ship chipmaking tools freely to its Xi’an memory plant on 31 December 2025 and replaced it with a licence for 2026.²² SK hynix says the same of its Chinese plants.⁴² The Commerce Department decides, and I found nothing on 2027. I put non-renewal at about one in five, because both companies were approved for 2026 on 30 December. Samsung does not say how much Xi’an makes. I assume the loss is 8% of memory revenue and the Chinese assets are written off. E (estimate)

Tariffs. Korea’s 2025 deal with Washington set 15% on many goods and left chips open. Seoul says the joint fact sheet promises chip tariffs no less favourable than other major countries get, and TSMC won an exemption after pledging $250B in US investment. The commerce secretary has said memory makers can pay 100% or build in America.⁴¹ The president decides. I put a rate that leaves Korea worse off than Taiwan at about one in four, and I model it as three points off the memory margin. E (estimate)

Taiwan. TSMC makes the accelerators that buy Samsung’s HBM. A blockade or war would stop them, cut memory demand and disrupt shipping, though Samsung’s foundry would gain some work. Beijing, Taipei and Washington decide, and I have no measure to offer. Under one in ten over five years is a judgement. The table’s fall uses my bear’s memory cycle with 30% less memory revenue and the base capital plans unchanged, so it overstates the loss if Samsung cuts spending. E (estimate)

North Korea sits outside my model. The table shows only what a permanently higher premium for holding Korean assets would do.

The price is ₩106,000 below my base. That already covers the China and tariff rows, alone or together. It does not cover Taiwan, and I would not expect any price to.