SK hynix, based in South Korea, makes DRAM, the working memory every processor needs, and HBM, DRAM stacked beside an AI processor to feed it fast. It also makes flash storage.
The company · 2 · What exactly does it sell?
DRAM is over three quarters of sales
Each $1 of sales, 2025
77%21%2%
Working memory and AI memoryFlash storageOther
AI memory is sold inside the DRAM figure, and is the fastest-growing part.
The company · 3 · How does it make money?
Memory sells at a market price, and profit follows it
80 centsof each sales dollar as operating profit, Q3
48.6%the same figure in 2025
Won 7,730 billionoperating loss in 2023, when prices fell
Factories cost the same whether prices are high or low, so most of a price rise becomes profit, and most of a fall becomes loss.
Is it a good business? · 4 · What gives it an edge?
Three firms make DRAM; it leads in AI memory
Almost 90%
of world DRAM from three companies
Over 50%
of the AI memory market
14
customers on long agreements
That keeps new rivals out but does not hold prices up: most of today's profit is the shortage.
Samsung grew its AI memory shipments faster in mid-2026.
Is it a good business? · 6 · What does its record show?
Price rises, not volumes, drive the profit
PriceUp high 20s%DRAM prices, Q3 on the quarter before
VolumeUp about 10%DRAM shipped
Last downturnWon 9,138 billionnet loss in 2023, when prices fell
Prices rose almost three times as fast as shipments, so price did most of the work. In 2023 it ran in reverse.
Is it a good business? · 7 · How much room is there to grow?
AI memory heading from $33B toward $86B
$33B → $86BAI memory sales, 2025 to 2027
$154.7Bworld DRAM sales in one 2026 quarter
Over 50%of its 2027 and 2028 output signed
Those are sales at today's prices. Nobody publishes how much memory the world needs once prices fall back.
Is it a good business? · 8 · Who is in charge?
A factory chief who took AI memory to volume
Since 2022Kwak Noh-jung chief executive
20.5%owned by SK square
10%of operating profit shared with staff
As a foreign company, its officers file no US insider trading reports.
The story right now · 9 · Why does AI need so much memory?
Every AI chip needs a stack of fast memory
Over 50%of AI memory made by SK hynix
14customers on five-year agreements
38%the shares below their high
AI processors work on huge amounts of data at once, and each needs fast memory beside it. That memory eats factory space, so ordinary memory runs short too.
The story right now · 10 · Why does memory profit swing so much?
Shortages lift prices, then new factories sink them
Demand outruns supplyPrices jump.
Profit soarsMakers build factories.
New factories produceSupply catches up.
Prices fallIn 2023, a loss.
Long agreements make this boom longer, but their prices still follow the market. We value SK hynix on a normal year, not today's peak.
The story right now · 11 · What changed lately?
A record Won 83.3 trillion of operating profit
Won 104.6 trillionquarterly sales, up 32%
80 centsof each sales dollar as operating profit
Won 32.6 trillioncurrency and investment losses
38%the shares below their high
Strong quarter
The shortage still lifts profit and the newest AI memory is shipping. The agreements' prices stayed private, and US export licences for its China factories are a worry.
The story right now · 12 · What is everyone arguing about?
Do AI memory and long agreements end the cycle?
Many investors think
A memory boom at its peak draws new supply
We think
AI memory and long agreements make this boom longer and safer
The market cut the shares three times in four months while profit rose. Against us: agreement prices still move with the market, and Samsung is catching up.
The price and our view · 13 · What is a share worth?
The price sits about a tenth above our value
$125what we think a share is worth · the last close is of that
$125our value
Attractive below $99.74
Neutral
Unattractive above $150
$61.33gloomy: growth halves and profit per $1 falls
$125our estimate: prices fall back to a normal year
$180sunny: faster growth, higher profit per $1
Value per share in three futures. None is a prediction of the share price.
We estimate what a share is worth from the cash the business should earn over many years. Our rule: buy only when the price is at least a fifth below our estimate, as a cushion for being wrong.
The price and our view · 14 · How could we be wrong?
The shortage ends, and Samsung returns
The shortage endsBiggest
In 2023 the company lost money.
Samsung in new AI memoryWatch
Could take share in the next AI chips.
China licencesShowing now
Tools for its China factories held up.
Agreement pricesUnknown
What they earn when prices fall.
Its largest customer was nearly a quarter of 2025 sales.
The price and our view · 15 · How will we know?
January shows whether profit keeps rising
Operating profit above Won 83.3 trillion, 75 cents per $1Good sign
DRAM shipped up 20% or moreGood sign
Below 60 cents per $1 as DRAM prices fallWe turn negative
JAN 202728
Fourth-quarter 2026 results, 28 January. Companies report every three months; each report is a test of the story.
Our view also changes on price alone: we turn positive if the shares fall a fifth below our estimate, and negative if they rise a fifth above it.
The price and our view · 16 · So, what do we think?
The AI memory leader, valued on a normal year
$125our value
Attractive below $99.74
Neutral
Unattractive above $150
NeutralMedium conviction
SK hynix leads in AI memory, and fourteen customers signed five-year agreements for over half of 2027 and 2028. But memory sells at a market price, and in 2023 the same business lost money. Neutral.
↑ ↓ to move · C for contents
What it does
More about the company
Who it is
SK hynix makes two kinds of memory chip. DRAM is the fast working memory every processor needs; HBM is DRAM chips stacked into a tower right beside an AI processor, to feed it data fast. NAND flash keeps data with the power off, much of it sold as data-centre drives through Solidigm, its US arm.
DRAM is about three-quarters of sales. The buyers are AI chip makers, cloud companies and phone and PC makers; without SK hynix they would buy from Samsung, Micron or, for ordinary chips, China's CXMT.
How it makes money
SK hynix sells memory by the bit, so sales are the bits shipped times the price per bit. Prices swing far more than volumes: in the third quarter DRAM shipments rose about a tenth while DRAM prices rose by over a quarter. Almost all of a price rise falls to profit, because the factories cost the same either way.
HBM is sold on yearly volume and price deals agreed in advance; ordinary chips mostly sell at the going price.
Growth and profit
What growth costs
Growth made SK hynix far more profitable, because research and overheads barely moved while prices soared, so they shrank to a sliver of sales. What does grow with profit is the factory bill and staff profit sharing. The lever works in reverse: in 2023, when prices fell, the company lost money.
Why it wins
The evidence
Memory is hard to make at scale: three companies make almost 90% of the world's DRAM, and each spends tens of trillions of won a year on factories. SK hynix leads the hardest part, HBM, with over half of the market, after being first to stack chips this way and developing the bonding method it uses.
That keeps new rivals out but does not hold prices up. Most of today's margin is the shortage, which is why customers signed long agreements to be sure of supply.
Who could beat it
The rivals in full
Samsung, the biggest DRAM maker, is shipping its own HBM4 and grew its shipments faster in mid-2026. Micron sells the same three products to the same buyers. And China's CXMT has grown to about a tenth of world DRAM sales in ordinary chips.
Each rival
Samsung Electronics HBM4005930.KS
the AI chip maker takes more of its next-generation HBM from Samsung instead of SK hynix
the GPU maker or cloud buyer moves part of its HBM or server memory volume to Micron
CXMT (ChangXin Memory Technologies) DDR5 and LPDDR5688469.SS
a phone, PC or Chinese server maker buys ordinary DRAM from CXMT, mostly in China today
Samsung, Kioxia, Micron and Sandisk data center SSDs005930.KS285A.TMUSNDK
the data center buyer fills its flash storage from another flash maker
How big it could get
The market and who has signed
The market
World DRAM sales reached $154.7B in a single quarter of 2026, and the top five flash makers sold $68.9B. Gartner, quoted by the company, expects the memory market to nearly triple in 2026, with HBM growing from $33B in 2025 toward $86B in 2027, and TrendForce sees contract prices still rising this quarter.
Those are sales at today's prices. Nobody publishes how much memory the world will need once prices fall back, so the market beyond 2027 is unmeasured.
Who has signed
Fourteen customers, including large cloud companies, have signed long-term agreements, usually for five years, covering over half of 2027 and 2028 output, management says. HBM volumes and prices for 2027 are agreed. The terms are not filed, and prices inside the agreements still move with the market within a range.
Who runs it
Leaders, pay and the cash
Who runs it
Kwak Noh-jung has been chief executive since 2022, after running manufacturing and technology, and took HBM3E and HBM4 to volume. SK square, controlled by the SK Group, owns about a fifth of the shares. Staff share 10% of operating profit, so pay rises and falls with it. As a foreign company, its officers file no US insider trading reports.
What they did with the money
SK hynix plans about Won 100 trillion of returns for 2026 in dividends and buybacks, with part of the buyback paying staff in shares. In July it sold 2.5% more shares as US-listed ADSs at $149 each.
The last report
The latest report
SK hynix's third-quarter operating profit rose 38% to a record Won 83.3 trillion
What it did to the thesis Strong quarter
It showed the shortage still widening margins and HBM4 shipping, which supports the view. It did not show the profit lasting: the agreements' prices stayed private, and net profit fell on currency and investment losses outside the business.
The report
In the third quarter of 2026 sales rose 32% from the quarter before to Won 104.6 trillion and operating profit 38% to Won 83.3 trillion, 80 cents of each sales dollar. Prices did most of the work: DRAM prices rose by over a quarter, shipments by about a tenth, and HBM4 began shipping.
Net profit fell to Won 39.8 trillion, because a weaker won and a markdown of investments it holds cost Won 32.6 trillion. That is why earnings per share missed forecasts by a third while sales beat.
What's happening now
The background
The shares are 38% below their high and under the $149 they were sold at in July. It is the third fall in four months, the latest after a record quarter, on worries about how long the boom lasts and about US export licences for its China factories.
The next test is fourth-quarter results on 28 January 2027, which will show whether profit keeps rising and shipments grow as guided.
Where we disagree
Both sides in full
The market
A memory boom at its peak: profits this high always draw new supply, so pay little for each dollar of today's profit.
We believe
HBM and the long agreements make this boom longer and safer than past ones, and the price needs far less than today's profit.
The argument
The market has cut the shares three times in four months while each quarter's profit rose, pricing the length of the cycle rather than the quarter. We agree the cycle matters, but HBM takes far more factory space per bit, only three firms make it, and over half of the next two years is signed.
Against us: the agreements' prices are not public and still move with the market, Samsung is catching up in HBM4, and in 2023 the same factories lost money.
What it's worth
How we value it
Valued on the analysts' revenue path for three years, cut by the price fall half its margin drop to a normal year implies because memory sells at a market price, each later year closing 30% of the gap to 4%, on official operating profit with stock pay a cost and tax paid in full.
What the price assumes
Sales growth in the first year, then slowing as ours does, at the same profit per sale.
The price assumed
9.9%
Our estimate
26%
Measured at the close of
6 Nov 2026
The three cases
Bear case$61.33
Growth runs at half the analysts' path, less the same price fall, 1%, minus 12% and minus 11%, the margin settles a fifth lower at 16.7%, and new capital earns only the 8% rate: the downturn test.
Our value$125
Revenue follows the analysts' path (C46), 60%, 28% and 29%, less the price fall half the margin drop from 74.5% to 20.9% implies: sales 49% of the path by year three, so 26%, 0% and 2%; then it fades toward 4%, closing 30% of the gap each year; the margin settles at its five-year average of 20.9%, not today's cycle peak.
Bull case$180
Growth runs a quarter above the analysts' path, less the price fall, 39%, 7% and 8%, the margin reaches 23.0%, and new capital earns 15%.
What could go wrong
Every risk
The risks
The shortage ends
Profits this high draw new factories; when supply caught up in 2023, the company lost money.
Agreements without public terms
Prices inside the agreements move with the market, and the prospectus says buyers can cut orders.
Samsung in HBM4
The biggest DRAM maker is shipping HBM4 and could take share in the next AI chips.
China
CXMT is growing in ordinary DRAM, and US licences for tools bound for SK hynix's own China factories are reported held up.
One big customer
Its largest customer was nearly a quarter of 2025 sales.
What we still do not know
The price ranges in the long-term agreements, and what margin they earn when market prices fall.
Whether SK hynix keeps over half of HBM once Samsung's HBM4 ships at volume.
How much of the third quarter's investment loss was the Kioxia stake, and whether more follows.
What changes our view
Every route to a new view
Every route to a new view
→ Attractive
The share price falls to about $99.74, 20% below what we think the business is worth.
→ Unattractive
The share price rises to about $150, 20% above what we think the business is worth.
→ Unattractive
Operating margin falls below 60% while DRAM contract prices fall. A customer cuts, defers or walks away from a volume in a long-term agreement.
The next test
It supports the case if
Operating profit above Won 83.3 trillion, margin of 75% or more, DRAM bits up 20% or more and the agreements still over half of 2027 capacity.
It pressures the case if
Operating profit below Won 83.3 trillion, margin below 75%, or DRAM bits up less than 20%.
We would also be wrong if
An outside count puts SK hynix below 40% of HBM sales in a 2027 quarter.
US export licences block new tools for its China factories for a year.
The verdict
Scores and how we weighed it
The scores
Growth
3 of 4
Business quality
3 of 4
Financial position
4 of 4
Execution record
3 of 4
Per share capture
3 of 4
Potential
3 of 4
Risk profile
3 of 4
Valuation
3 of 4
Evidence quality
3 of 4
How we weighed it
ValueMixed
The price is about at our normal-year value, with the bear case far below.
BusinessPositive
The leader in high-bandwidth memory with long agreements signed.
GrowthMixed
Profit is a shortage peak, and the value assumes it falls back.
RisksNegative
The shortage ends, the agreements' prices are private, a bigger rival ships the newest memory and China grows.
RivalsMixed
There is no usable filed cohort; its own history mixes currencies.
The groupNegative
Memory names move together on the shortage; its profit is the group's cycle.
Latest newsPositive
The latest report showed the shortage still widening margins and the newest memory shipping.
In plain words
HBM
High-bandwidth memory: memory chips stacked on top of each other and placed right beside an AI processor, so data moves between them fast. It takes far more factory space per bit than ordinary memory, which tightens supply for everything else.
In plain words
Our estimate of value
We forecast how much cash the business can earn over the next ten years and beyond, then ask what that stream is worth today to an investor who wants an 8% yearly return. We do it three times: a gloomy future, an expected one and a sunny one. We call a share Attractive only when its price is at least 20% below the expected estimate, and Unattractive when it is 20% above.