Deep Dive — Samsung sold out HBM for 2027 and is spending $63B

Samsung Electronics just posted the largest quarterly operating profit any technology company has recorded: 107.4 trillion won for July to September, up 783% from a year earlier, on revenue of 195.3 trillion won. It is the first company in the sector to clear 100 trillion won in a single quarter, according to its own earnings deck.

The more consequential sentence came later on the call, from memory executive vice-president Jaejune Kim: Samsung's 2027 high-bandwidth memory supply is already sold out, at agreed target prices that reflect "our product competitiveness," while customers keep asking for more than the contracts cover.

Record profit plus sold-out forward supply plus a stock that fell as much as 6.6% in Seoul is the entire memory cycle in one morning. The profit is the past. The sold-out 2027 is the next two years. The share price is investors asking what happens in 2028, when the capacity being funded by these exact profits starts producing bits.

What 106.5 trillion won of chip profit looks like

Almost all of the money came from one division. Device Solutions — Samsung's chips arm — earned 106.5 trillion won of the group's 107.4 trillion won operating profit, roughly 99% of it, on 147.1 trillion won of sales. A year ago that division earned 7 trillion won on 33.1 trillion won, a fifteen-fold jump in profit on a 4.4-fold rise in sales.

The rest of Samsung shrank under the weight of its own success. Mobile and networks swung from a 3.6 trillion won profit a year earlier to a 0.6 trillion won loss, because the memory inside a Galaxy phone got dramatically more expensive. Visual display, home appliances and network gear all lost money or barely broke even. Samsung Display earned 1.1 trillion won; Harman 0.6 trillion won. The conglomerate that sells phones, televisions and washing machines is now a memory company with some side businesses attached, and those side businesses are being squeezed by the component prices the memory company is charging.

Prices did the work. Samsung's DRAM average selling price rose about 20% quarter over quarter and NAND rose by a high-teens percentage, the company said on the call, with increases across every application and servers leading. Bit shipments grew only in the low-to-mid single digits for DRAM and mid-single digits for NAND — nearly all of the revenue growth came from price, not volume. That is what a shortage looks like from the inside.

HBM4 is the product that matters — the fourth generation of the stacked memory that sits next to an AI accelerator and feeds it bandwidth no conventional DRAM can reach, which our explainer on what HBM actually is covers in plain terms. Sales of HBM4 more than tripled from the second quarter and it now accounts for about 60% of Samsung's total HBM revenue, the company said, consistent with the target it gave in July. By its own accounting Samsung has already pulled its HBM market share level with its overall DRAM share — a milestone it had targeted for the second half of the year. For two years the knock on Samsung was that SK Hynix owned HBM and Samsung was qualifying late. That gap is closed on Samsung's numbers.

Sold out is a contract, not a promise

"2027 is sold out" sounds like a demand statement. It is really a pricing statement, and a supply one.

Samsung has now concluded multi-year supply contracts with 10 major accounts and is negotiating with 10 more, several in the final stage, Kim said on the call. The customers have broadened beyond hyperscalers and server OEMs to include mobile, automotive and networking buyers. The company's stated goal is to have multi-year contracts cover 60% to 70% of its mid- and long-term capacity by 2030.

Locking volume two years out is exactly what buyers do when they expect to be short. It is also what sellers do when they expect the spot market to get worse: a fixed target price protects both sides from a spike, which means Samsung gave up the upside it would have captured by selling 2027 output at whatever the market paid. Anyone modelling memory earnings as a pure spot-price derivative should note that the largest supplier has hedged a big share of next year.

Why more capex will not produce more bits

Samsung raised 2026 capital spending to 84.4 trillion won, about $63 billion, an increase of roughly 60% on the year, with 76.8 trillion won of it going to the chip division, as Bloomberg reported. Purchases of property and equipment already hit 26.9 trillion won in the third quarter alone, against 10.8 trillion won a year earlier.

The interesting part is Samsung's own explanation of why that money will not loosen the shortage soon. On the call, Kim argued that much of the 2027 capex increase goes into buying cleanroom space rather than tools, that construction costs per unit of space are higher than in past cycles, and that equipment prices and lead times are stretched by the same shortage everyone else is trying to buy their way out of. A fab takes years from groundbreaking to wafers. By Samsung's reading, capex added in 2027 will not move bit output much in 2028.

Then there is the mix shift working against supply. HBM4E and HBM5 dies are larger than the parts they replace, and they need significantly more metal process layers, so every wafer converted to HBM yields fewer bits of conventional DRAM than before. The same wafer capacity that used to supply laptop memory now supplies a fraction of the HBM that AI accelerators want. Samsung expects tight supply to persist into 2028 on that basis, and Micron has said the same.

The price data agrees that the shortage is real but that the rate of increase is slowing. TrendForce expects conventional DRAM contract prices to rise 10% to 15% in the fourth quarter and NAND 15% to 20%, per its September forecast — respectable numbers that look tame next to the roughly 60% jump in the second quarter. Analysts polled by Reuters expect Samsung's fourth-quarter profit to grow 8.2% sequentially, down from 20% in the third. The cycle has not turned. It has stopped accelerating.

The part the market disliked

Both headline numbers came in a hair under consensus: revenue of 195.25 trillion won against 195.79 trillion expected, operating profit of 107.44 trillion against 107.66 trillion, CNBC reported. The chip division's 106.5 trillion won missed the 108.5 trillion won analysts had modelled. That is a rounding error on a record quarter, and it was enough to knock the stock down more than 5%.

What actually hurt was the payout. Of an estimated 90 trillion to 110 trillion won in total 2026 shareholder returns, 35 trillion won is already paid or committed, leaving 55 trillion to 75 trillion won — and Samsung will decide what to do with it at a board meeting in late January 2027, alongside its next three-year return policy. For years the standing complaint about Samsung stock has been that it hoards cash and trades below book. A quarter that produces 107 trillion won of profit and defers the question of who gets it, while raising capex by 60%, is the old Samsung showing up in the middle of the new one.

There is a second, quieter signal. Samsung's shares are still more than 25% below their June record despite four consecutive record quarters, Reuters noted after the preliminary guidance — the first of those four we covered when Samsung projected a 100 trillion won quarter on AI memory demand. The market is not disputing the shortage. It is disputing the durability of a business whose profits quadrupled on price rather than volume, and asking how much of the coming capex wave arrives just as the shortage ends.

What this means if you are not buying memory

Three groups pay for this quarter.

Phone and PC makers do, directly. TrendForce expects the share of industry capacity available for PC memory to fall again in 2027; notebook makers can cut configurations, eat the cost or raise prices, and Samsung's own mobile division has chosen the third option — its component bill turned a 3.6 trillion won profit into a 0.6 trillion won loss in a single year. If you are wondering why a flagship phone costs more this year, this earnings report is the answer.

Cloud providers do, at the margin. Compute is memory-hungry: every accelerator needs HBM stacks, every inference server needs high-capacity SSDs, and enterprise SSD bit demand is forecast to grow more than 80% in 2026 as inference replaces training as the dominant workload. If HBM and server DRAM are contracted years ahead at fixed prices, the cost of a token stops falling as fast as the model efficiency gains suggest. We have already seen the downstream effects in hardware pricing — the memory squeeze has US NAND makers weighing new fabs, and the shortage is now the binding constraint on AI buildouts rather than GPU allocation.

And buyers of anything with a DRAM chip in it do, eventually. Consumers see it last and feel it longest, because contract prices negotiated this quarter land in retail shelves two or three quarters later.

The contrarian case

The bear argument is not that AI demand is fake. It is that memory has punished every generation of investors who confused a shortage with a permanent condition.

Start with what sold-out 2027 actually protects. If prices are fixed with customers for next year, Samsung's 2027 memory revenue is effectively capped on the upside while its cost base is not: 84.4 trillion won of capex lands as depreciation whether or not the AI buildout keeps its pace. Fixed price, rising fixed cost is the shape of a business that looks brilliant in the up-cycle and terrible in the down one.

Then note where the marginal supply comes from. Samsung is spending 76.8 trillion won on chips this year and has said outright that it is securing cleanrooms ahead of demand so it can install tools later if the market weakens. That is a company positioning to flood the market quickly if it decides the peak has passed. SK Hynix and Micron are making the same calculations. Chinese DRAM capacity is growing too, mostly in the commodity segments where Samsung's mobile customers buy.

Finally, the demand side is now one customer group's capex line. Alphabet spent 51.36 billion dollars on capital projects in the third quarter and posted negative free cash flow of 9.07 billion dollars doing it; Microsoft spent 45.2 billion dollars and came in below its own plan. Memory's record profits are a direct transfer from those budgets. If either of them blinks on 2027 spending — and both have now been punished by investors for spending too much rather than too little — "sold out" becomes "over-committed" in one earnings cycle.

What to watch

Three dates and one number. The January board meeting in Seoul settles the shareholder return policy and the 2027 capital plan in the same session — that is when we learn whether the windfall goes to buybacks or to fabs. HBM4E mass production is slated for the first half of 2027 with volume ramp in the second half, the point at which Samsung either holds its recovered HBM share or gives it back. The first fab at Taylor, Texas has begun cleanroom work and equipment installation, and Samsung expects 2-nanometer orders to more than double this year, with high-performance computing passing half of foundry revenue next year.

The number is the fourth-quarter DRAM contract price. Samsung guided to mid-to-high-teens bit growth for DRAM in the fourth quarter, and TrendForce sees a 10% to 15% price rise. If price growth falls below bit growth, the shortage is ending and the 84.4 trillion won starts working against the company that spent it. Until then, this is what the top of a cycle looks like: everyone sold out, nobody able to add capacity fast enough, and the supplier's own shareholders asking for the cash before it goes into concrete.

How much of a memory bill should a chipmaker be allowed to fix two years ahead — and does locking 2027 prices help buyers or just hide the next spike? Write to us via our contact page.