Samsung Electronics posted preliminary operating profit of 107.4 trillion won for the third quarter of 2026, disclosed on Oct. 8. The number reads like a company-wide figure. It isn't, quite. Meritz Securities estimated that the memory division alone produced about 109 trillion won in operating profit, split roughly 80.1 trillion won from DRAM and 28.7 trillion won from NAND flash. The rest of Samsung, meaning the foundry, System LSI, and the phone business, was under water by several trillion won.
Volume did almost none of the work. Meritz put Samsung's DRAM bit shipments up 3.4% from the previous quarter and NAND up 4.2%. Average selling prices moved 20.1% for DRAM and 19% for NAND over the same period. Shipments crept. Prices ran.
The part that matters for anyone sourcing electronics components over the next eighteen months sits in the 2027 contract calendar now being negotiated. TrendForce raised its 2027 blended high-bandwidth memory average selling price forecast to 121% above this year's level on Sept. 29, citing supply tightness and a heavier HBM4 mix. Samsung has proposed mid-$4 per gigabit for 2027 HBM4 supply, reported The Korea Herald, which is more than three times the roughly $1.50 per gigabit it charges for HBM3E. Whether those numbers survive negotiation is the question of the next four quarters.
The commodity parts are the surprise
DDR4 is obsolete. It is also up more than fourfold.
DRAMeXchange put the September 2026 contract price for 8-gigabit DDR4 at $26, against $6.3 in September 2025. The same firm tracked 128-gigabit multi-level cell NAND to $30.6 from $3.8, an eightfold move in twelve months. Neither part is new, and neither runs on a node any of the big three would choose to build today.
Allocation explains it. SK hynix, Samsung, and Micron have spent two years pulling wafer starts toward HBM and server DRAM, where the price curve actually rewards the capital. Kim Tae-woo, executive vice president of Samsung's memory business, said wafers assigned to HBM will climb from about 20% of global DRAM capacity today to close to 30% next year. KB Securities goes further on the commodity side, projecting that non-HBM products fall to 59% of global DRAM capacity in 2027, down from 65% this year and 73% in 2025.
The arithmetic is unforgiving. Producing a given number of gigabits as HBM takes roughly three times the wafer output of the same volume in conventional DRAM, because of the smaller dies, the through-silicon vias, the wafer thinning, and the stacking steps. Micron has confirmed the same three-to-one conversion ratio between HBM and DDR5 wafer capacity in its 2026 disclosures. Every HBM line added is three commodity lines gone.
HBM4 resets the reference price
The figures being quoted in 2027 talks are large enough to carry Samsung's entire memory business.
KB Securities expects Samsung's HBM selling prices to rise more than 100% year on year in 2027, and expects HBM4 to account for about 80% of the company's HBM revenue next year, up from roughly 40% this year. On industry estimates, a 12-high, 36-gigabyte HBM4 stack goes from about $600 to around $1,300.
Micron got to that number first. On its fiscal fourth-quarter call on Oct. 1, chief executive Sanjay Mehrotra said the company had completed contracts covering the bulk of its 2027 HBM bit supply, and that next year's HBM prices would be far higher than in 2026. He also said the supply-demand imbalance would run through 2028 and that he could not predict when supply would meet demand, as The Korea Herald reported. Once one of three suppliers books a higher price in a live contract, the other two negotiate off that number rather than under it. One vendor's disclosure becomes a market-wide move.
The competitive map shifted underneath all of this. SK hynix held about 50% of HBM share in the second quarter of 2026, down from 58%, with Micron at roughly 18% and Samsung taking the remainder. Samsung's constraint through the first half was yield. It confirmed approximately 80% HBM4 yield in late August, which is why it is now winning share at exactly the moment the price curve resets upward.
Buyers are already pushing back
Not every customer will pay. The price forecasts tend to skip that part.
Conventional DRAM contract momentum has cooled to roughly 0 to 5% quarter on quarter for the fourth quarter of 2026, after gains of 13 to 18% in the third, according to a review of Q4 DRAM contract negotiations. The reason is mundane. PC and phone makers hit price points that consumers refused to accept. TrendForce, meanwhile, expects conventional DRAM contract prices to rise another 10 to 15% quarter on quarter in Q4, so those two figures sit close together and uncomfortably. The January release settles it.
Buyers are also reshaping their commitments rather than trimming them. AMD chief executive Lisa Su, speaking to reporters in Seoul on Oct. 7, said the company values supply relationships lasting three to five years or more over partnerships of one to three years, noted in Samsung's quarterly coverage. That is a buyer volunteering to lock volume in exchange for certainty, which is the only real bargaining position left when all three suppliers are sold out.
Micron's balance sheet shows where the money went. The company reported a record $54 billion fourth quarter and disclosed roughly $100 billion in hyperscaler contracts for future supply, including about $22 billion in advance commitments. Long-term agreements with prepayment have replaced the quarterly spot booking, and they work for hyperscalers with balance sheets. They work less well for an industrial equipment maker trying to source a 16-gigabyte DDR5 module.
What to watch in January
TrendForce's official fourth-quarter conventional DRAM contract release lands in early January 2027. It will show whether the flat-to-slightly-up guidance for Q4 2026 held, or whether the HBM4 reallocation dragged commodity pricing up behind it.
The second figure is Micron's fiscal first-quarter 2027 gross margin against a guide near 86%. A second consecutive decline would be the first hard evidence in this cycle that pricing power has moved back toward buyers. A re-acceleration would mean the fourth-quarter pause was a breather rather than a turn.
One observation worth stating plainly. Samsung's rivals keep committing tens of billions to new logic capacity that arrives years late, but the 2026 memory squeeze was not caused by a shortage of fabs. It was caused by fabs choosing to build HBM instead of commodity DRAM. That choice takes a spreadsheet to reverse and years of cleanroom time. Until a supplier with spare wafers offers commodity allocations at a sane price, everyone outside the AI accelerator business keeps paying for it.
More coverage from the desk is in our Semiconductors section, with archive access at /archive.