Samsung's results expose the capacity trade-off behind the memory shortage

Samsung Electronics expects third-quarter sales of approximately KRW195 trillion and operating profit of approximately KRW107.4 trillion, according to preliminary guidance issued on 8 October.

The company will publish its full divisional results on 29 October. The guidance does not yet disclose how much of the increase came from memory, foundry, mobile or other operations, but analysts have connected the earnings rise primarily to stronger prices and demand for AI-related memory.

Reuters reported that Samsung and Micron expect the imbalance between memory supply and demand to persist into 2028. Samsung's high-bandwidth-memory bit shipments were estimated by analysts to have risen by roughly 50 per cent from the previous quarter.

HBM competes for more than packaging capacity

High-bandwidth memory is assembled as vertically stacked DRAM dies connected through through-silicon vias. Its production depends on advanced DRAM fabrication, known-good-die testing, stacking, bonding and specialist packaging capacity.

The logic base die used in newer products adds another advanced-manufacturing requirement. Samsung's HBM4 architecture combines its 1c-generation DRAM with a base die manufactured on a 4nm logic process.

This means HBM demand does not sit entirely apart from conventional memory production. Suppliers must decide how wafer starts, process equipment and investment are divided between HBM and other DRAM products. Advanced packaging and test capacity then place additional constraints on the volume that can be shipped.

TrendForce expects conventional DRAM contract prices to rise by between 10 and 15 per cent quarter on quarter in the fourth quarter of 2026. It forecasts a 15 to 20 per cent rise for NAND flash contract prices, citing AI-server demand and manufacturers' prioritisation of advanced production.

The effect reaches beyond AI accelerators

AI systems consume HBM directly, but their servers also require large quantities of conventional DRAM and enterprise storage. The resulting demand reaches memory categories used in networking, industrial computing and other embedded systems.

Not every part follows the same price cycle. Device density, interface generation, package type, qualification status and supplier allocation can create different availability conditions within the broad DRAM or NAND market.

Samsung's detailed results should provide more information about its memory mix, HBM ramp and foundry performance. Until then, the preliminary figures show the scale of the current upturn but not its complete divisional composition.