Returns too indiscriminate for a normal re-rating
The returns the memory companies this wiki tracks (its nodes) posted in 2025-26 are hard to read as a normal industry re-rating. Kioxia rose +1,082% in 2025, Nanya Technology +985%, and SanDisk +884% in 2026. The majors moved in step: SK hynix +361% in 2025 and +195% in 2026, Micron +240% then +326%, and Kioxia again +332% in 2026. NAND or DRAM, bleeding-edge HBM or legacy DDR4, it did not discriminate. Even Samsung Electronics, whose business is far more than memory, was dragged up +213% in 2025.
The AI-chain rise began at the chips and spread layer by layer (that order is confirmed in Did the AI cycle propagate along the value chain?), and the layer where the propagation ran most extreme is memory. The re-rating was far more violent in memory than in the glamorous logic chips.
What rose was not a company but an asset class
That indiscriminateness is the key clue: something acted on the memory asset class as a whole, beyond any one company's results. Look at what came just before and the character sharpens. Nanya was −56% in 2024 before swinging to +985% the next year; Micron sat at −46% in 2022 and Samsung at −26% in 2024 before they leapt. The drop from trough to peak was that large.
The cause starts at a single point and balloons outward. A single AI accelerator demands tens of gigabytes of HBM (high-bandwidth memory). HBM is built by stacking DRAM dies vertically, so for the same bits stored it eats far more wafer area and yield loss than ordinary DRAM. That demand soaks up the limited wafer capacity of the three makers (Samsung Electronics, SK hynix, Micron) disproportionately, and because DRAM supply is effectively their oligopoly, there is no way around them.
The bottleneck sits in HBM, but the price rose across the whole stack
As the three redirected wafers to HBM and high-density DDR5, supply of ordinary DDR4, LPDDR4 and NAND shrank. The absorption shows up in the books: at SK hynix a single largest customer (NVIDIA) is about 27% of revenue (roughly ₩11 trillion), and at Micron one cloud customer is 17%, so demand tilted hard toward AI.
Less supply means higher prices. So even Nanya Technology, which makes no HBM at all, and GigaDevice, a niche-memory specialist in NOR and the like (+148% in 2025, +145% in 2026), benefited from a shortage of products they do not even build. Nanya shows this absorption most clearly: it went from −56% in 2024 to +985% in 2025 without selling a single HBM stack, rising nearly tenfold on the legacy-DRAM price left behind in HBM's wake. The bottleneck sits at a single point in HBM, yet the price rose across the entire memory stack, and that indiscriminateness proves where the bottleneck is. The same pressure showed up in China's parallel ecosystem, where niche and legacy demand runs large.
The shortage moves into adjacent chips and tools
The shortage did not stop at the memory chip. First the adjacent chips that drive and control memory rose with it. Montage, which makes the RCD interface chip that conditions DDR5 module signals, grew revenue +50% (¥5.46B) and its stock +169% in 2025, while Phison, the top NAND-controller vendor with server and AI over 70% of its mix, jumped +432%.
Then came the equipment. The capex meant to close the gap flowed on to etch and deposition tools: memory-heavy Lam Research rose +140% in 2025, Kokusai Electric, world number one in batch ALD at ~70% share, +165%, and Wonik IPS, over half of whose sales go to Samsung's chip arm, +409%. A single point of HBM demand spread through the whole memory chain: chip, interface, controller, equipment.
Verdict
A single bottleneck at HBM claimed the bulk of the wafers of Samsung Electronics, SK hynix and Micron, and that supply squeeze re-rated the entire memory asset class, including companies that make no HBM at all. What rose was the asset class rather than any particular product. But memory is, among the many clocks, the most cyclical asset of all, so behind triple-digit gains sits triple-digit give-back risk. The very nodes that just ran triple digits were, a year or two earlier, posting down years: Nanya −56%, Micron −46%.
Fill memory, and the bottleneck simply moves. As wafers expand and DRAM and NAND supply normalizes, the next constraint moves to the space between chip and board: the substrate and advanced packaging that stack the HBM and tie it to logic. The bottleneck moves from silicon to package follows that migration.