AINEWS
High bandwidth memory (HBM) production is cannibalizing conventional DRAM fab capacity, so the suppliers sitting on the most conventional capacity, Samsung and Micron, are collecting the bigger windfall this quarter.
The memory chip cycle has a counterintuitive leader, and it is not the company with the most high-bandwidth memory share. Samsung Electronics and Micron are collecting the bigger windfall from the current DRAM price spike, because the very boom that built SK Hynix into the dominant HBM supplier, the stacked memory chips that ride into NVIDIA's AI accelerators, has starved its own conventional DRAM business. The data behind that divergence landed in the same week in early August.
In Q2 2026, conventional DRAM contract prices rose quarter over quarter while HBM prices fell year over year, according to Counterpoint Research via GamesBeat. The conventional-DRAM move happened in the same quarter when HBM3E pricing dropped and HBM4 launches slipped, a Counterpoint read that trade coverage of the same window confirms. Conventional memory is back to being a seller's market. The AI-grade memory that drives accelerator stacks is not.
The mechanism is fab capacity. The DRAM and HBM lines at Samsung, SK Hynix, and Micron all use the same 300-millimeter-class fabs. Every HBM die is built by stacking DRAM wafers, and the equipment, clean room space, and process engineering that go into an HBM line are taken out of the line that would otherwise be feeding conventional PC and smartphone modules. DigiTimes, Counterpoint's read, and a secondary explainer on the 2026 shortage all point to the same constraint: AI training clusters and the HBM lines that feed them are bidding the same wafers up against the rest of the memory market.
That allocation is the variable that explains the current ranking. SK Hynix is widely estimated to hold around 50 to 55 percent of the HBM market, with Samsung at roughly 35 to 40 percent and Micron at 5 to 10 percent. SK Hynix was the first to mass-produce HBM3E and remains the lead supplier into NVIDIA. By that scoreboard, it is the AI memory winner. By this quarter's contract pricing, it is not. DigiTimes reported on August 5 that Samsung and Micron are benefiting more strongly from rising conventional DRAM and NAND prices than SK Hynix, because SK Hynix's HBM focus has slowed its conventional DRAM expansion. The company winning the AI race is paying for it with a thinner conventional book.
The Q3 2026 forecast splits the same way. TrendForce, via BigGo Finance, projected DRAM contract price increases would moderate to 13 to 18 percent in the third quarter, slightly below Samsung's stated 20 percent target, with weak consumer demand cited as the reason. A separate BigGo Finance read on the same window put the Q3 contract-price move at more than 25 percent, fueling expectations for stronger earnings at both Samsung and SK Hynix. Both reads treat the conventional DRAM line as the pricing engine. The difference is the magnitude, and which side of the bid the analyst is on.
The same window also showed a reset on the AI memory side. HBM3E contract pricing dropped in Q2 2026, and HBM4 launches slipped on the schedule that suppliers had been quoting to customers. That is the second beat of the same mechanism: when HBM allocations cool, the fabs that were taken out of conventional DRAM can come back. Whether they do, and how fast, is what the next two earnings cycles will answer.
Downstream cost pressure is the second-order effect the structure predicts. Phones and PCs compete for the same conventional DRAM capacity that AI has crowded. Secondary coverage has circulated 2026 figures of smartphone shipment declines near 13 percent and PC shipment contractions near 11 percent, with DRAM spot quotes up several-fold year over year. Those magnitudes are not yet confirmed at primary sources, but the direction matches what a fab-capacity squeeze predicts. The bill lands on the next buying cycle for whoever is assembling servers, phones, and PCs.
The HBM share table and the conventional DRAM contract table no longer line up, because the variable that drives both is fab allocation. A quarter that is good for HBM can be neutral or worse for the same supplier's conventional book, and a quarter that is good for conventional DRAM is great for the supplier that kept capacity on the conventional line. Samsung and Micron are collecting on the conventional book this cycle. SK Hynix is positioned on the AI side. The next memory cycle will be a different bet.