HBM turns AI demand into a broader DRAM supply squeeze

By Nathan Brooks

SK Hynix reported third-quarter operating profit of around 83 trillion won, roughly seven times the level of a year earlier, on an operating margin of around 80%. Its headline results cite higher DRAM and NAND prices, but high-bandwidth memory, or HBM, drove much of the surge. The arithmetic is exceptional, even if the market prefers tidy explanations.

What does HBM actually do?

HBM is a form of DRAM, the working memory that holds data while a processor uses it. Conventional DRAM sits on separate modules some distance from the chip; HBM stacks a dozen or more thin DRAM dies vertically and places the stack beside the processor. Thousands of tiny vertical links create a wide, short data path, delivering more bandwidth per second with less power.

That design addresses a stubborn accelerator problem: AI chips often calculate faster than their memory can feed them. Training and running large models requires enormous volumes of data to move into and out of memory, and insufficient bandwidth leaves an expensive processor idle. Every leading AI accelerator, including Nvidia GPUs, uses HBM, which has made it one of the most valuable components in an AI server.

Why are conventional DRAM prices rising?

Industry estimates put HBM at about three times the wafer capacity of standard DRAM per bit of output. Its larger dies, complex stacking process and lost output mean makers must devote substantially more production to each delivered bit. When production lines shift toward HBM, less capacity remains for conventional DRAM used in PCs, phones and servers.

SK Hynix reported DRAM selling prices rising at a high-20% rate in a single quarter. AI demand therefore lifts prices across the memory market rather than only at the high end, supporting earnings at SK Hynix, Samsung Electronics and Micron. It also weighs on margins at the PC, phone and server companies that buy those memory components.

Who controls the HBM market?

Only SK Hynix, Samsung Electronics and Micron manufacture HBM at scale. SK Hynix built an early lead, but the field has tightened. Counterpoint Research estimated SK Hynix at around 50% of HBM revenue in Q2 2026, down from around 64% a year earlier.

Samsung held around 33% and Micron around 18% in that estimate. Samsung’s share rose as it caught up on the previous HBM3E generation. The concentrated market also makes the Kospi and the wider Korean market unusually sensitive to AI spending, while changes in share between SK Hynix and Samsung can move both stocks sharply.

What does HBM4 change?

HBM4 doubles the width of the connection to the processor and is being built into Nvidia’s Vera Rubin platform. Nvidia chief executive Jensen Huang said in June: “All three suppliers are in production for Vera Rubin,” with first systems due to ship in the third quarter. Analysts estimate SK Hynix supplies the majority of that HBM4 volume, although Nvidia has not disclosed a split.

HBM4 broadens Nvidia’s supply choices without making the market less concentrated. All three suppliers remain in the race, and both production volume and pricing will test whether SK Hynix can retain the commercial advantage implied by its early lead. That is a scheduling question, not a victory lap.

Where could the memory cycle break?

HBM helps explain why the three suppliers are earning record profits and returning cash to shareholders. It also exposes the sector’s familiar cyclical weakness. Some forecasters warn that HBM prices could soften once new capacity and competition catch up with demand.

The main risk is that added HBM4 capacity and a three-way contest among SK Hynix, Samsung and Micron arrive together. If they do, pricing power could fade quickly, as it has in previous memory cycles. For now, HBM4 volumes and pricing, the pace of Samsung’s and Micron’s share gains, and the growth rate of AI spending will determine whether this unusually strong upcycle extends or turns.