How HBM became the AI memory behind SK Hynix's profit jump and rising DRAM prices

The strength of the memory upcycle owes much to HBM. Because the technology soaks up production capacity, AI-driven demand raises prices across the whole DRAM market, not only at the premium end; that bolsters earnings for SK Hynix, Samsung and Micron while squeezing margins for PC, phone and server makers who buy memory. With supply concentrated among a few players, the Kospi and the broader Korean market are especially exposed to AI investment, and shifts in HBM market share between SK Hynix and Samsung can push both stocks sharply. The big danger is the cycle reversing: pricing power could evaporate quickly, as in earlier memory cycles, if fresh capacity and three-way rivalry in HBM4 overtake demand.

---

Earlier came a striking earnings report:

• SK Hynix posts 83tn won Q3 operating profit and 100tn won shareholder payout plan

--- HBM is the memory that supplies data to AI chips, and its heavy appetite for production capacity is raising prices for almost every memory chip worldwide.

Key points:

• Thin DRAM dies are stacked vertically in HBM and set beside the processor, giving much higher bandwidth and lower power consumption.
• AI chips frequently hit limits set by data-feeding speed, which is why HBM is used in every major AI accelerator, Nvidia’s GPUs among them.
• By industry estimates, HBM consumes about three times the wafer capacity per bit that standard DRAM does, so a production shift toward it tightens ordinary-memory supply and pushes prices up.
• Just three companies—SK Hynix, Samsung and Micron—manufacture HBM in volume. Counterpoint places SK Hynix’s HBM revenue share near 50% in Q2 of 2026, versus about 64% a year earlier, and Samsung’s near 33%.
• All three are producing HBM4 for Nvidia’s Vera Rubin. Analysts believe most of that supply comes from SK Hynix, although Nvidia has not revealed the split.
• New capacity, competition and a possible easing in HBM prices after supply catches up are the risks to monitor.

SK Hynix’s most recent earnings supplied a remarkable figure for the AI boom: operating profit near 83 trillion won in the third quarter, about seven times the year-earlier level, with an operating margin close to 80%. The company’s top-line numbers point to DRAM and NAND prices, yet the driver behind much of that jump is high-bandwidth memory, or HBM. What follows is an explanation of what that memory is and why it matters for markets.

HBM belongs to the DRAM family, the working memory in which data is held while a processor works on it. Conventional DRAM is located on separate modules positioned away from the chip it serves. HBM, by contrast, piles a dozen or more thin memory dies into a stack and puts that stack next to the processor, with thousands of tiny vertical links making the connection. That creates a wide, short route for data, yielding far more bandwidth—more data per second—while drawing less power.

This is significant because the constraint on AI chips is usually the speed at which they receive data rather than their calculation speed. Large AI models require huge volumes of information to be moved into and out of memory while they are trained and run, and an expensive processor stands idle if bandwidth is insufficient. That is why HBM comes packaged with every leading AI accelerator, Nvidia’s GPUs included, and why the memory is now among the most valuable parts of an AI server.

HBM is also behind the climb in prices of standard memory. Producing HBM consumes much more wafer capacity per unit of memory than normal DRAM does—industry estimates often say roughly three times as much—because the dies are bigger, stacking is complicated and some output is wasted in the process. When manufacturers move production lines toward HBM in response to AI demand, the conventional DRAM used in PCs, phones and servers has less capacity left. That tightness is part of the reason SK Hynix recorded DRAM selling prices rising at a high-20% pace over one quarter.

HBM is produced at scale by just three companies: SK Hynix, Samsung Electronics and Micron. SK Hynix gained an early edge and continues to hold the top share, though competition is growing. Counterpoint Research data show SK Hynix’s HBM revenue share dropping to around 50% in Q2 2026 from close to 64% a year earlier; Samsung’s climbed to about 33% as it closed the gap on the prior generation, HBM3E. Micron’s share was roughly 18%.

HBM4 is the next contest: it doubles the connection width to the processor and is being integrated into Nvidia’s new Vera Rubin platform. Jensen Huang, Nvidia’s chief executive, said in June that all three suppliers have begun production for Vera Rubin, and the first systems are scheduled to ship in the third quarter. SK Hynix is estimated by analysts to supply most of the HBM4 volume, but Nvidia has not released a split.

For investors, HBM is central to several major uncertainties. It is a main reason memory makers are posting record profits and handing cash back to shareholders, yet the business remains cyclical. Some forecasters have cautioned that HBM pricing could ease after competition and fresh capacity catch up with demand. The key indicators are HBM4 volumes and prices, the pace at which Samsung and Micron close in, and whether AI spending expands quickly enough to soak up additional supply.