HBM explained: the AI memory behind SK Hynix’s record profits and rising chip prices

HBM helps explain why the memory upcycle has been so strong. Because it absorbs production capacity, AI demand lifts prices across the whole DRAM market, not just at the high end, which supports earnings at SK Hynix, Samsung and Micron and weighs on margins for PC, phone and server makers that buy memory. The concentration of supply makes the Kospi and the wider Korean market unusually sensitive to AI spending, and changes in HBM market share between SK Hynix and Samsung can move both stocks sharply. The main risk is the cycle turning: if new capacity and three-way competition in HBM4 catch up with demand, pricing power could fade quickly, as it has in past memory cycles.---Earlier, stunning results:SK Hynix Q3 operating profit beats at 83tn won, unveils 100tn won shareholder payout plan--- HBM is the memory that keeps AI chips fed with data, and because it eats so much production capacity, it is pushing up prices for nearly every memory chip in the world.Summary:HBM stacks thin DRAM dies vertically and places them right next to the processor, delivering much higher bandwidth with less power.AI chips are often limited by how fast they can be fed data, so every leading AI accelerator, including Nvidia’s GPUs, uses HBM.HBM uses around three times the wafer capacity per bit of standard DRAM, by industry estimates, so shifting production to it squeezes supply of ordinary memory and lifts prices.Only SK Hynix, Samsung and Micron make it at scale. Counterpoint puts SK Hynix at around 50% of HBM revenue in Q2 2026, down from around 64% a year earlier, with Samsung at around 33%.All three are in production of HBM4 for Nvidia’s Vera Rubin. Analysts estimate SK Hynix supplies the majority, though Nvidia has not disclosed a split.The risks to watch are new capacity, competition and a potential softening in HBM prices once supply catches up. SK Hynix’s latest results put a striking number on the AI boom: 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%. The company’s headline figures talk about DRAM and NAND prices, but the product behind much of that surge is high-bandwidth memory, or HBM. Here is what it is and why it matters for markets.HBM is a type of DRAM, the working memory that holds data while a processor uses it. Ordinary DRAM sits on separate modules some distance from the chip it serves. HBM instead stacks a dozen or more thin memory dies on top of each other and places the stack right beside the processor, connected by thousands of tiny vertical links. The result is a very wide, very short path for data, which delivers far more bandwidth, meaning more data moved per second, while using less power.That matters because AI chips are often limited less by how fast they can calculate than by how quickly they can be fed data. Training and running large AI models means shuttling enormous amounts of information in and out of memory, and without enough bandwidth, an expensive processor sits idle. That is why every leading AI accelerator, including Nvidia’s GPUs, is packaged with HBM, and why the memory has become one of the most valuable components in an AI server.HBM also explains why prices for ordinary memory are rising. Making HBM uses far more wafer capacity per unit of memory than standard DRAM, with industry estimates commonly putting it at around three times as much, because the dies are larger, the stacking process is complex and some output is lost along the way. As manufacturers shift production lines toward HBM to meet AI demand, less capacity is left for the conventional DRAM used in PCs, phones and servers. That squeeze helps explain why SK Hynix reported DRAM selling prices up by a high-20% rate in a single quarter.Only three companies make HBM at scale: SK Hynix, Samsung Electronics and Micron. SK Hynix established an early lead and still holds the largest share, but competition is intensifying. According to Counterpoint Research, SK Hynix’s share of HBM revenue fell to around 50% in the second quarter of 2026 from around 64% a year earlier, while Samsung’s rose to around 33% as it caught up on the previous generation, known as HBM3E. Micron held around 18%. The next battleground is HBM4, which doubles the width of the connection to the processor and is being built into Nvidia’s new Vera Rubin platform. Nvidia chief executive Jensen Huang said in June that all three suppliers are in production for Vera Rubin, with the 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 any split.For investors, HBM sits at the centre of several big questions. It is a key reason memory makers are earning record profits and returning cash to shareholders, but it is also a cyclical business. Some forecasters warn that HBM prices could soften once competition and new capacity catch up with demand. The things to watch are HBM4 volumes and pricing, how quickly Samsung and Micron gain ground, and whether AI spending keeps growing fast enough to absorb new supply.  This article was written by Eamonn Sheridan at investinglive.com.