The high-bandwidth memory market could soon stop being a three-company race, at least for individual customers.
Micron executive vice president Sumit Sadana told the KeyBanc Technology Leadership Forum that the cost and time needed to qualify HBM designs will push chip buyers toward working with just two suppliers, or even one, per project. Today all three memory makers – SK Hynix, Samsung and Micron – ship HBM4 for Nvidia’s Vera Rubin.
The shift toward made-to-order parts begins with HBM4E, Sadana said, with Micron already planning custom SKUs with customers. He also laid out the arithmetic behind the memory crunch: every 100 bits of HBM3E output costs around 300 bits of conventional DDR supply, a ratio that worsens to nearly 4:1 with HBM4E.
Micron frames its HBM3E as the efficiency leader, claiming 30 percent lower power draw than the nearest competitor at full production volume.
Its first HBM4E, due to ramp in 2027, will be a JEDEC-standard part, with tailored versions following. Micron confirmed that TSMC will handle the logic dies for the HBM4E family, standard and custom alike, and that the memory moves to a newer DRAM process, operations chief Manish Bhatia said in May.
Rival SK Hynix is building a similar playbook. It is hiring for a Next Memory Strategy unit focused on tailored parts, and its One Team approach splits the work three ways: Nvidia sets the HBM requirements, SK Hynix does the stacking, and TSMC takes the logic die plus packaging.
