Aug 12 - At the Keybanc Technology Conference on Aug 11 (ET), Micron CBO Sumit Sadana made a striking claim: custom HBM memory will officially launch as HBM4 evolves to HBM4E.
Sadana noted HBM4’s Base Die uses logic CMOS, currently integrating memory interface IP. For HBM4E, some clients want to "offload" specific AI xPU circuits to HBM Base Die to boost xPU area efficiency—spawning "custom HBM." This makes HBM more than just off-chip cache; it becomes part of xPU logic, with custom HBM incompatible with JEDEC standard HBM.
Custom HBM Base Die costs are high. Beyond AI chip leaders like NVIDIA, most xPU firms can’t partner with all three HBM makers, limiting supply to 1-2 vendors. This shift will reshape the market, with widespread "exclusive" custom HBM partnerships emerging.
On the same day, Micron raised Q4 FY2025 (ending Aug 28) guidance: revenue from $10.7±0.3B to $11.2±0.1B. Sadana attributed the upgrade to higher average selling prices (unchanged shipment forecasts), driven by strong AI/data center demand and HBM squeezing non-HBM DRAM capacity.
ICgoodFind: Micron’s HBM4E insights—custom HBM and "exclusive" models—could steer industry trends, demanding attention.
