Samsung unveils zHBM and zNAND-O 3D memory promising 8x performance over HBM5

At FMS 2026, Samsung revealed a next-generation 3D memory roadmap aimed squarely at AI infrastructure, unveiling concept models for zHBM and zNAND-O. Samsung claims zHBM delivers roughly 8x the performance of next-generation HBM5 and more than 10x memory density, achieved through advanced wafer-bonding techniques — a potentially major step for the memory bandwidth that increasingly bottlenecks large-model training and inference.

Memory has become one of the hardest constraints in AI systems: as models and context windows grow, feeding data to compute fast enough is often the limiting factor, which is why HBM has become a strategic chokepoint dominated by Samsung, SK Hynix and Micron. Samsung also introduced an industry-first V10 BV-NAND with more than 400 layers and reaffirmed plans to ramp HBM4 production in the second half of 2026.

The announcements slot into a week thick with hardware news — AMD's Taalas acquisition for inference-specialized silicon and NVIDIA's Vera Rubin endorsement — underscoring that the AI race is now as much about memory and packaging as about model architectures. If zHBM's performance claims hold, it could reshape the economics of feeding data to accelerators like Vera Rubin.

The caveat: these are concept models and roadmap signals, not shipping products, and memory roadmaps often slip. Watch for concrete timelines, whether zHBM's 8x claim survives independent scrutiny, how it interacts with NVIDIA's and AMD's accelerator roadmaps, and whether SK Hynix and Micron respond with competing next-gen architectures.