SK hynix used the Future of Memory and Storage (FMS) 2026 conference in Santa Clara from August 4-6 to lay out a full-stack, tiered AI memory portfolio pointed squarely at NVIDIA's next accelerator generation and the appetite of agentic AI workloads.
16-Layer HBM4 Next To Vera Rubin
The centerpiece of SK hynix's booth was a physical 16-layer 48GB HBM4 module placed alongside a model of NVIDIA's next-generation Vera Rubin accelerator, underlining how tightly the two roadmaps are now aligned. SK hynix also displayed physical units of the 12-layer 48GB HBM4E it began sampling in July and the 12-layer 36GB HBM3E currently in mass production.
First Wafer-Bonded 375-Layer 4D NAND
On flash, SK hynix debuted its first NAND product built with wafer bonding: a 375-layer 4D NAND that delivers 2.5x the performance-per-watt of the previous generation. Enterprise SSDs based on the new NAND are slated for mass production in the first half of 2027. The company also showed the PS1101 QLC eSSD, up ~55% on the 176-layer generation, with sample shipments to major hyperscalers scheduled to begin this month. Sandisk shared the stage with SK hynix earlier in the week to unveil BiCS10 332-layer QLC NAND for AI workloads.
CMM-Ax And Tiered Memory Pitch
SK hynix used its FMS keynote to argue for a tiered stack of HBM, DRAM, NAND (via High Bandwidth Flash) and SSD, minimizing data movement for agentic AI. It also demoed CXL Pooled Memory for distributed KV cache, CMM-Hybrid for prefetch, and the new CMM-Ax compute-in-memory module, which embeds compute inside CXL memory to accelerate long-context LLM inference. The pitch dovetails with Samsung's parallel zHBM roadmap unveiled at FMS 2026.
Reporting based on coverage from StorageReview and the SK hynix newsroom.