FMS 2026: Samsung zHBM, 400+ Layer NAND, and the New AI Memory Stack

Alongside zHBM, Samsung also showed a zNAND-O concept architecture designed for edge AI inference .

Samsung V10 BV-NAND — Industry-First 400+ Layer Wafer-Bonded NAND, in Mass Production

Samsung introduced V10 BV-NAND (Bonding V-NAND), the industry's first NAND flash built with wafer bonding technology, separating the memory array from peripheral logic and bonding the wafers together .

• Layer count: More than 400 active layers, the highest Samsung has ever disclosed for its V-NAND line .
• Density: Approximately 58% higher memory density than the previous V9 generation .
• Energy and thermal efficiency: The wafer-bonding approach roughly triples energy efficiency and cuts thermal resistance, according to Samsung .
• Production status: V10 BV-NAND is already in mass production, making it the first NAND to pass the 400-layer mark commercially .

SK hynix High Bandwidth Flash (HBF) — Open Standard Backed by Google and Tenstorrent

SK hynix (along with SanDisk) released the first open technical specification for High Bandwidth Flash (HBF) through the Open Compute Project (OCP) at FMS 2026 .

• Specifications: The initial standard supports capacities up to 512GB using 8-high and 16-high NAND stacks, with up to 3.0 TB/s bandwidth via a UCIe interface .
• Ecosystem backing: Google DeepMind and Tenstorrent have joined the HBF alliance .
• Purpose: HBF is intended to serve as a capacity-focused storage tier between high-bandwidth DRAM and conventional NVMe SSDs, addressing AI inference memory bottlenecks .

Industry Consensus: HBM Alone Is Insufficient — Hierarchical Memory Is the Future

Across keynotes and panels, a strong industry consensus emerged that HBM on its own cannot meet AI's growing demands. Multiple presenters argued that hierarchical memory structures — combining HBM, HBF-class flash, CXL-attached memory, and NVMe storage in a tiered architecture — will be central to future AI infrastructure . Marvell's presentations, for instance, explicitly focused on agentic AI inference infrastructure requiring a diversified memory and storage portfolio beyond just HBM .

What Was Not Verified

Two claims from the original question — that Nvidia predicted AI agents would exceed the global population and shift memory pricing to cost-per-I/O, and that Samsung, SK hynix, and Micron have sold out their entire 2027 DRAM and HBM production with customers receiving only 60-70% of requested volumes — could not be confirmed through any of the available published sources from FMS 2026. These statements may have been made in a keynote or panel that has not been captured in indexed reports.