Samsung’s Next-Gen Memory Roadmap at FMS 2026: zHBM, 400+ Layer V-NAND, and a New AI Business Model

zHBM — Vertical HBM Concept Architecture

Perhaps Samsung's most attention-grabbing reveal was zHBM, a concept architecture that stacks high-bandwidth memory vertically directly above an AI accelerator, rather than beside it on a silicon interposer . Advanced wafer-bonding technology makes this arrangement possible . Samsung claims zHBM could deliver ~8x the performance of HBM5, over 10x the memory density, 3x the energy efficiency, and reduce thermal resistance by more than half . It is important to note that zHBM was shown as a physical concept model (mock-up), not a shipping product, and no commercialization timetable was provided . The architecture, if realized, would eliminate the silicon interposer that has underpinned every HBM generation to date .

zNAND-O — Edge AI 3D NAND Concept

Samsung also introduced zNAND-O, a concept that combines V-NAND technology with through-silicon vias (TSV) to vertically stack 4 or 8 NAND chips into a single package . The architecture is designed for edge AI workloads, with Samsung highlighting improvements in space utilization, I/O performance, and latency . Like zHBM, zNAND-O was a concept model without a specific commercialization date .

HBM4E Samples and HBM5 Models

Samsung showcased HBM4E samples — the next step in its HBM roadmap — and HBM5 concept models at its booth . The company began industry-first HBM4 mass production in February 2026 using its 1c DRAM and 4nm chip technology, and by May 2026 became the first company to deliver HBM4E samples to global customers . Mass production of HBM4E is targeted for 2027, using Samsung's 4th-generation Hybrid Bonding process .

LPDDR5X-PIM — Processing-in-Memory

Samsung demonstrated the LPDDR5X-PIM, which it described as the industry's first LPDDR memory with on-chip processing-in-memory (PIM) capability . The design executes partial data operations inside the memory itself, reducing data movement between memory and the CPU/GPU for AI inference workloads, which can improve efficiency and reduce power consumption .

Enterprise SSD Platforms

Samsung also showcased its next-generation enterprise SSD platforms, the PM1763 and BM1773, designed for AI infrastructure workloads .

Market Context: Samsung Regains DRAM Leadership

Samsung's Q2 2026 DRAM Market Share Lead

According to Counterpoint Research, Samsung held 39% of the global DRAM market by revenue in Q2 2026, overtaking SK Hynix which fell to 26% (Micron was at 25%) . The shift was driven by strong demand for commodity DRAM and Samsung's expanding HBM output, while SK Hynix faced stagnant HBM pricing . Despite SK Hynix's revenue growing 214% year-over-year, its market share declined from 39% in Q2 2025 to 26% in Q2 2026 . This marks Samsung regaining the No. 1 spot after SK Hynix briefly led in prior quarters .

Shift Toward Long-Term AI Customer Agreements (LTAs)

During its Q2 2026 earnings call on July 30, Samsung announced a major strategic shift: it plans to allocate 60–70% of its total memory production capacity under long-term supply agreements (LTAs) .

Key details of the LTA structure:

• "5-year rolling" contract structure: a 5-year base term with annual renewals .
• Contracts include upfront deposits (Samsung said it had already received approximately one-quarter of the total advance payments) and price caps that limit quarterly price reductions to a maximum of 5%, while upside pricing flexibility is essentially unlimited .
• Samsung has already signed LTAs with five major global data center customers (reportedly AWS, Microsoft, Google, Meta, and Oracle) and is in late-stage talks with five more .
• By contrast, Micron targets ~50% of its sales under its Strategic Customer Agreements .

According to a Bank of America Merrill Lynch survey, the contract terms clearly favor the supplier, with price cuts capped at 5% per quarter while upside potential is unrestricted .

Broader Industry Momentum at FMS 2026

Samsung was not the only company making news at FMS 2026:

• SK Hynix and Sandisk jointly unveiled the first High Bandwidth Flash (HBF) standard, a new storage tier designed to bridge the performance gap between HBM and SSDs . HBF specifications cover capacities up to 512 GB (in 8-high and 16-high NAND die stacks) with bandwidth ranging from ~0.4 TB/s to 3.0 TB/s, using the Universal Chiplet Interconnect Express (UCIe) interface . Google and Tenstorrent are participating in the HBF consortium .
• Kioxia showcased its GP1 Series PCIe 6.0 NVMe SSDs optimized for GPU direct access, and co-presented with Nvidia on AI infrastructure flash storage . Kioxia also featured its CM10 Series PCIe 6.0 enterprise SSDs and NX1 Series SSDs, both supporting direct liquid cooling .
• Micron already has the 9650 PCIe 6.0 enterprise SSD in mass production since February 2026 — the first commercial PCIe 6.0 drive .
• SK Hynix also demonstrated its tenth-generation (V10) 375-layer 4D NAND at its booth, claiming a 2.5x increase in power efficiency .

Bottom line: Samsung's FMS 2026 message was about architectural reinvention — zHBM and zNAND-O represent radical departures from current memory designs — backed by strong DRAM market momentum and a new LTA-driven business model that locks in 60–70% of capacity for AI customers. The broader industry is clearly aligning around tiered memory architectures (HBM, HBF, SSD) and the PCIe 6.0 interface, signaling a period of intense competition and innovation in AI infrastructure memory.