AI agents are driving up demand for KV cache and token storage, reshaping memory architecture. Cost considerations are pushing token storage down from expensive HBM and DRAM to SSDs. NAND suppliers are actively promoting next-generation solutions such as HBF, SCM, and SLC to address endurance and latency challenges, while wafer consumption climbs—making advance investment in leading-edge capacity a key competitive factor.

Key Highlights

• Architectural Shift: AI Agents feature long-term memory and autonomous planning capabilities, driving up Token and KV Cache storage demand and triggering a memory hierarchy re-architecture.
• Cost-Driven Offloading: Due to high costs and supply constraints of HBM and DRAM, offloading non-real-time data to SSDs has become an industry consensus.
• Technology Innovation: To overcome traditional SSD endurance and latency bottlenecks, vendors are advancing specialized high-performance products like HBF, SCM, and SLC/pSLC.
• Capacity Challenges: High-end NAND significantly increases wafer consumption, prompting major manufacturers to expand cleanroom space early (such as Kioxia's K3 Fab) to secure a competitive edge.

Table of Contents

1. Planning Ahead—AI Agents Drive Token Storage Demand, Positioning Next-Generation NAND Flash as a Crucial Component in Memory Hierarchy Restructuring
2. Offloading Token Storage from HBM/DRAM to SSDs Emerges as an Inevitable Trend
3. Requirements for NAND Endurance and Latency Catalyze Next-Generation High-Performance Dedicated SSDs
4. SLC Production Consumes About Three Times the Wafers of TLC; Kioxia Gains a Strategic Edge Through Early K3 Fab Expansion
   • Analysis of Products Required by AI Agents

<Total Pages: 5>