Sandisk sees 60–90GB of NAND per CPU in AI agent workloads

Sandisk VP Rajeev Nagabhirava said at the SEMICON Taiwan 2026 Memory Summit that AI inference and agent-orchestration workloads can require roughly 60–90GB of NAND flash per CPU. The company is pushing HBF-based KV cache management as customers rethink how they allocate memory and storage across AI systems.

• Sandisk estimates 60–90GB of NAND per CPU for AI agent workloads
• The figure covers inference and agent orchestration, where memory roles are being reassessed
• The company is promoting HBF-based KV cache management
• The remarks were made at the SEMICON Taiwan 2026 Memory Summit