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

This is a storage-hierarchy disruption, not a capacity upgrade. AI agents require persistent state, tool-call memory, and cross-session context — 60–90 GB per CPU signals NAND migrating from cold-archive to quasi-working-memory, directly encroaching on DRAM's economic territory. The technical cascade is immediate: enterprise SSD controllers must hit sub-microsecond latency, forcing CXL 3.0 and PCIe Gen6 adoption roughly 18 months ahead of schedule. Upstream, 300+ layer V-NAND becomes the minimum viable node, concentrating wafer capacity in China Taiwan and Japan. Strategically, SanDisk publicly quantifying demand is a pricing anchor. If AI workloads absorb 30–40% of enterprise NAND wafer starts by 2026, the 2023–24 price collapse locks in as a structural floor. Samsung will likely rebrand its QLC enterprise line as "agent memory" to capture premium ASPs. Micron, with HBM3E consuming fab space, risks ceding enterprise SSD share to Kioxia and the SanDisk JV. The 12–24-month tail: NAND pricing decouples from smartphone and PC cycles. A distinct "AI NAND" category emerges with 3–5× ASP premium over consumer TLC. The deeper variable — as per-CPU NAND density scales, the DRAM-to-NAND cost ratio (currently ~10:1) compresses toward 3:1, rewriting server BOM economics and making memory the next compute bottleneck after HBM.