Pichai Balaji, ScaleFlux | theCUBE + NYSE Wired: AI Data Pipeline Forum 2026
summary
Claims that NAND must handle 95% of GPU-side memory are backed by concrete numbers: GPUs hit 20–22 TB/s, DRAM sits roughly 37–40× slower and traditional storage about 800× slower. A combined SSD plus CXL memory controller approach turns TLC NAND into a memory layer, demonstrated with 200 FDP streams and sustained endurance of 7–10 drive writes per day on standard NAND. A NAND‑agnostic reference design and contract manufacturing path is used to mitigate supply constraints.
AI-written summary · about
