CT Sun, AIC | theCUBE + NYSE Wired: AI Data Pipeline Forum 2026
summary
Argues that memory matters more than compute for modern AI, claiming tiered architectures from HBM to cooler storage keep hot context close while offloading less urgent data, and KV caches preserve inference context so systems avoid costly recompute. Notes practical design tradeoffs: thousandx performance gaps between tiers, necessity of different SSD components, and a move to disaggregated converged storage connected via NVMe over Fabrics so compute and data scale independently.
AI-written summary · about
