Charles Giancarlo, Pure Storage | AI Infrastructure Silicon Valley - Executive Series
summary
Claims that writing speed, not just read throughput, is the critical bottleneck for large model training because frequent checkpointing bottleneck pauses cost days of work; Pure touts RAW flash firmware and data reduction that cut power and footprint while enabling near-instant writes. Describes a Fusion virtualized pool that presents on‑prem and cloud storage as one API driven layer for RAG and inference. Cites deployments including half an exabyte in Meta supercluster with 24,000 GPUs, 18% growth, and a hyperscaler design win to replace spinning disk.
AI-written summary · about
