Bernie Wu, MemVerge | theCUBE + NYSE Wired: AI Data Pipeline Forum 2026
summary
Argues that memory architecture is the primary AI bottleneck and that CXL becomes a new midtier between DRAM and SSDs, expanding the hierarchy; describes software that glues heterogeneous hardware into CXL tier deployments and a GPU-agnostic, storage-agnostic layer; highlights KV caching and reuse of precomputed embeddings to keep expensive GPUs utilized; warns that fault tolerance and persistence return as recompute costs rise; demos an abstraction called GPU surfing and cites a PCIe 6 CXL card as a concrete hardware trend.
AI-written summary · about
