AI Infrastructure: Memory Systems, Model-Assisted Design, and Artifact Delivery
SK hynix presented CXL pooling and tiered memory ideas for KV-cache reuse; an OpenAI model reportedly improved an upcoming AMD GPU design’s PPA by up to 14% in chip-design tests. Hugging Face reports nearly 1 EB monthly CDN delivery and about 420 GB/s, but architecture and cost details remain limited; Vertiv’s 800-volt DC pilot also lacks deployment and efficiency results.
Sources (2)
Updated Nov 22, 2026