CoreWeave Deploys Multi-Plane Network Fabric and Claude Agent Sandboxes

CoreWeave has deployed NVIDIA's Spectrum-X Multiplane into production at its Vera Rubin racks, a multi-plane Ethernet network that NVIDIA states scales to 512,000 GPUs and delivers 1.6x better AI networking performance than off-the-shelf Ethernet. CoreWeave also announced support for running Claude agents inside sandboxed containers controllable from a mobile phone, with access to custom tools, MCPs, Slurm, and configurable network policies. CoreWeave stated that agentic AI creates fundamentally different network traffic patterns, with thousands of concurrent agent chains producing unpredictable, latency-sensitive traffic alongside training and inference workloads on the same fabric, and distilled its experience into five lessons. The Wall Street Journal described agentic AI as creating two distinct computing challenges: efficiently processing large amounts of context and generating tokens with extremely low latency. Groq's 3 LPX processor addresses the latency problem through deterministic compiler scheduling, preplanned chip-to-chip transfers, and 128GB of SRAM across the rack.