CoreWeave adds HGX B300
CoreWeave (hpcwire.com) NVIDIA HGX B300 (Blackwell Ultra) instances and added Weights & Biases tooling, and the provider has already landed customers including Perplexity. That makes Blackwell‑class cloud access public and tightens the market for low‑latency, high‑scale inference and agent workloads.
HGX B300 nodes on CoreWeave ship with 2.1 TB of HBM3e memory — about a 50% increase versus HGX B200 — which CoreWeave says enables single‑node training on models with 100B+ parameters and long‑context inference. (coreweave.com) The platform includes next‑generation NVIDIA Quantum‑X800XDR InfiniBand that CoreWeave says doubles node‑to‑node bandwidth to reduce interconnect bottlenecks, and every HGX B300 server on CoreWeave Cloud is managed with liquid cooling for sustained peak performance. (coreweave.com) CoreWeave completed its acquisition of Weights & Biases on May 5, 2025, and has since folded W&B features into the cloud stack — including a Serverless RL capability that CoreWeave claims delivers up to 1.4× faster training at as much as 40% lower cost versus self‑managed clusters. (prnewswire.com) Perplexity signed a multi‑year agreement announced March 4, 2026 to run inference on CoreWeave, initially deploying on CoreWeave Kubernetes Service and using W&B Models for training, fine‑tuning and model management while leveraging GB200 NVL72‑powered clusters for scale. (businesswire.com) CoreWeave says it will be among the first to field NVIDIA’s Vera Rubin NVL72 platform and NVIDIA Vera CPU racks in production in the second half of 2026 as part of its roadmap to support next‑gen inference and agent workloads. (investing.com) To bolster production credentials, CoreWeave points to industry benchmark performance — citing MLPerf results and top Platinum ClusterMAX™ rankings — and customer references that it ran production workloads on HGX B200 before moving to newer Blackwell Ultra hardware. (businesswire.com)
