CoreWeave adds HGX B300

HGX B300 nodes ship with 288 GB of HBM3e per GPU in 8‑GPU configurations, yielding roughly 2.1 TB of aggregated HBM3e per node (stated as a 50% memory increase versus B200 in CoreWeave’s release). (supermicro.com) The platform uses 5th‑generation NVLink with a 1.8 TB/s NVLink GPU‑to‑GPU fabric plus NVIDIA Quantum‑X800XDR InfiniBand to double node‑to‑node bandwidth for multi‑node reasoning and inference workloads. (supermicro.com) CoreWeave says every HGX B300 server on its cloud is liquid‑cooled to avoid thermal throttling and sustain sustained peak performance, and Supermicro’s reference 4U liquid‑cooled HGX B300 chassis lists redundant 6,600W titanium power supplies for dense racks. (coreweave.com) The March 16 announcement bundled platform updates with integrated Weights & Biases features — including environment‑free reinforcement learning, W&B Inference, and Mission Control observability — following CoreWeave’s completed acquisition of Weights & Biases (reported at about $1.7 billion). (coreweave.com) CoreWeave re‑affirmed plans to be among the first providers to deploy NVIDIA’s Rubin NVL72 rack and Vera CPU racks in production in the second half of 2026, positioning the company to offer both HGX B300 nodes now and rack‑scale Rubin systems later this year. (coreweave.com) Customer and partner signals cited in the release include a quote from Aman Sanger of Cursor attesting to prior production runs on HGX B200, and early integrations such as Cline’s use of CoreWeave’s W&B Inference to power autonomous coding workflows. (coreweave.com)