CoreWeave Rolls Out NVIDIA HGX B300 ‘Blackwell Ultra’ Across Its AI Cloud
The move targets deployment‑era workloads with tooling for production reinforcement learning.
Overview
- At NVIDIA GTC, CoreWeave announced general availability of HGX B300 instances across its AI‑native cloud.
- The Blackwell Ultra hardware provides 50% more memory and doubles node‑to‑node bandwidth to support long‑context inference and models exceeding 100 billion parameters.
- Every server is liquid‑cooled to prevent thermal throttling and sustain peak performance for mission‑critical tasks.
- A new partnership with Weights & Biases brings environment‑free reinforcement learning, enabling on‑the‑job agent training with automated scoring and failure detection.
- CoreWeave says it expects to be among the first to deploy NVIDIA’s Vera Rubin NVL72 in H2 2026, as analysts recently adjusted price targets reflecting strong demand signals alongside profitability concerns.