CoreWeave adds HGX B300 — and new tools
CoreWeave announced general availability of NVIDIA HGX B300 in its cloud and rolled out new tooling to simplify buying high‑performance GPU capacity, while also landing a Perplexity deal (hpcwire.com) and (x.com) product updates. The move validates HGX adoption among neoclouds even as CoreWeave publicly wrestles with site and power issues — a mix of growth and execution risk.
HGX B300 nodes on CoreWeave are provisioned with 2.1 TB of HBM3e memory — about a 50% increase over HGX B200 — designed to run long‑context inference and single‑node training of 100B+ parameter models. (coreweave.com) CoreWeave says every HGX B300 server on its cloud uses liquid cooling to remove thermal throttling and sustain peak Blackwell Ultra performance during long inference bursts. (finance.yahoo.com) The company also launched CoreWeave Flexible Capacity Plans — branded Flex Reservations and Spot — a unified consumption framework intended to move customers beyond strict reserved vs on‑demand choices. (investors.coreweave.com) CoreWeave disclosed a multi‑year strategic partnership to host Perplexity’s inference workloads and pilot joint services, with the agreement first reported March 4, 2026. (axios.com) Executives have underwritten that expansion with heavy capital commitments: CoreWeave guided 2026 capital expenditures of roughly $30 billion to $35 billion to add data center and power capacity. (bloomberg.com) That growth comes with operational friction—CoreWeave’s public status dashboard logged five incidents and six maintenances in the last 30 days, and lists a March 17, 2026 scheduled networking maintenance across multiple US clusters. (status.coreweave.com) CoreWeave’s platform expansion sits alongside a commercial safety net: Nvidia agreed to purchase unsold CoreWeave capacity under a deal valued initially at $6.3 billion, a backstop that runs through April 13, 2032. (finance.yahoo.com)
