CoreWeave goes big on B300 + Blackwell
CoreWeave announced an HGX B300 expansion and posted benchmark claims — 3.42× faster token generation for reasoning models and 270 GB HBM3e per GPU — while also deploying PhaidraAI agents across liquid‑cooled NVIDIA Grace Blackwell nodes for huge thermal and response gains. The company’s social posts tout 80% fewer thermal spikes and sub‑10s agent response times, but a separate social thread flagged a $66.8B backlog, power delays, capex cuts and a class action — growth and ops risk at once. ( )
CoreWeave’s HGX B300 rollout was announced in a company news release tied to NVIDIA’s GTC event on March 16, 2026, marking the platform as generally available on CoreWeave Cloud. (coreweave.com) The company’s blog post named benchmark workloads Kimi K2.5 and DeepSeek‑R1 in side‑by‑side comparisons against prior Blackwell systems to illustrate the new B300 configuration gains. (coreweave.com) System integrator and vendor datasheets list differing per‑GPU HBM3e capacities for Blackwell Ultra hardware — Supermicro’s B300 materials quote up to 288 GB of HBM3e per GPU while NVIDIA reference materials show ~279 GB for GB300/Blackwell Ultra configurations. (supermicro.com) Phaidra’s announcement with CoreWeave and Applied Digital describes a self‑learning RL cooling agent that uses real‑time rack power telemetry as a leading indicator to predict and prevent thermal spikes in liquid‑cooled deployments. (phaidra.ai) A securities class action filed by Hagens Berman alleges CoreWeave understated data‑center delays tied to a Denton, Texas cluster, and the firm publicly disclosed the investigation via a PR Newswire release in March 2026. (prnewswire.com) CoreWeave reported full‑year 2025 revenue of roughly $5.1 billion in its February 26, 2026 earnings release while calling out a materially larger multi‑year contract backlog in the same filing. (investors.coreweave.com) The company has been funding capacity growth with large financing deals — including a $2.6 billion secured debt facility announced in 2025 — alongside multi‑billion dollar customer agreements such as an expanded OpenAI contract disclosed in 2025. (investors.coreweave.com)
