CoreWeave brings up first NVIDIA GB300 NVL72 in Dell rack-scale system

What CoreWeave deployed

CoreWeave describes the NVIDIA GB300 NVL72 as a rack-scale, liquid-cooled platform built for demanding AI reasoning workloads. According to the company, each system combines 72 NVIDIA Blackwell Ultra GPUs, 36 Arm-based NVIDIA Grace CPUs and 18 NVIDIA BlueField-3 DPUs in a single platform.

The deployment continues a pattern CoreWeave points to in its own history. It says it was the first AI cloud provider to offer access to the NVIDIA HGX H100 system, the NVIDIA H200 and the NVIDIA GB200 NVL72. The GB300 NVL72 is housed in Dell's integrated rack scale system, and CoreWeave names Dell Technologies, Switch and Vertiv as partners on this deployment.

Performance claims and hardware details

CoreWeave states that the GB300 NVL72 delivers up to a 10x boost in user responsiveness and a 5x improvement in throughput per watt compared with the previous generation NVIDIA Hopper architecture. The company says this adds up to a 50x increase in output for reasoning model inference. It also cites 1.5x more dense FP4 performance and 2x higher NVIDIA Quantum-X800 InfiniBand speeds.

The company lists these platform features:

• Blackwell Ultra GPUs: 1.5x more AI compute FLOPS than Blackwell GPUs.
• HBM3e memory: up to 21TB of high-bandwidth GPU memory per rack, which CoreWeave says is 1.5x more than GB200 NVL72 and allows larger batch sizes and larger models.
• Fifth-generation NVLink: 130TB/s of aggregate bandwidth between GPUs in the system.
• InfiniBand networking: NVIDIA Quantum-X800 switches and ConnectX-8 SuperNICs provide 800 gigabits per second of dedicated connectivity to each GPU.
• Multi-tenant networking: NVIDIA DOCA running on BlueField-3 provides line-speed tenant networking up to 200Gb/s and high-performance GPU data access.

CoreWeave says the system is engineered for test-time scaling inference. For context, it also recalls that its GB200-based instances delivered 2.86x better per-GPU inference performance than H100-based instances, a result it attributes to its focus on AI workloads.

Software and observability additions

CoreWeave says it optimized its cloud stack for the new hardware, including its Kubernetes service (CKS), its Slurm on Kubernetes offering (SUNK), its observability tooling and its custom-designed Rack LifeCycle Controller (RLCC).

New with this launch are the Cabinet Wrangler and Cabinet Details dashboard, which give insight at rack scale. Hardware-level data and cluster health events are now streamed directly through Weights & Biases models. According to CoreWeave, this lets AI researchers correlate infrastructure issues with their training runs and resolve disruptions faster.

Moonvalley, a generative filmmaking company, is quoted in the announcement. Co-Founder and CEO Naeem Talukdar said that CoreWeave's ability "to quickly bring it to market via its AI cloud platform" is "a major enabler for AI-driven filmmaking to become a reality for our industry."

What it means for customers

For customers, the announcement signals early access to the GB300 NVL72 on a cloud platform that CoreWeave says is built to train, optimize and deploy multi-trillion parameter models. The company frames the platform as a step up for reasoning and agentic workloads.

The release does not give pricing, regional availability or a date for wider access. CoreWeave only says to stay tuned for updates as it moves toward broader availability. All performance figures in the announcement are the company's own claims.