CoreWeave has deployed the first NVIDIA GB300 NVL72 system in a production environment, marking a significant step forward in AI infrastructure. Housed within a Dell-integrated rack-scale system, the platform features 72 NVIDIA Blackwell Ultra GPUs, 36 Grace CPUs, and 36 BlueField-3 DPUs, all interconnected by NVIDIA’s latest Quantum-X800 InfiniBand and NVLink technologies. This build pushes AI inference performance dramatically, delivering a 10x increase in user responsiveness and 5x better throughput per watt compared to Hopper-based systems.
The GB300 NVL72 is optimized for advanced reasoning workloads and massive multi-trillion parameter model deployments. CoreWeave’s custom AI cloud stack—including Kubernetes-based orchestration (CKS), Slurm on Kubernetes (SUNK), and new rack-scale observability tools—has been fine-tuned to maximize platform performance. Enhancements include real-time infrastructure-to-training correlation using Weights & Biases integrations. Moonvalley, an AI-driven filmmaking company, is among the first users of the system.
The deployment was achieved in collaboration with Dell Technologies, Switch, and Vertiv. CoreWeave’s ecosystem-driven strategy, combined with rapid platform integration, positions it as a key infrastructure provider for large-scale AI workloads. The GB300 NVL72 features 21TB of GPU memory per rack, 130TB/s of NVLink bandwidth, and 800Gbps of GPU network connectivity via ConnectX-8 SuperNICs.
- First cloud deployment of NVIDIA GB300 NVL72 by CoreWeave
- Platform features 72 Blackwell Ultra GPUs, 21TB HBM3e memory, and next-gen InfiniBand
- Delivers up to 10x boost in AI reasoning performance and 50x output vs Hopper
- Includes 130TB/s NVLink, 800Gbps RDMA per GPU, and secure multi-tenant DOCA networking
- CoreWeave stack integrates CKS, SUNK, RLCC, and Weights & Biases for deep observability
“At Moonvalley, we are building the future of generative filmmaking… The bringup of the GB300 on CoreWeave is incredibly exciting,” said Naeem Talukdar, CEO of Moonvalley.