CoreWeave Launches NVIDIA GB200 NVL72 Systems for Advanced AI Workloads

Details

• CoreWeave is the first cloud provider to make NVIDIA GB200 NVL72 systems widely available, with Cohere, IBM, and Mistral AI among early adopters starting in April 2025.
• The NVIDIA GB200 NVL72 platform incorporates 72 Blackwell GPUs and 36 Grace CPUs per rack, supported by NVLink and 400Gb/s InfiniBand to scale up to 110,000 GPUs.
• Initial users have achieved 2-3x faster training speeds for models with over 100 billion parameters compared to previous Hopper GPU systems, even before specific Blackwell optimizations.
• CoreWeave’s launch follows its February 2025 availability announcement and strong MLPerf benchmark results, including 30x faster inference speeds.
• IBM is deploying these systems within its Storage Scale platform, while Mistral AI is using the infrastructure to advance open-source AI model development.

Impact

This rollout significantly accelerates AI model development by enabling 25% lower inference costs and more efficient FP4 precision. CoreWeave’s early adoption of Blackwell technology gives it an edge against larger cloud rivals, particularly for running massive models in agentic AI applications. The platform helps drive new breakthroughs in real-time reasoning and positions both CoreWeave and NVIDIA as key players in the next generation of AI infrastructure.