CoreWeave has become one of the first cloud providers to offer NVIDIA GB200 NVL72 systems at scale, giving its customers access to thousands of the newly launched NVIDIA Grace Blackwell GPUs. Early adopters including Cohere, IBM, and Mistral AI are already leveraging this infrastructure to develop, train, and deploy next-generation Artificial Intelligence models and applications, accelerating advancements across the industry.
The NVIDIA GB200 NVL72 platform is engineered for large-scale reasoning and Artificial Intelligence agents, featuring a rack-scale architecture with 72 interconnected GPUs via NVIDIA NVLink and advanced memory capabilities. Cohere is using these systems to power secure enterprise Artificial Intelligence applications through its North platform, reporting up to three times better performance in training 100-billion-parameter models compared to prior-generation GPUs. Further performance gains are anticipated as Cohere unlocks Blackwell-specific optimizations, taking advantage of unified memory and improved floating-point precision.
IBM has deployed thousands of Blackwell GPUs on CoreWeave’s infrastructure to accelerate training of its open-source Granite model family, which aims to support enterprise Artificial Intelligence with a focus on safety, speed, and cost efficiency. These models drive solutions such as IBM watsonx Orchestrate, enhancing workflow automation and deployment of capable enterprise agents. The integration with IBM’s high-performance Storage Scale System ensures reliable and rapid data access within core Artificial Intelligence workflows.
Mistral AI, based in Paris and a leader in open-source Artificial Intelligence, is utilizing its first thousand Blackwell GPUs via CoreWeave to build advanced language models like Mistral Large, citing a twofold improvement in dense model training performance even before optimizations. The GB200 NVL72’s scale and connectivity enable fast iteration and deployment for high-demand language applications. CoreWeave continues to expand availability, offering rack-scale NVIDIA instances and preparing to scale up to 110,000 GPUs with advanced InfiniBand networking, thus empowering the next generation of Artificial Intelligence agents and large-scale reasoning systems for innovators across the globe.
