Companies like Cohere, IBM, and Mistral AI are already utilizing the new NVL72 systems in the cloud to train large-scale AI agents.
CoreWeave has become one of the first cloud providers to bring NVIDIA GB200 NVL72 systems online at scale, providing access to thousands of Grace Blackwell GPUs for its clients. Leading companies in artificial intelligence such as Cohere, IBM, and Mistral AI are already leveraging this infrastructure to train and deploy advanced AI models with unprecedented performance.
The new NVL72 systems, designed for AI agent workloads and reasoning, offer an architecture with 72 GPUs interconnected via NVLink and 36 Grace CPUs, enabling massive data processing with low latency, energy efficiency, and unmatched scalability.
Cohere: Up to 3 Times the Performance for 100 Billion Parameter Models
Cohere, a company specializing in enterprise AI solutions, has already begun using these systems to train massive models and deploy custom agents through its North platform. According to the company, the NVL72 systems offer up to 3 times the performance of previous Hopper GPUs, even before applying specific optimizations for Blackwell.
Autumn Moulder, Vice President of Engineering at Cohere, emphasizes that the integration has been seamless and that unified memory and FP4 precision of the new system are enabling more efficient and cost-effective inferences. “We look forward to seeing even greater improvements with future optimizations,” she stated.
IBM: Driving Granite Models with Secure and Efficient AI
IBM has also accessed thousands of GB200 GPUs through CoreWeave to accelerate the training of its Granite model family, which will be open source and focused on enterprise applications with an emphasis on security, speed, and cost.
These models form the basis of solutions like IBM watsonx Orchestrate, aimed at automating workflows using AI agents. IBM is also using its IBM Storage Scale System, integrated with CoreWeave’s cloud, to ensure I/O performance that meets these workloads.
“This collaboration with CoreWeave will enable us to build high-performance, low-cost models for the next generation of enterprise AI,” said Sriram Raghavan, Vice President of AI at IBM Research.
Mistral AI: Doubling Performance for Open Source Models
The French startup Mistral AI, known for its open-source models with high reasoning power, is receiving its first thousand GB200 GPUs to accelerate the development of new generations of linguistic models.
According to Thimothee Lacroix, co-founder and CTO of Mistral, “without additional optimizations, we have already seen a 2× improvement in training dense models,” which opens up new possibilities for scaling their models and real-time inferences.
CoreWeave and Mass Deployment of NVIDIA Blackwell
CoreWeave not only offers these instances for long-term enterprise customers but also makes them available to other users through its cloud infrastructure. Each instance includes rack-scale NVLink, with NVIDIA Quantum-2 InfiniBand networks, and can scale up to 110,000 GPUs, positioning itself as one of the most powerful platforms in the world for developing artificial intelligence.
This advancement marks a new milestone in the evolution of accelerated computing, allowing companies to train and deploy increasingly complex and powerful AI models with unprecedented efficiency.
With this implementation, CoreWeave and NVIDIA reinforce their leadership in the race to enable the infrastructure needed for next-generation artificial intelligence.
Source: Nvidia