CoreWeave First to Make NVIDIA GB200 NVL72 Instances Generally Available
CoreWeave says it is the first cloud provider to make NVIDIA GB200 NVL72-based instances generally available. The cluster is built on the NVIDIA GB200 Grace Blackwell Superchip and targets customers who train, deploy and scale complex AI models.
- CoreWeave says it is the first cloud provider with generally available NVIDIA GB200 NVL72-based instances.
- Clusters can scale to up to 110,000 GPUs with 400Gb/s bandwidth per GPU via Quantum-2 InfiniBand, according to CoreWeave.
- The company claims up to 30X faster real-time LLM inference and up to 4X faster LLM training versus previous generations.
- CoreWeave is also set to deliver a GB200-based AI supercomputer to IBM for training Granite models.
- Interested customers must contact CoreWeave; no pricing is given.
What CoreWeave is offering
According to CoreWeave, which calls itself the AI Hyperscaler™, the GB200 NVL72-powered cluster is meant to help customers rapidly train, deploy and scale large AI models. The release is dated Feb. 4, 2025.
Brian Venturo, co-founder and Chief Strategy Officer of CoreWeave, described the launch as "another achievement of our series of firsts." He pointed to the company's cloud services, including CoreWeave Kubernetes Service, Slurm on Kubernetes (SUNK) and its Observability platform, which CoreWeave says are built to make it easier to run, manage and scale AI workloads on cutting-edge hardware.
Networking and performance claims
CoreWeave argues that the scalability of advanced models can be constrained by server limits, especially memory capacity and the speed of communication between GPUs. Its GB200 NVL72 instances use rack-level NVLink connectivity and NVIDIA Quantum-2 InfiniBand networking. The company says this delivers 400Gb/s of bandwidth per GPU through a rail-optimized topology, for clusters of up to 110,000 GPUs.
The setup also uses NVIDIA Quantum-2's SHARP In-Network Computing technology, which CoreWeave says can further optimize collective communication and result in ultra-low latency and faster training.
The company lists the following figures for the platform, compared with previous generations:
- Up to 30X faster real-time large language model (LLM) inference.
- Up to 25X lower Total Cost of Ownership and 25X less energy for real-time inference.
- Up to 4X faster training of LLM models.
Track record and partners
CoreWeave places the launch in a line of earlier hardware milestones. Last August, it says, it was among the first to offer NVIDIA H200 GPUs to train the fastest GPT-3 LLM workloads. In November it was one of the first to demo NVIDIA GB200 systems in action. Earlier in the same month as the release, it announced that it will deliver one of the first NVIDIA GB200 Grace Blackwell Superchip-enabled AI supercomputers to IBM for training its next generation of Granite models.
Priya Nagpurkar, VP, Hybrid Cloud and AI Platform Research at IBM, said the partnership, which includes IBM Spectrum Scale Storage, shows IBM's commitment to "advancing a hybrid cloud strategy for AI." Ian Buck, Vice President of Hyperscale and HPC at NVIDIA, said scaling for inference and training is "one of the largest challenges for organizations developing next generation AI workloads."
What it means for customers
For organizations that need large-scale GPU capacity, the announcement means GB200 NVL72 instances can now be requested from a cloud provider rather than only seen in demos. The source gives no pricing or regional availability. Interested customers are asked to contact CoreWeave for more information.
The performance and cost figures are CoreWeave's own claims, stated as maximums against earlier generations, so actual results will depend on the workload.
