CoreWeave has become the first AI cloud provider to deploy Nvidia's Vera Rubin NVL72 system. The company received the hardware from Dell on May 31 and made it operational within 24 hours. The stock price of CoreWeave rose by approximately 14% following the announcement, with trading volume doubling its three-month average.
Each NVL72 rack features 72 Rubin GPUs and 36 Vera CPUs, utilizing NVLink 6 interconnects with 260 TB/s bandwidth. The system provides up to 3.6 exaFLOPS of NVFP4 inference capability per rack. According to internal benchmarks, the architecture offers 10 times the inference throughput per megawatt compared to the Blackwell generation.
The company claims the new system reduces the number of GPUs required for equivalent workloads by half and lowers costs to approximately one-tenth per million tokens. Clusters can scale to thousands of GPUs using Spectrum-X Ethernet or Quantum-X800 InfiniBand networking.
CoreWeave introduced proprietary management tools, including Valvey for thermal performance monitoring and Racky for unified rack control. Additionally, the company launched LOTA, a storage solution providing data transport speeds of up to 7 GB/s per GPU. Full production availability of the clusters is expected later in 2026.
