Quick Overview
- Shares of CoreWeave advanced approximately 3% in premarket hours Wednesday following news of a multi-rack Nvidia Vera Rubin NVL72 cluster deployment on its cloud platform.
- The infrastructure links hundreds of Rubin GPUs together into one unified scale-out system optimized for agentic AI applications.
- CoreWeave is positioning itself as the inaugural AI cloud platform to successfully validate and deploy a Vera Rubin NVL72 system.
- Additionally, the firm unveiled two enhanced AI Object Storage capabilities: write acceleration across regions and a new Archive storage tier.
- Nvidia (NVDA) shares increased roughly 1% following the announcement.
CoreWeave (CRWV) shares gained approximately 3% during Wednesday’s premarket session following the company’s announcement of a multi-rack Nvidia Vera Rubin NVL72 cluster deployment on its cloud platform.
CoreWeave, Inc. Class A Common Stock, CRWV
This development builds upon CoreWeave’s June achievement, when it established itself as the inaugural AI cloud platform to activate a single Vera Rubin NVL72 rack. Wednesday’s reveal represents a significant expansion of that capability.
The newly implemented multi-rack configuration integrates hundreds of Rubin GPUs into one unified scale-out system. This represents a substantial increase in computational power for enterprises operating intensive AI applications.
Each individual Vera Rubin NVL72 rack incorporates 72 Rubin GPUs paired with 36 Vera CPUs, alongside Nvidia NVLink 6 technology, ConnectX-9 SuperNICs, and BlueField-4 DPUs. The multi-rack configuration links these units together through Nvidia’s Spectrum-X Ethernet infrastructure.
The ConnectX-9 SuperNICs provide 1.6 Tb/s scale-out bandwidth per GPU through multiplane, multirail pathways. This architecture can accommodate approximately 128,000 GPUs per rail within a non-blocking network fabric.
The flexible architecture allows additional racks to be integrated without requiring a complete infrastructure overhaul. This adaptability becomes critical when enterprises need to expand rapidly.
Chen Goldberg, EVP of product and engineering at CoreWeave, noted that the multi-rack Vera Rubin system provides organizations developing agentic AI with “greater scale, faster iteration, and higher productivity as models and agents continuously learn and improve.”
The system is engineered to accommodate both model training and inference operations at enterprise scale. Agentic AI applications, which demand that models continuously adapt and respond, place substantial requirements on computing infrastructure.
Enhanced Storage Capabilities Unveiled
In conjunction with the cluster announcement, CoreWeave rolled out two additional capabilities for its AI Object Storage solution.
The first enhancement enables cross-region write acceleration, allowing data to be written with local-level latency while simultaneous background replication occurs to remote regions. The second addition is an Archive storage tier designed for economical storage without retrieval fees, early deletion charges, or reading costs.
CoreWeave’s Local Object Transport Accelerator (LOTA) technology enables read operations at local NVMe performance levels. The platform claims it cuts latency by as much as 8x when compared to conventional storage cluster configurations.
Storage Benchmark Data
According to CoreWeave, LOTA delivers throughput reaching 7 GB/s per GPU. This performance level is achieved via managed caching implemented on every CoreWeave Kubernetes Service node.
The Archive tier serves as an economical alternative for datasets requiring infrequent access while remaining available within CoreWeave’s infrastructure environment.
Nvidia shares moved approximately 1% higher on Wednesday in tandem with CoreWeave’s announcement. The partnership between these two companies draws significant industry attention, as Nvidia hardware forms the foundation of CoreWeave’s cloud operations.
CoreWeave emphasized that it remains the sole AI cloud provider to have successfully validated and implemented the Vera Rubin NVL72 architecture across multiple racks.





