CoreWeave Deploys Multi-Rack Nvidia Vera Rubin NVL72 Cluster for AI Workloads

CoreWeave (NASDAQ:CRWV) has announced the deployment of a multi-rack Nvidia Vera Rubin NVL72 cluster on CoreWeave Cloud, connecting hundreds of Nvidia Rubin graphics processing units (GPUs) into a single computing cluster.

The infrastructure is designed to support agentic artificial intelligence workloads, which involve AI systems performing sequences of tasks and interacting with tools.

The company also introduced two additions to its AI Object Storage service, aimed at improving data access across regions and providing an option for long-term storage.

Multi-Rack Configuration Connects Hundreds of GPUs

A single Nvidia Vera Rubin NVL72 rack combines 72 Rubin GPUs with 36 Vera CPUs, alongside Nvidia NVLink 6, ConnectX-9 SuperNICs and BlueField-4 data processing units.

CoreWeave’s multi-rack deployment connects these systems through Nvidia Spectrum-X Ethernet networking, allowing hundreds of GPUs to operate within a single scale-out cluster.

Each Rubin GPU includes two ConnectX-9 SuperNICs, providing 1.6 terabits per second of connectivity per GPU.

According to the company, the networking architecture supports approximately 128,000 GPUs per rail.

“With multi-rack Vera Rubin, we are connecting hundreds of Rubin GPUs as a single scale-out cluster,” said Chen Goldberg, executive vice president of product and engineering at CoreWeave.

CoreWeave said it was the first AI cloud provider to validate and deploy a single-rack Vera Rubin NVL72 system before extending the configuration across multiple racks. This claim was made by the company.

CoreWeave Adds Cross-Region Storage Acceleration

Alongside the computing deployment, CoreWeave announced cross-region write acceleration for its AI Object Storage service.

The feature allows data to be written locally within one region while replication to a second region takes place in the background.

According to CoreWeave, this approach removes the need for training jobs to wait for cross-region writes to complete, reducing delays associated with transferring data between locations.

The company also introduced an Archive storage tier intended for lower-cost, long-term data retention.

CoreWeave said the new tier does not impose retrieval, early deletion or read fees.

Local Object Transport Accelerator Targets Data Access Latency

CoreWeave also outlined the performance of its Local Object Transport Accelerator (LOTA), which is designed to deliver object storage reads at speeds comparable to local NVMe storage.

The company claims LOTA reduces latency by a factor of eight compared with traditional storage clusters and supports throughput of up to 7 gigabytes per second per GPU.

These figures represent CoreWeave’s reported performance claims.

Cécile Robert-Michon, director of internal infrastructure at Cohere, said the storage service provides “a unified dataset footprint across regions with reads cached locally, so nothing waits on the network.”

The storage additions are intended to support AI workloads that require access to large datasets across multiple computing regions.

Infrastructure Expansion Follows Nasdaq Listing

CoreWeave, which completed its Nasdaq public listing in March 2025, provides cloud infrastructure designed for AI computing.

The Vera Rubin deployment expands its computing configuration from a single rack to a multi-rack system, while the storage updates address data transfer and retention requirements.

The company did not disclose the financial cost of the deployment, customer commitments associated with the new cluster or expected revenue contributions from the infrastructure and storage additions.

CoreWeave stock price