Oracle and Nvidia have announced the expansion of Oracle Cloud Infrastructure Superclusters to multi-gigawatt capacity using the Vera Rubin platform. The hardware, including Vera Rubin NVL72 racks and Vera CPUs, is currently available for rent on the Oracle cloud platform.
The NVL72 rack configuration integrates multiple GPUs using the NVLink 6 interconnect. Each GPU features HBM4 memory with up to 288 GB capacity and 22 TB/s bandwidth. The infrastructure also incorporates Nvidia BlueField-4 DPUs to manage networking, storage, and security tasks.
The partnership introduces HGX Rubin NVL8 systems, which are eight-GPU configurations designed for training frontier models and agentic AI workloads. The new architecture allows for independent scaling of CPU and GPU resources, moving away from fixed hardware bundles.
Nvidia projects that the Vera Rubin platform will provide up to 30 times higher inference throughput per watt and four times the training performance compared to the Blackwell generation. Production shipments of the platform began in the fall of 2026.
Nvidia estimates a revenue opportunity of $40 billion per gigawatt for the Vera Rubin platform. This deployment follows Oracle's previous integration of Nvidia Blackwell NVL72 systems into its cloud infrastructure.
