On October 27, 2026, NVIDIA and Oracle announced the expansion of Oracle Cloud Infrastructure (OCI) Superclusters to a multi-gigawatt scale powered by NVIDIA’s next-generation “Vera Rubin NVL72,” alongside the official launch of the “Vera CPU” on OCI. This partnership expansion focuses on enhancing not only advanced training and inference for large language models (LLMs) but also addressing data processing bottlenecks encountered by agentic AI during tool execution and data retrieval, while scaling power and compute resources to the next level.
Combining Multi-Gigawatt OCI Superclusters with Advanced Architecture
According to the announcement, the companies are deploying the rack-scale Vera Rubin NVL72 compute system within OCI Supercluster environments equipped with multi-gigawatt power and cooling infrastructure. The system leverages NVIDIA NVLink interconnect technology to combine 72 Vera Rubin GPUs per rack, delivering immense computing performance. OCI’s multiplanar network architecture supports high-speed connectivity for hundreds of thousands of GPUs as the system scales.
The OCI Supercluster integrates “NVIDIA Spectrum-X” Ethernet networking, utilizing Spectrum-6 switches and BlueField-4 DPUs, alongside Oracle’s custom “Oracle Acceleron” network interface. This expansion builds upon the existing GB200 and GB300 NVL72 infrastructure technology in OCI, with a design that ensures interoperability with existing workloads via a common CUDA software foundation, even as hardware is upgraded.
Supporting Agentic AI and Resolving Bottlenecks with Vera CPU
In this announcement, the companies confirmed the general availability of the dedicated “NVIDIA Vera CPU” within the Oracle Cloud Infrastructure (OCI) Compute environment. Modern AI agents require more than simple response generation; they involve CPU-intensive tasks such as external tool calling, code execution, and database searches. In existing infrastructures, if these computations lag, GPUs often sit idle between model calls, limiting overall efficiency.
The Vera CPU combines high single-thread performance with robust memory bandwidth to help execute complex agent processes rapidly. Using the OCI Kubernetes Engine (OKE), users can independently scale CPU and GPU resources, reducing idle capacity. Furthermore, the “Oracle Acceleron” network enhances security by providing hardware-level isolation for the execution of external code generated by AI.
Introducing x86-based ‘HGX Rubin NVL8’ and Specifications
In addition to the NVL72 rack-scale systems, NVIDIA and Oracle announced plans to roll out the x86-based “NVIDIA HGX Rubin NVL8” system, which features eight Rubin GPUs in a single server. This model adopts a lightweight 8-GPU configuration, allowing enterprises to adopt next-generation Rubin GPUs while maintaining existing x86-compatible application environments.
The key specifications for the HGX Rubin NVL8, as detailed in the press release, include:
- Compute Accelerator: 8 NVIDIA Rubin GPUs interconnected via NVLink
- Processor and Memory: Dual 64-core Intel Xeon 6 CPUs and 4TB of DDR5 system memory
- Network Bandwidth: 6.4Tbps high-speed RDMA (Remote Direct Memory Access) network interface
This system enables enterprise customers to choose compute resources tailored to specific workload requirements, including fine-tuning, frontier AI training, and small-scale inference.
Real-World Validation with OpenAI’s ‘GPT-6 Astra’
The announcement highlighted the work of global AI lab OpenAI as a flagship implementation of the combined NVIDIA and Oracle infrastructure. OpenAI reported that it utilized over 100,000 NVIDIA Blackwell GPUs at the OCI Supercluster in Abilene, Texas, to train and serve its next-generation frontier model, “GPT-6 Astra.”
Sreekanth Pothanis, leader of compute infrastructure engineering at OpenAI, noted that the OCI-based Vera Rubin system provides the scalability required for the demanding compute needs of next-generation models. Furthermore, OpenAI’s CTO of Compute, Uday Ruddarraju, explained that based on early access results with the Vera CPU in the OCI environment, they are deploying the Vera CPU into production to handle the rapidly increasing code execution and tool-handling requirements of their Agentic Sandbox workloads.
Infrastructure Flexibility and Technical Disclosure
Ian Buck, Vice President of Hyperscale and HPC at NVIDIA, and Mahesh Thiagarajan, Executive Vice President of OCI at Oracle, emphasized that AI factories must be built with the durability to handle continuous software optimization and future model shifts. The companies added that the combination of Vera Rubin GPUs, Vera CPUs, and Oracle Acceleron technology will maximize GPU utilization and improve power efficiency.
This information is based on a single press release provided by NVIDIA and Oracle on October 27, 2026; third-party independent performance verification and detailed power consumption metrics were not included in the original material. The companies intend to provide updates on construction schedules and regional rollout plans as cloud commercialization progresses.
Official Source: