Nvidia and Oracle Expand OCI Partnership

Nvidia and Oracle said on the 27th, local time, that they had officially launched Nvidia Vera Rubin NVL72 systems and standalone Vera CPUs on Oracle Cloud Infrastructure (OCI). They also announced plans to expand OCI Superclusters to multiple gigawatts (GW), citing demand for model training and inference and the growth of agentic AI systems that use multiple tools.

The partnership centers on scaling GPU and CPU resources independently according to the work each handles. GPUs train and run large models, while CPUs handle code execution, information retrieval and software tools between an agent's model calls. The companies said slow CPU work can leave GPUs waiting to start their next task. They argue that expanding only the resources needed for actual workloads can reduce unused capacity compared with scaling CPUs and GPUs together.

Oracle said Vera CPUs handle CPU-intensive work such as agents' code execution and tool calls, with high single-thread performance and memory bandwidth. OCI Kubernetes Engine can scale CPU and GPU resources separately, while Oracle Acceleron networking uses hardware-based isolation for environments that run untrusted AI-generated code. OpenAI is among Vera's early adopters. Uday Radhakrishnan, its CTO for computing, said the company is deploying Vera in OCI production environments to run code and tools for agent sandbox workloads.

Nvidia said Vera CPUs are up to 80% faster than major x86 processors on agent-related CPU tasks. That figure comes from Nvidia platform benchmarks; it does not mean the same improvement has been verified on OCI instances. The article notes that lower GPU utilization does not always result from CPU bottlenecks. Companies should compare existing and new configurations on the same agent workloads, measuring completion rates, end-to-end response times, CPU and GPU use, and total cost.

OCI Superclusters use NVL72 systems, Spectrum-6 switches and BlueField-4 DPUs to support large-scale model training and inference. Each NVL72 rack connects 72 GPUs with NVLink, and OCI's network is designed to scale to hundreds of thousands of GPUs. Oracle is adding the Vera Rubin platform after infrastructure based on GB200 and GB300 NVL72 systems, and plans to offer HGX Rubin NVL8, which connects 8 Rubin GPUs. NVL8 remains at the planning stage, not an official launch; it targets smaller-scale training, fine-tuning and inference with servers based on Intel Xeon 6 CPUs.

OpenAI said it trains and serves GPT-6 Astra on an Oracle Abilene data center Supercluster equipped with more than 100,000 Nvidia Blackwell GPUs. In a separate announcement, Oracle said it plans to use systems from AI chip startup Cerebras to provide OCI cloud services starting in the first half of 2027. Cerebras develops computers built around large chips that use entire silicon wafers and says they can help reduce inference latency. The companies did not disclose the contract value or investment amount.

Nvidia and Oracle said on the 27th, local time, that they had officially launched Nvidia Vera Rubin NVL72 systems and standalone Vera CPUs on Oracle Cloud Infrastructure (OCI). They also announced plans to expand OCI Superclusters to multiple gigawatts (GW), citing demand for model training and inference and the growth of agentic AI systems that use multiple tools.

The partnership centers on scaling GPU and CPU resources independently according to the work each handles. GPUs train and run large models, while CPUs handle code execution, information retrieval and software tools between an agent's model calls. The companies said slow CPU work can leave GPUs waiting to start their next task. They argue that expanding only the resources needed for actual workloads can reduce unused capacity compared with scaling CPUs and GPUs together.

Oracle said Vera CPUs handle CPU-intensive work such as agents' code execution and tool calls, with high single-thread performance and memory bandwidth. OCI Kubernetes Engine can scale CPU and GPU resources separately, while Oracle Acceleron networking uses hardware-based isolation for environments that run untrusted AI-generated code. OpenAI is among Vera's early adopters. Uday Radhakrishnan, its CTO for computing, said the company is deploying Vera in OCI production environments to run code and tools for agent sandbox workloads.

Nvidia said Vera CPUs are up to 80% faster than major x86 processors on agent-related CPU tasks. That figure comes from Nvidia platform benchmarks; it does not mean the same improvement has been verified on OCI instances. The article notes that lower GPU utilization does not always result from CPU bottlenecks. Companies should compare existing and new configurations on the same agent workloads, measuring completion rates, end-to-end response times, CPU and GPU use, and total cost.

OCI Superclusters use NVL72 systems, Spectrum-6 switches and BlueField-4 DPUs to support large-scale model training and inference. Each NVL72 rack connects 72 GPUs with NVLink, and OCI's network is designed to scale to hundreds of thousands of GPUs. Oracle is adding the Vera Rubin platform after infrastructure based on GB200 and GB300 NVL72 systems, and plans to offer HGX Rubin NVL8, which connects 8 Rubin GPUs. NVL8 remains at the planning stage, not an official launch; it targets smaller-scale training, fine-tuning and inference with servers based on Intel Xeon 6 CPUs.

OpenAI said it trains and serves GPT-6 Astra on an Oracle Abilene data center Supercluster equipped with more than 100,000 Nvidia Blackwell GPUs. In a separate announcement, Oracle said it plans to use systems from AI chip startup Cerebras to provide OCI cloud services starting in the first half of 2027. Cerebras develops computers built around large chips that use entire silicon wafers and says they can help reduce inference latency. The companies did not disclose the contract value or investment amount.