NVIDIA and Oracle are expanding their cloud partnership, with plans for multi-gigawatt AI infrastructure on Oracle Cloud Infrastructure (OCI) that combines systems for AI training and inference with dedicated compute for autonomous AI agents.

Oracle’s Acceleron networking and AI-optimised data centres will support the expansion, which is designed to provide the networked GPU capacity needed to train and deploy increasingly large AI models.

NVIDIA Vera Rubin NVL72 rack-scale systems and standalone Vera CPUs are now generally available on OCI, where customers can scale CPU and GPU allocations independently through OCI Kubernetes Engine to reduce idle capacity and improve resource use.

OpenAI has used OCI Superclusters with more than 100,000 NVIDIA GPUs in Abilene, Texas, to train and serve frontier models, and is integrating Vera CPUs into production for agentic sandbox workloads; Uday Ruddarraju, CTO of Compute at OpenAI, said they will support code execution and tool use.

Agentic AI systems take on multi-step work—such as writing and testing code, searching databases and using software tools—rather than simply answering prompts; pairing Vera CPUs with GPUs is intended to keep GPUs productive while background tasks run.

Independent CPU and GPU scaling is intended to reduce resource waste and computational costs, while hardware-secured sandboxing is designed to let AI-generated code run without exposing sensitive enterprise data or systems.

OCI is also adding the NVIDIA HGX Rubin NVL8, an x86-based server with eight Rubin GPUs connected via NVLink, dual 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory and 6.4 Tbps RDMA networking, designed to support existing x86 software.