Nvidia, Oracle Supercharge AI Infrastructure

Nvidia and Oracle have expanded their partnership to enhance AI infrastructure, with Vera Rubin NVL72 systems and standalone Vera CPUs now generally available on Oracle Cloud Infrastructure (OCI). This expansion includes plans to scale their Superclusters to a multi-gigawatt level.

The announcement addresses the dual needs of AI: the computational power for training sophisticated models and the CPU capacity required for AI agents to execute code and utilize software tools between model calls.

Nvidia and Oracle propose that allowing customers to scale CPU and GPU resources independently can significantly reduce unused capacity, a critical consideration given recent findings of low enterprise GPU utilization. VentureBeat Intelligence surveys indicate that a large percentage of organizations report GPU usage at 50% or below.

This development prompts enterprises to question whether a more balanced infrastructure approach can yield greater productivity from their investments. For organizations experiencing GPU idleness due to CPU bottlenecks, faster CPUs could offer a solution, whereas issues with remote services or application design might require different fixes.

Oracle's announcement details Vera CPUs, described by Nvidia as crucial for agent coordination, code execution, data processing, and supporting reinforcement learning. While Oracle cites Nvidia's benchmarks of up to 80% faster performance on agent-related CPU workloads compared to leading x86 processors, buyers are advised to treat these as vendor benchmarks rather than guaranteed application improvements.

OpenAI has announced its adoption of Vera on OCI for agent sandbox workloads, highlighting a clear benefit for foundation-model developers. However, the utility extends to enterprises running significant volumes of agent code or retrieval operations, even without training their own frontier models.

Oracle also plans to offer a smaller, eight-GPU Rubin NVL8 configuration using Intel Xeon processors, intended for training, fine-tuning, and inference. The OCI infrastructure is designed to accommodate evolving Nvidia platforms and shift between training and inference workloads, offering customers long-term flexibility.

Buyers are cautioned that full workload-specific pricing and performance evidence are still developing, and regional availability for Vera CPUs needs confirmation. The core decision for enterprises will rest on demonstrating how this enhanced infrastructure flexibility translates into tangible completed work and cost savings.