NVIDIA Vera Rubin and Vera CPUs on OCI: What Is Available Now

NVIDIA and Oracle announced general availability of Vera Rubin NVL72 systems and NVIDIA Vera CPUs on Oracle Cloud Infrastructure on October 27, 2026. They also outlined a larger capacity expansion, but that future buildout should not be confused with the infrastructure available now.

The useful change for AI teams is the combination of GPU capacity for models and separate CPU capacity for the tools those models use. Buying more of one does not automatically remove a bottleneck in the other.

What is generally available on OCI?

The joint announcement separates three developments:

Infrastructure	Announced status
Vera Rubin NVL72	Generally available on OCI
NVIDIA Vera CPUs	Generally available on OCI
Multi-gigawatt Supercluster expansion	Planned
HGX Rubin NVL8 with x86 CPUs	Planned introduction

The proposed NVL8 configuration uses eight Rubin GPUs with Intel Xeon 6 CPUs. It offers a smaller deployment unit than the 72-GPU rack-scale design, while retaining an x86 host environment.

The announcement does not provide a regional availability table or a complete hourly price list for the new offerings. Teams still need to confirm capacity, region, quota and commercial terms with Oracle. General availability is a product milestone, not a guarantee of spare capacity in every region.

Why a CPU release matters for AI agents

NVIDIA's Vera architecture overview describes an Arm-compatible processor with 88 Olympus cores, 176 threads and high-bandwidth LPDDR5X memory. It is designed for work such as code execution, data processing and running many software environments alongside models.

Consider a coding agent that generates a patch, runs a test suite and then reads the result before deciding what to do next. Faster inference helps the first step. It cannot make a slow compiler, overloaded test runner or blocked retrieval job finish sooner.

This makes CPU time part of the cost of delivering an agent result. If a service keeps expensive inference capacity reserved while tools repeatedly stall, the application can spend more without completing proportionally more work. The practical measurement is elapsed time and cost per successful task, with tool execution included.

Vera's published performance claims are NVIDIA's comparisons, not an independent benchmark of an application running on OCI. A team should test its own dependencies and runtime before assuming an Arm migration or a newer CPU will improve the whole workflow.

Networking is part of the execution path

Oracle's Acceleron documentation explains how its SmartNIC handles networking and infrastructure processing outside application CPUs. It also describes hardware-enforced isolation and policy-based communication controls.

Those capabilities matter when one agent task touches a sandbox, a retrieval service and a model server. The machines need to exchange data, and the sandbox needs boundaries even when the generated code appears harmless. Infrastructure isolation supports that design; it does not replace application permissions or careful handling of credentials.

What to measure before moving a workload

Start with a trace of the existing agent rather than a hardware specification. How much time goes into model generation, tool execution, data movement and waiting in queues? Which component becomes saturated when several users arrive together?

That breakdown determines whether separate CPU capacity helps. A workflow dominated by a remote API's response time may see little benefit. A workload spending most of its time compiling or processing data could have a more relevant bottleneck to investigate.

Oracle's release gives teams another way to provision those stages. The next useful result will be evidence that the complete agent finishes more reliable work for the same budget, including the time it spends outside the model.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain infrastructure launches and technical tradeoffs in content that developers and business buyers can understand and share.

Book a 30-minute call.