NVIDIA and Oracle announced on October 27, 2026, that NVIDIA Vera Rubin NVL72 systems and Vera CPUs are generally available on Oracle Cloud Infrastructure (OCI). Alongside that launch, the companies outlined plans to expand OCI Superclusters to multi-gigawatt scale and introduce smaller HGX Rubin NVL8 systems.
Those are different milestones. Customers have two newly available infrastructure offerings, according to the companies’ joint announcement. The much larger capacity expansion and eight-GPU configuration belong to the roadmap, with no firm delivery dates disclosed.
The collaboration also addresses two different parts of an AI workload: GPUs for training and serving models, and CPUs for the tool calls, code execution and data retrieval that happen between model requests. The ability to scale those resources independently is arguably the most useful operational detail in the announcement.
What Is Available and What Is Planned
The release separates four elements:
- Vera Rubin NVL72 on OCI: Generally available, bringing NVIDIA’s 72-GPU rack-scale systems to Oracle’s cloud.
- Vera CPUs on OCI: Generally available as CPU compute resources for workloads including agent tool execution and code testing.
- Multi-gigawatt OCI Supercluster expansion: Planned, using Vera Rubin NVL72 systems.
- HGX Rubin NVL8 on OCI: Planned, providing eight Rubin GPUs in an x86-based server configuration.
Oracle’s announcement makes the same distinction between general availability and future expansion. These are two publications of the joint announcement, not separate independent confirmations.
For infrastructure buyers, I’d treat the general-availability statement as the starting point for a deployment conversation, then ask about the specific region, allocation and commercial terms. The release provides no regional availability list, pricing or customer capacity guarantees. General availability tells readers the products have moved beyond an early-access announcement; it does not reveal how much capacity a particular customer can secure.
Multi-Gigawatt Scale Describes an Expansion Goal
OCI’s planned Superclusters combine NVIDIA NVL72 racks, Oracle Acceleron networking, NVIDIA Spectrum-X Ethernet, Spectrum-6 switches and BlueField-4 data processing units. NVIDIA and Oracle describe that infrastructure as supported by gigawatt-class power and cooling.
Each NVL72 rack connects 72 GPUs through NVLink, NVIDIA’s scale-up interconnect. The companies say OCI’s broader, multiplanar network scales across hundreds of thousands of GPUs. The Vera Rubin expansion builds on existing OCI infrastructure that includes NVIDIA GB200 and GB300 NVL72 systems, according to the NVIDIA release.
Two details help put that scale in context.
First, a gigawatt is one billion watts. In this announcement, gigawatts describe the infrastructure’s power scale, not a measure of model throughput or a specific number of delivered GPUs. The release gives no final GPU count or site-by-site power allocation for the expansion.
Second, the rack and the wider cluster have different networking roles. NVLink connects the 72 GPUs within a rack; the broader network connects infrastructure across the cluster. A reference to hundreds of thousands of GPUs should not be read as a claim that all those GPUs share one rack-level NVLink connection.
My reading is that the power headline communicates the size of Oracle’s infrastructure ambition. It offers much less help in estimating a training job’s completion time. That requires workload-specific information about communication, utilization and the resources actually allocated to the job.
HGX Rubin NVL8 Offers a Different Deployment Shape
OCI’s planned HGX Rubin NVL8 configuration combines eight Rubin GPUs with Intel host processors. It gives customers a smaller GPU grouping and an x86-compatible option for training, fine-tuning and inference.
The published configuration includes:
| Component | Planned HGX Rubin NVL8 Specification |
|---|---|
| GPUs | Eight NVIDIA Rubin GPUs connected through NVLink |
| Host CPUs | Two Intel Xeon 6 processors, each with 64 cores |
| Host memory | 4 TB of DDR5 |
| Networking | 6.4 Tbps of high-speed RDMA networking |
| Availability | Planned; no firm delivery date disclosed |
Two 64-core processors give the server 128 CPU cores in total. The 4 TB figure describes its DDR5 host memory, not GPU memory; the announcement does not specify GPU memory capacity for this offering.
The network figure also needs careful reading. 6.4 Tbps means 6.4 terabits per second, not terabytes. Dividing by eight gives a nominal equivalent of 800 gigabytes per second before protocol overhead and other constraints. That is a unit conversion from the advertised networking figure, not measured application throughput.
RDMA stands for remote direct memory access, a mechanism for transferring data between systems’ memory with reduced CPU involvement. The release does not provide measured RDMA throughput or workload results for this configuration.
NVL8 also has a distinct host-CPU choice: its announced configuration uses Intel Xeon 6 processors, while Vera CPUs are a separate OCI offering. For teams with x86-based applications, that compatibility may be a practical reason to consider NVL8.
I’d avoid treating the eight-GPU and 72-GPU options as interchangeable building blocks. GPU count alone cannot establish equivalent performance, networking behavior or cost. The announcement supplies a useful configuration outline, but no like-for-like comparison between the two.
Why Vera CPUs Matter Between Model Calls
An AI agent can spend part of its workflow generating a response on a GPU, then shift to CPU-based work: running a tool, testing generated code or retrieving information. A slow tool step can delay the next model request.
Frequently Asked Questions
3 questions
1Which NVIDIA Systems Are Generally Available on OCI?
NVIDIA and Oracle say Vera Rubin NVL72 systems and Vera CPUs are generally available. HGX Rubin NVL8 remains a planned offering.
