AIdataNewsGlobal General: NVIDIA and Oracle Split Capacity Tracks with Vera Rubin NVL72 and Vera CPUs
Oracle Cloud Infrastructure and NVIDIA are moving capacity planning for large AI systems in two directions at once. A multi-gigawatt scale-up is described for OCI Supercluster capacity, while GPU supply and CPU supply are separated into distinct tracks through Vera Rubin NVL72 systems and Vera CPUs. The practical question for buyers is whether they can tune accelerated compute and ordinary compute resources according to different workload phases instead of treating them as one fixed bundle.
For readers tracking AI infrastructure announcements, the useful signal is not only the product name but the architectural choices: rack-scale GPU systems, independent CPU capacity, and networking components that connect large clusters. The following fact sheet separates what is being expanded, what is newly named, and what already has a customer reference.
Key disclosed facts, summarized for general readers
| Layer | Disclosed item | Reading note |
|---|---|---|
| Capacity target | multi-gigawatt expansion for OCI Supercluster capacity serving AI training, inference, and agentic applications. | Frames the update as infrastructure scale, not only a product addition. |
| GPU building block | NVIDIA NVL72 rack-scale systems, with NVLink connecting 72 GPUs inside each rack and an OCI network designed to extend across hundreds of thousands of GPUs. | Ties rack-level acceleration to cluster-level reach. |
| Expanded GPU system | Vera Rubin NVL72 systems are planned for wider use in OCI Supercluster capacity. | Watch whether multi-gigawatt planning becomes operational delivery. |
| CPU path | OCI Compute with NVIDIA Vera CPUs supports single-thread performance and memory bandwidth for agent tasks. | Makes CPU capacity a separate lever from GPU capacity. |
| Networking | Oracle Acceleron plus Spectrum-X Ethernet with Spectrum-6 switches and BlueField-4 DPUs. | Network isolation and data-path components are named alongside compute. |
| Existing base | GB200 systems and GB300 NVL72 systems remain part of the NVIDIA-based OCI infrastructure. | The Rubin expansion builds on an installed NVIDIA platform. |
| Software layer | CUDA is cited as the software foundation, with continued software optimization. | Software portability is part of the capacity story. |
| x86 deployment variant | HGX Rubin NVL8 connects eight Rubin GPUs through NVLink in an x86 server, with two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory, and 6.4 Tbps of RDMA networking. | Offers a more conventional server layout while retaining NVLink-based GPU connectivity. |
| Customer reference | OpenAI in Abilene trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 Blackwell GPUs. | Shows frontier-model work already occurring on comparable OCI infrastructure. |
The architectural idea: split CPU capacity from GPU capacity
The core planning point is separability. AI training and model inference usually draw attention because GPUs dominate the cost curve, but the disclosed plan also gives explicit weight to CPU work. In agent applications, the model may issue a call, then the system may execute tools, test code, or retrieve data before the next model step. If the CPU side cannot complete those steps promptly, the GPU side may spend time waiting rather than advancing the task.
The CPU offering addresses that waiting risk by pairing single-thread performance with memory bandwidth. Fast single-thread behavior can help individual tool calls and code-execution tasks move sooner. Memory bandwidth can help keep processor cores supplied when multiple operations run concurrently. OCI Compute is the service layer for these processors, and OCI Kubernetes Engine allows CPU and GPU resources to be scaled independently. Oracle Acceleron is also associated with hardware-based isolation for workloads running untrusted, AI-generated code.
Cluster composition: rack, network, power, cooling
The described OCI Supercluster combines NVIDIA NVL72 rack-scale systems, Oracle Acceleron, and Spectrum-X Ethernet networking. Within each rack, NVLink connects 72 GPUs. Across the broader environment, the OCI network is engineered to extend across hundreds of thousands of GPUs. The system also depends on large-scale power and cooling capacity for training and serving models.
The networking package is not generic. It names Spectrum-X Ethernet, Spectrum-6 switches, and BlueField-4 DPUs. These components are described as part of the infrastructure that carries data and control traffic for accelerated workloads. The software foundation is CUDA, with continued optimization intended to help customers adapt as models and workloads move across hardware generations.
The x86 route: HGX Rubin NVL8
The update also includes a planned HGX Rubin NVL8 configuration. This design places eight Rubin GPUs in an x86-based server and connects them through NVLink. The disclosed build combines two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory, and 6.4 Tbps of RDMA networking. Its stated role is a deployment path for AI training, fine-tuning, and inference, while preserving compatibility with x86-based applications.
For readers comparing infrastructure choices, that configuration is useful because it is not described only as a rack-scale system. It gives customers an eight-GPU layout, anchored by x86 host components, while retaining NVLink GPU interconnect.
Installed base and reference example
The planned expansion is not starting from an empty platform. The disclosed base includes GB200 systems and GB300 NVL72 systems, which are part of OCI's existing NVIDIA-based infrastructure. CUDA and ongoing software optimization are cited as continuity mechanisms as models and workloads shift across hardware generations.
A concrete reference point appears in Abilene, where OpenAI trained and served GPT-6 Astra on an OCI Supercluster equipped with more than 100,000 Blackwell GPUs. Customers granted early access to NVIDIA Vera CPUs on OCI, including OpenAI, are also evaluating the processors across AI workloads.
FAQ: What the NVIDIA and Oracle infrastructure update covers
What are the two main compute additions?
Vera Rubin NVL72 systems and Vera CPUs. The first is linked to expanded GPU capacity in OCI Supercluster plans; the second is linked to CPU capacity for agent-heavy tasks.
Why are GPU and CPU resources treated separately?
The disclosed rationale is workload matching. GPU capacity handles model training and inference, while CPU capacity handles tool execution, code testing, and data retrieval around those model calls. Independent sizing can reduce idle capacity.
Which cluster scale figures are mentioned?
The text includes multi-gigawatt expansion, 72 GPUs in each rack, and an OCI network designed for hundreds of thousands of GPUs. The OpenAI example mentions more than 100,000 Blackwell GPUs.
What is HGX Rubin NVL8?
A planned x86-based configuration that connects eight Rubin GPUs through NVLink. The disclosed server profile includes two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory, and 6.4 Tbps of RDMA networking.
Which existing NVIDIA platforms are mentioned?
GB200 systems and GB300 NVL72 systems are listed as part of the current infrastructure. CUDA is described as the software foundation.
What should readers watch next?
Operational delivery of multi-gigawatt capacity, general availability wording for the named systems, adoption of independent CPU and GPU scaling, and how x86 compatibility for HGX Rubin NVL8 is used in training, fine-tuning, and inference workloads.
Bottom line for the General briefing
The announcement can be read as a modular scaling plan. NVIDIA and Oracle are describing a stack where GPU scale, CPU responsiveness, networking, isolation, software compatibility, and existing deployments each have named components. The relevant numbers are cluster-level: multi-gigawatt ambition, 72 GPUs per rack, hundreds of thousands of GPUs in the network design, and more than 100,000 Blackwell GPUs in the disclosed OpenAI example. The relevant product names are Vera Rubin NVL72, Vera CPUs, OCI Compute, Kubernetes Engine, HGX Rubin NVL8, GB200, GB300 NVL72, CUDA, Oracle Acceleron, Spectrum-X Ethernet, Spectrum-6, and BlueField-4.
For a global AI and data audience, the main value is interpretive clarity: this is not simply a faster chip notice. It is a capacity architecture note, showing how cloud providers may let customers tune GPU and CPU resources according to different workload phases while keeping the underlying NVIDIA software and Oracle cloud environment aligned.