Oracle and NVIDIA announced general availability of NVIDIA Vera Rubin NVL72 systems and Vera CPUs on Oracle Cloud Infrastructure on October 27, 2026. The launch adds GPU infrastructure for training and inference alongside standalone CPU capacity for the tool execution, code testing and retrieval work that AI agents perform between model calls. The companies also announced plans to expand OCI Superclusters to multi-gigawatt scale. (NVIDIA announcement)
These announcements carry different purchasing implications. General availability applies to Vera Rubin NVL72 and Vera CPUs. The larger infrastructure buildout and a smaller HGX Rubin NVL8 configuration are planned additions.
For infrastructure buyers, the consequential detail is the ability to provision CPU services separately from GPU infrastructure. An agent that spends substantial time running tools needs a different resource mix from a service that mainly generates tokens.
What Is Available Now, and What Is Planned?
The announcements describe four related offerings and expansion plans:
| Offering or plan | Announced status | What it provides |
|---|---|---|
| Vera Rubin NVL72 on OCI | Generally available | Rack-scale systems connecting 72 GPUs through NVIDIA NVLink |
| OCI Compute with Vera CPUs | Generally available, with regional rollout language | Standalone CPU capacity through bare-metal and Flex virtual machine instances |
| Multi-gigawatt OCI Supercluster expansion | Planned | Additional Vera Rubin NVL72 infrastructure, networking, power and cooling |
| HGX Rubin NVL8 | Planned | An eight-GPU Rubin server configuration using Intel Xeon 6 CPUs |
NVIDIA’s release establishes the product and expansion status. Oracle’s CPU-specific announcement supplies the deployment options.
General availability still needs a regional check. Oracle’s CPU post announces GA but closes by saying Vera CPU instances will be available in select regions “this quarter.” It also says 150,000 CPUs will initially be made available across Oracle’s regions. Read together, those statements describe a product launch with a regional rollout, not proof that the entire announced allocation is immediately provisionable everywhere.
The reviewed announcements provide no regional inventory, pricing, customer quotas or firm delivery dates for the planned configurations. I would treat the GA label as the point to start a capacity discussion, not as a substitute for confirming availability in the region where the application must run.
Vera CPUs Give Agent Services Their Own Compute Pool
Vera serves two roles in this launch: it is a host CPU in Vera Rubin NVL72, and it is available as a standalone OCI compute platform. The second role lets customers run agent services without acquiring Rubin GPU capacity alongside them.
Oracle’s detailed Vera CPU post lists the following specifications for each processor:
- 88 Arm-compatible NVIDIA Olympus cores
- 176 threads, enabled by NVIDIA Spatial Multithreading
- Up to 1.2 TB/s of memory bandwidth
- 3.4 TB/s of bisectional bandwidth in the second-generation NVIDIA Scalable Coherency Fabric
The two bandwidth figures describe different parts of the processor. Memory bandwidth concerns movement between memory and the processor; the coherency-fabric figure concerns the internal interconnect that supports coherent operation. They should not be added together or treated as interchangeable measures.
For OCI’s published dual-processor configuration, Oracle lists:
- 176 CPU cores and 352 threads
- 1 TB of LPDDR5X system memory
- 7.68 TB of local NVMe storage
- 200 Gbps Acceleron networking
These are system-level specifications, separate from the per-processor figures above. Oracle offers both bare-metal and Flex VM deployment options; buyers should verify the resources assigned to the particular instance shape they intend to order.
The Arm-compatible architecture also creates a practical migration checkpoint. Teams moving existing x86 services should check container images, compiled extensions and native dependencies before assuming that an orchestration or sandbox stack will transfer unchanged. That is a compatibility consideration inferred from the architecture, not a migration problem reported in the announcement.
Independent Scaling Targets the Gaps Between Model Calls
OCI Kubernetes Engine enables customers to scale CPU and GPU resources independently, according to the NVIDIA announcement. Combined with standalone Vera instances, this gives operators separate capacity pools for model inference and the services surrounding it.
An agent workflow illustrates why that matters:
- A model generates a tool request.
- A CPU service executes code, queries data or runs a test.
- The service returns its result.
- The model resumes inference using that result.
A slow second step increases the time before the next model call. Across many concurrent agents, CPU queues can become a bottleneck even when sufficient inference capacity exists.
Independent scaling gives operators a way to add tool workers or sandbox capacity without expanding the GPU pool at the same time. Conversely, rising inference demand can justify additional GPU capacity without requiring a matching increase in every supporting CPU service.
This is a resource-allocation capability, not a guarantee of higher GPU utilization. A shared inference service may already fill one agent’s waiting time with work from other requests. Faster CPU execution can still improve that agent’s completion time, but its effect on aggregate GPU use depends on scheduling and workload concurrency.
My view is that this separation is the launch’s most useful operational feature for agent developers. It lets capacity decisions follow the actual bottleneck. Before buying more hardware, I would measure CPU queue time, tool-execution latency, model-serving latency and completed tasks per second. A slow external API will not become faster because the orchestration service gains more cores.
Performance Claims Need Workload-Level Checks
Oracle reports NVIDIA’s claim of up to 80% faster agentic CPU performance than leading x86 CPUs. Its CPU post explicitly says the result comes from NVIDIA platform benchmarks, not OCI instance-level testing, and that performance varies with workload and configuration. The reviewed material does not supply enough detail to map that percentage to a particular customer application. (Oracle CPU announcement)
Oracle’s broader launch post also advertises up to 30 times higher inference throughput per watt and four times the training performance for Vera Rubin compared with the previous Blackwell generation. These are vendor claims, with different metrics, not independent measurements of an OCI deployment.
The practical comparison is therefore a completed workload under defined conditions: model, precision, context length, concurrency, CPU task mix and infrastructure cost. Peak platform figures alone cannot establish the cost of completing an agent task.
Acceleron Adds Hardware Isolation for Generated Code
Oracle Acceleron provides hardware-based isolation for workloads running untrusted, AI-generated code, according to NVIDIA’s release.
Generated code should be treated as untrusted input when an agent can execute it. A sandbox may need to run a test suite, inspect files or invoke a tool without gaining access to unrelated workloads or sensitive infrastructure.
Hardware isolation is relevant because increasing sandbox density also increases the importance of the boundary between workloads. However, the announcement does not detail Acceleron’s threat model, the isolation boundary for each instance type or the configuration required to apply it.
I would make those details part of the deployment review. Buyers should ask which components share a host, how network access is restricted, where credentials are exposed and how execution environments are reset. Isolation is one layer; permissions, secrets handling, egress policy and execution limits still determine what generated code can reach.
The release supports a hardware-isolation claim. It does not provide enough detail to rank that protection against competing cloud sandbox designs.
OpenAI Says Vera Is Entering Production for Sandboxes
Frequently Asked Questions
3 questions
1Are Vera Rubin and Vera CPUs available on OCI now?
Oracle and NVIDIA announced general availability on October 27, 2026. Oracle’s Vera CPU post also specifies a select-region rollout during the quarter, so buyers should confirm local capacity.
