On October 27, 2026, NVIDIA and Oracle announced general availability of Vera CPU instances and Vera Rubin NVL72 systems on Oracle Cloud Infrastructure (OCI) Superclusters. Vera is offered as a standalone CPU service, while Oracle and NVIDIA described a multi-gigawatt expansion of GPU-based Superclusters as a plan.
What the standalone Vera CPU is for
Oracle Cloud describes Vera as an 88-core Arm CPU for agent orchestration, sandboxing, retrieval and GPU-fleet control planes. That makes the offering distinct from a GPU system: a customer can use Vera CPU instances without treating them as part of a Rubin GPU configuration. NVIDIA attributes a memory-bandwidth capability of 1.2 TB/s to Vera.
Oracle also describes OCI Kubernetes Engine as able to scale CPU and GPU resources independently. That lets customers adjust the two resource pools separately. Oracle says its Acceleron technology provides hardware-based isolation for workloads running untrusted AI-generated code.
The reported Supercluster network stack combines NVIDIA Spectrum-X Ethernet, Spectrum-6 switches and BlueField-4 DPUs with Oracle Acceleron networking. These are descriptions of the announced infrastructure.
What the Vera benchmark results measure
Tom’s Hardware ran its Vera tests on a local server, not on OCI. Its published figures therefore describe the tested local configurations, rather than Vera instances running in Oracle’s cloud. For Stream Triad, higher bandwidth is better; for PyBench and SQLite, lower test times are better.
| Test and condition | CPU configuration | Result |
| Stream Triad, OpenMP spread | 2× Vera | 2,343 GB/s |
| Stream Triad, no OpenMP spread label | 2× Vera | 1,646 GB/s |
| Stream Triad, no OpenMP spread label | 2× Intel Xeon 6980P | 1,182 GB/s |
| Stream Triad, OpenMP spread | 2× Intel Xeon 6980P | 949 GB/s |
| PyBench, average test time | 2× Vera | 499 ms |
| PyBench, average test time | 2× AMD EPYC 9755 | 520 ms |
| PyBench, average test time | 2× AMD EPYC 9965 | 571 ms |
| PyBench, average test time | 2× Intel Xeon 6980P, both listed configurations | 585 ms |
| SQLite, 128 threads/copies | 2× Vera | 5.4 s |
| SQLite, 128 threads/copies | 2× AMD EPYC 9965 | 5.9 s |
| SQLite, 128 threads/copies | 2× AMD EPYC 9755 | 7.8 s |
| SQLite, 128 threads/copies | 1× Intel Xeon 6980P | 12.4 s |
| SQLite, 128 threads/copies | 2× Intel Xeon 6980P | 20.6 s |
The Stream Triad figures vary with the OpenMP setting: the Vera result is 1,646 GB/s without the spread label and 2,343 GB/s with it, while the Xeon results are 1,182 GB/s and 949 GB/s, respectively. The comparison depends on the setting, so a single multiplier would obscure that difference.
OpenAI’s stated use and the planned Rubin system
OpenAI was named among OCI customers with early access. OpenAI compute CTO Uday Ruddarraju said the company was bringing Vera into production on OCI for code execution and tools in agentic sandbox workloads. His statement describes an intended production transition.
The separately planned HGX Rubin NVL8 configuration pairs eight Rubin GPUs with two 64-core Intel Xeon 6 CPUs. Its reported specifications also include 4 TB of DDR5 memory and 6.4 Tbps of high-speed RDMA networking. Oracle and NVIDIA described the broader OCI Supercluster expansion as targeting multi-gigawatt scale.
