In one sentence
On October 27, 2026, NVIDIA Vera CPUs reached GA on OCI for CPU-intensive agent tasks between model calls; teams can evaluate them in five steps: region check, instance deploy, x86 baseline, independent CPU/GPU scaling via OKE, and per-task cost accounting.
When to use this guide
- Teams running agent workloads on OCI that need CPU-heavy work between model calls: tool execution, test code, data retrieval.
- Enterprises with low GPU utilization that want to scale CPU and GPU resources independently and cut GPU idle time between model calls.
- Architecture and cost teams running side-by-side evaluation and end-to-end per-task cost modeling.
Context: Oracle and NVIDIA announced at Oracle AI World (Las Vegas, October 27, 2026) that NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPU reached general availability on OCI.
Steps
- Check region availability: confirm Vera CPU instance availability in your target region in the OCI Console; Oracle says rollout expands to selected regions this quarter.
- Deploy the instance: provision Vera CPU on bare metal or flexible VMs, and verify the high single-thread performance and memory-bandwidth configuration that matches agent code-execution and tool-call loads.
- Benchmark against x86: run the same agent workflow on existing mainstream x86 processors as a baseline, measuring task completion rate, end-to-end latency, and CPU/GPU consumption. VentureBeat reports NVIDIA claims Vera is up to about 80 percent faster than mainstream x86 on agent-related CPU loads — a vendor platform benchmark, not an OCI instance measurement.
- Scale independently with OKE: use Oracle Kubernetes Engine to scale CPU and GPU resources independently; Oracle says this reduces GPU idle time between model calls. OpenAI Compute CTO Uday Ruddarraju said OpenAI is bringing Vera CPU into OCI production to support agent sandbox workloads.
- Account for total cost: add model API fees and Oracle Acceleron hardware-isolation overhead to compute end-to-end cost per task; Oracle says Acceleron provides hardware-level isolation for running untrusted AI-generated code.
Notes
- The up-to-about-80 percent figure is a vendor platform benchmark cited by NVIDIA via VentureBeat, not an OCI instance measurement; use your own workload baseline.
- Regional coverage is described by Oracle as expanding to selected regions this quarter; the exact region list and quota follow the OCI Console and official docs.
- OpenAI's production rollout is a public statement by its Compute CTO; deployment scale and timing on OpenAI's side were not disclosed.
- Oracle Acceleron SKUs, pricing, and availability domains follow Oracle's official documentation.