Oracle and NVIDIA used Oracle AI World on October 27, 2026 to announce that NVIDIA Vera CPUs and Vera Rubin NVL72 systems are now generally available on Oracle Cloud Infrastructure (OCI). Oracle says it is making an initial 150,000 Vera CPUs available and calls itself "the first hyperscaler to bring Vera CPUs to market." OpenAI is the headline customer: its compute CTO says it is "bringing Vera into production on OCI" for agentic sandbox workloads.
This post sorts the announcement into what is shipping, what is planned and what is only a vendor claim. It follows our earlier coverage of Vera CPUs in Europe for agent sandboxes and the independent memory measurements from Chips and Cheese.
TL;DR: what is real, what is planned
| Item | Status per Oracle and NVIDIA |
|---|---|
| NVIDIA Vera CPUs on OCI | Generally available; initial deployment of 150,000 CPUs |
| Vera Rubin NVL72 on OCI | Generally available |
| OpenAI on Vera | "Bringing Vera into production" for agent sandboxes; early-access customer |
| HGX Rubin NVL8 (two 64-core Intel Xeon 6, 4 TB DDR5) | Planned, not shipping |
| Multi-gigawatt Vera Rubin expansion | Planned |
| Pricing | Not stated in the press release or the Oracle blog we read |
Why does an AI cloud care about CPUs?
The press release makes the argument plainly: agents "call tools, execute code and retrieve information," and slow execution of those CPU-heavy steps "can delay results and leave GPUs waiting between model calls." The pitch is that fast single-thread performance and memory bandwidth reduce GPU idle time, and that OCI Kubernetes Engine lets customers scale CPU and GPU capacity independently.
This is the same logic behind the agent-sandbox shift we have tracked: the model call is only one part of an agent loop, and the rest runs on ordinary compute. Whether a custom Arm CPU beats the x86 instances people use today is exactly what independent testing has to show, and our critique of the Vera whitepaper explains why vendor numbers deserve caution.
What is in the Vera CPU?
Per Oracle's OCI blog, Vera is NVIDIA's first fully custom CPU, with a ground-up core called Olympus "built around the agent loop rather than around raw floating point throughput." Each CPU has 88 cores and 176 threads through spatial multithreading and up to 1.2 TB/s of LPDDR5X memory bandwidth. Oracle lists the target workloads:
- agentic orchestration and tool serving;
- sandbox execution at density, which Oracle says drew the most direct pull from AI-native companies;
- control plane and orchestration for GPU fleets;
- vector search and retrieval;
- in-memory databases and columnar analytics;
- general operational services on the same Arm platform.
Oracle says the database result is where "the independent benchmarks are strongest for Vera," but the post does not name the benchmarks, so we cannot assess that claim here.
What did OpenAI actually say?
Two OpenAI executives are quoted. Uday Ruddarraju, chief technology officer of compute, said Vera CPUs "deliver the world-class loaded performance needed to execute code and run tools across our rapidly growing agentic sandbox workloads," and that OpenAI is bringing Vera into production on OCI. Sreekanth Pothanis said Vera Rubin on OCI gives OpenAI "the computing performance and scale to accelerate model training and serve increasingly demanding AI workloads."
Both are vendor-published quotes without performance numbers. They confirm intent and early-access use, not a measured speed-up. The press release describes OpenAI as one of the customers with early access, so "in production" is a stated plan to scale, not a disclosed fleet size.
Oracle also repeats that OpenAI trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs in Abilene, as announced at launch; see our GPT-6 Astra launch coverage and the earlier piece on Oracle and OpenAI's ChatGPT and Codex workspace on OCI.
What about Vera Rubin NVL72 and Rubin NVL8?
Oracle says Vera Rubin NVL72 delivers up to 30 times higher inference throughput per watt and a 4x leap in training performance compared with the prior Blackwell generation, using HBM4 and NVLink 6. Those are NVIDIA-sourced comparisons; our analysis of NVIDIA's Vera Rubin fewer-GPUs claim covers how such figures are framed.
The planned HGX Rubin NVL8 pairs eight Rubin GPUs with two 64-core Intel Xeon 6 CPUs, 4 TB of DDR5 and 6.4 Tbps of RDMA networking, in an x86 configuration for customers who need x86 compatibility. It is explicitly a plan.
Networking runs on Oracle Acceleron: 102.4 Tb/s NVIDIA Spectrum-X Ethernet switches, ConnectX-9 SuperNICs and BlueField-4 DPUs. Oracle describes a multiplanar fabric using Multipath Reliable Connection, with two-tier networks supporting more than 131,000 GPUs. Oracle also says Acceleron provides hardware-based isolation for workloads running untrusted, AI-generated code, which matters for sandbox use.
How does this fit the wider Vera rollout?
Oracle is not the only cloud adding Vera Rubin. Google Cloud's A5X Vera Rubin general availability and CoreWeave's first Vera Rubin cluster came earlier on the GPU side. The new part here is the CPU: Oracle's claim to be first among hyperscalers with Vera as a standalone compute product, which is a claim we could not verify against other clouds' announcements. Oracle is also adding Google's chips, as covered in Oracle's Google TPU plans for 2027, so the picture is a multi-vendor cloud, not an NVIDIA-only one.
AMD's counter is the CPU-heavy rack we covered in the Helios agentic CPU piece; the competition for agent-orchestration compute is now explicit.
How should you read 150,000 CPUs?
The number sounds large, but it needs context. Oracle calls it an "initial deployment," which is a count of processors, not of servers or sockets sold to a single customer, and the announcement does not say how much of it is already allocated to early-access customers such as OpenAI. With 88 cores per CPU, 150,000 CPUs would be on the order of 13 million cores if every unit were counted, but that is our arithmetic from Oracle's two figures, not a number either company published, and it says nothing about delivered performance.
It also helps to compare it with the GPU side. Agent sandboxes are usually many small, short-lived jobs: clone a repository, run tests, execute a script, tear the environment down. They are bounded by single-thread speed, memory bandwidth and how many isolated environments fit on a host, not by matrix-multiply throughput. That is why Oracle stresses "sandbox execution at density." The open question for buyers is cost per sandbox-second against the x86 alternatives, and for that the missing price list matters more than the headline CPU count.
Finally, the timing is notable. NVIDIA first presented Vera as a CPU for agents earlier this year, independent testers have since begun publishing memory measurements, and now a hyperscaler is offering it as rentable compute. The gap between unveiling and general availability is short for server silicon, which is part of why vendor claims and early-access quotes are all we have so far.
Other items in Oracle's blog
Oracle lists several software pieces alongside the hardware: NVIDIA Dynamo bringing disaggregated inference to Oracle Kubernetes Engine; Oracle Database Index Service v2 using NVIDIA cuVS for 5x faster index builds compared with CPU offloading (Oracle's figure); AI Accelerator Packs with NVIDIA cuOpt; NVIDIA Nemotron powering a cohorting agent in Oracle Health Data Intelligence, planned for EMEA with wider deployment in Q1 next year; and joint optimization of Oracle AI Media Models on GB300, available in the US only.
What should builders do with this?
If you run agent sandboxes today. Treat Vera on OCI as something to benchmark, not adopt on faith. Measure tool-call latency, cold-start time and cost per task on your own workload against your current x86 instances. Oracle has not published prices in the materials we read.
If you build on Arm. The platform is Arm, so check that your containers and toolchains build for arm64.
If you care about security. Sandboxes that run AI-generated code are an attack surface. Oracle's hardware isolation claim is worth testing, and our explainer on NVIDIA's agent-security hardware covers the adjacent layer. Teams that need to stop agents before they take dangerous actions can look at AgentBeam, from the explainx.ai team.
If you track the market. OpenAI's quote adds to evidence that agent workloads are shifting demand toward CPUs, not only GPUs. Independent verification will come from measurements like the ones in our Chips and Cheese write-up.
What is not confirmed
- No prices, instance shapes or regional availability appear in the press release or the Oracle blog we read; one search result suggested select regions this quarter, which we could not confirm in the primary sources.
- No performance numbers accompany the OpenAI quotes.
- The "independent benchmarks" Oracle cites for database workloads are not named.
- Rubin NVL8 and multi-gigawatt scale are plans covered by forward-looking disclaimers.
- "First hyperscaler" is Oracle's own claim.
We will update this post as pricing, regions or independent tests appear.
Related reading
- NVIDIA Vera CPUs in Europe: agent sandboxes and early-testing claims
- Vera CPU measured: 1.2 TB/s memory and 309 ns cross-socket
- NVIDIA Vera CPU whitepaper critique
- Google Cloud A5X Vera Rubin NVL72 general availability
- Oracle and Google TPUs on OCI in 2027
- Oracle, OpenAI ChatGPT and Codex workspace on OCI
- GPT-6 Astra launch, benchmarks and pricing
Details reflect Oracle's and NVIDIA's October 27, 2026 announcements and may change.