Oracle Becomes First Hyperscaler to Sell Nvidia Vera CPU Alongside Rubin NVL72

The competition among cloud providers for artificial intelligence workloads has mostly been counted in GPUs, but the processor that feeds those accelerators is becoming a product in its own right. At Oracle AI World in Las Vegas on October 27, 2026, Nvidia and Oracle put that shift on stage. Nvidia Vera Rubin NVL72 systems and Nvidia Vera CPUs are now generally available on Oracle Cloud Infrastructure, with an initial deployment of 150,000 Vera CPUs across Oracle regions, according to Oracle.

Vera is not a supporting part in the usual sense. Oracle describes it as Nvidia's first fully custom CPU, built on the ground-up Olympus core and designed around the agent loop rather than floating-point throughput. The Register reported on August 1, 2026 that Vera is the follow-on to Nvidia's Grace CPU and carries 88 custom Armv9.2 Olympus cores with 176 threads, support for up to 1.5 TB of LPDDR5X memory, and availability as a standalone platform independent of Nvidia's GPUs. That independence is the part Oracle is now selling.

The rack system is equally central to the announcement. Each Vera Rubin NVL72 rack connects 72 GPUs through NVLink, and OCI Superclusters pair those racks with Oracle Acceleron networking, Nvidia Spectrum-X Ethernet using Spectrum-6 switches and BlueField-4 DPUs, and gigawatt-class power and cooling. Oracle says its multiplanar network can scale across hundreds of thousands of GPUs. The build-out sits on top of existing GB200 and GB300 NVL72 capacity and Nvidia's CUDA software stack, which matters because customers are being asked to extend an existing Nvidia environment rather than start from a clean sheet.

OpenAI trained and served GPT-6 Astra in Abilene, Texas on an OCI Supercluster with more than 100,000 Nvidia Blackwell GPUs, according to Nvidia. Sreekanth Pothanis, an engineering leader for compute infrastructure at OpenAI, said Vera Rubin on OCI gives the company the performance and scale to accelerate training and serve demanding workloads. Oracle's Mahesh Thiagarajan and Nvidia's Ian Buck framed the deal around AI factories that must stay productive, durable and fungible as models change. The new claim is narrower and more consequential: Oracle says it is the first hyperscaler to bring Vera CPUs to market.

Key Facts

NVIDIA Newsroom reported on October 27, 2026 that both Vera Rubin NVL72 and Vera CPUs are generally available on OCI and that the companies plan to expand OCI Superclusters with Vera Rubin NVL72 systems to multi-gigawatt scale. Oracle Cloud Infrastructure Blog stated on October 27, 2026 that OCI Compute with Vera CPUs is generally available and that Oracle is the first hyperscaler to bring Vera CPUs to market, with an initial 150,000 CPUs made available across Oracle regions to meet demand for agentic AI workloads.

The published specifications are substantial. Oracle lists each Vera CPU at 88 Olympus cores, 176 threads, up to 1.2 TB/s of LPDDR5X memory bandwidth and 3.4 TB/s of bisectional fabric bandwidth. A two-socket configuration offers 176 cores, 352 threads and 1 TB of LPDDR5X system memory. Oracle also says Nvidia reports up to 80% faster agentic CPU performance than leading x86 CPUs, but that figure is a vendor benchmark rather than an independent test.

Customers can buy the chips as Bare Metal instances or Flex Virtual Machine instances, so agentic services can run independently of GPUs. OCI Kubernetes Engine lets CPU and GPU capacity scale separately, a detail that matters for workloads that spend more time coordinating tools and data than running matrix math. Oracle says Vera CPUs on OCI will be available in select regions this quarter. It has not published a region-by-region list, pricing, or an independent utilization study.

Data Phoenix reported on October 27, 2026 that Oracle has not named the exact regions where Vera CPU instances can be ordered immediately, though one Oracle product post says select-region availability will occur this quarter. The outlet also noted that Oracle did not clarify whether the 150,000 figure counts physical processors, provisionable instances or another unit of capacity. The multi-gigawatt Supercluster expansion and the HGX Rubin NVL8 configuration are future offerings, not capacity already deployed, with no deployment date or location for the Superclusters and no availability date for NVL8.

The future NVL8 design is a smaller alternative to the rack-scale system. It pairs eight Rubin GPUs joined by NVLink with an x86 server carrying two 64-core Intel Xeon 6 CPUs, 4 TB of DDR5 memory and 6.4 Tbps of RDMA networking. Nvidia positions it for training, fine-tuning and inference in a smaller form factor. The Register reported on August 1, 2026 that Vera also supports single- and dual-socket configurations and that the Vera CPU Superchip pairs two Vera CPUs over NVLink-C2C at 1.8 TB/s of bidirectional bandwidth. Those details support Oracle's pitch of a general-purpose cloud CPU.

Analysis

Oracle has bought differentiation rather than raw capacity. OCI already had GB200 and GB300 NVL72 systems, so GPU capacity alone is not the differentiator. Selling Vera as a standalone CPU gives OCI a claim that its rivals do not yet have, at least according to Oracle's first-hyperscaler statement. It also gives customers a way to add CPU capacity for agentic sandboxes without buying another GPU rack.

What this really means is that Oracle is using Nvidia's first custom CPU to reposition OCI for agentic workloads, while Nvidia gets a hyperscale partner that can show Vera running as standalone infrastructure rather than only as a companion to GPUs. Oracle gains a differentiated processor and a closer tie to Nvidia's roadmap. Nvidia gains a distribution channel for a CPU business that must prove it can stand apart from the GPU business.

The evidence for the performance story is still thin. Nvidia's 80% figure is a vendor benchmark, and Data Phoenix stressed that the launch material includes no independent OCI workload test, no pricing and no total-cost comparison. The 150,000-unit figure is also ambiguous because Oracle has not said whether it counts physical processors, provisionable instances or another measure. Those gaps do not make the launch unimportant. They do mean that the strongest claims in the announcement are capacity and availability claims, not proven efficiency claims.

The competitive read is equally measured. Vera's Armv9.2 design, 88 cores and high memory bandwidth give Oracle a credible alternative to x86 instances on paper, and the ability to scale CPU and GPU capacity separately is useful for agentic services that are not always GPU-bound. But no independent OCI comparison against x86 instances has been published, and the 80% figure is a vendor benchmark. The figures cited in the announcement show why the comparison is not settled: the advantage is about agentic CPU performance, a workload class that is still being defined, and it comes from the vendor whose chip is being sold.

Why It Matters

For cloud buyers, the practical change is that agentic workloads can be provisioned on a CPU product that Oracle says is designed for the agent loop, with 88 cores, 176 threads and up to 1.2 TB/s of memory bandwidth per CPU. Bare Metal and Flex VM options give teams different levels of control, and OKE's separate CPU and GPU scaling avoids forcing every agentic service into a GPU allocation. The limitation is availability: select regions this quarter is not the same as global capacity.

For the cloud market, the first-hyperscaler claim gives Oracle a marketing advantage that is easy to understand and hard for rivals to match without their own Vera supply. It also deepens Oracle's dependence on Nvidia at a time when cloud providers are trying to differentiate their silicon stacks. The multi-gigawatt Supercluster plan raises the stakes further, but it is a plan, not deployed capacity, and Oracle has not given a date or location for it.

For Nvidia, Vera is a test of whether it can sell a CPU on its own merits. The Register reported on August 1, 2026 that Vera supports PCIe 6.4 and CXL 3.1 and uses eight LPDDR5X memory controllers with SOCAMM2 modules up to 9,600 MT/s, features that position it for modern data center designs. Oracle's 150,000-unit deployment is a meaningful reference, but the absence of pricing and independent benchmarks means the market still lacks the data needed to judge total cost.

Next Up

The immediate next step is regional availability. Oracle says Vera CPUs will be available in select regions this quarter, but it has not named those regions or published prices. The multi-gigawatt Supercluster expansion and HGX Rubin NVL8 also lack deployment dates and locations. Customers evaluating OCI for agentic workloads should watch for the region list, instance pricing and any independent benchmark that tests Vera against x86 instances on real OCI workloads.

OpenAI's production use is the other signal to watch. Uday Ruddarraju, OpenAI's compute CTO, said Vera on OCI delivers the loaded performance needed for its agentic sandbox workloads and that OpenAI is bringing Vera into production at scale. That establishes a named customer deployment, but it does not validate Oracle's performance claims. The next credible test will be published utilization, pricing and total-cost data, not another launch-stage benchmark.