Oracle puts Vera Rubin and Vera CPUs on sale, with multi-gigawatt growth planned

Nvidia and Oracle said on October 27 that Vera Rubin NVL72 systems and Nvidia’s Vera CPUs are now generally available on Oracle Cloud Infrastructure (OCI). The two companies also plan to grow OCI Superclusters built on Vera Rubin to multi-gigawatt scale, according to the official announcement.

Racks, networking and a planned eight-GPU server

The release, issued from Oracle AI World in Las Vegas, describes OCI Superclusters as a combination of NVL72 rack-scale systems, Oracle’s Acceleron networking, Nvidia Spectrum-X Ethernet with Spectrum-6 switches, and BlueField-4 DPUs, all inside gigawatt-class power and cooling. Each rack links 72 GPUs over NVLink, and Nvidia says OCI’s network can span hundreds of thousands of GPUs. The expansion builds on the GB200 and GB300 NVL72 systems Oracle already runs.

Oracle also plans to add HGX Rubin NVL8, an x86 server with eight Rubin GPUs connected by NVLink. The planned configuration pairs them with two 64-core Intel Xeon 6 processors, 4TB of DDR5 memory and 6.4 Tbps of RDMA networking. No date was given for that system, and the release gives no price for any of the new offerings.

On the customer side, Nvidia points to OpenAI. The release says OpenAI trained and served GPT-6 Astra in Abilene on an OCI Supercluster with more than 100,000 Nvidia Blackwell GPUs. That describes the previous generation of hardware, so it shows the scale OCI already operates at rather than anything Vera Rubin has done yet.

Vera CPUs aimed at the agent bottleneck

The less expected part is the CPU. OCI Compute with Vera CPUs is pitched at the work that surrounds a model: tool calls, code execution and many agent tasks running at once. According to the release, high single-thread speed and memory bandwidth should cut the time GPUs sit idle between model calls. OCI Kubernetes Engine lets customers scale CPU and GPU capacity separately, and Acceleron supplies hardware-based isolation for untrusted code that an AI system generates.

Early-access customers, OpenAI among them, are testing Vera CPUs across AI workloads, the companies say. OpenAI’s Uday Ruddarraju, who leads compute technology there, is quoted saying the CPUs suit agentic sandbox workloads. Nvidia’s Ian Buck framed the pitch for operators: “AI factories must be productive, durable and fungible.”

Oracle joins Google in the Vera Rubin race

Oracle is not alone. Six days earlier, Nvidia’s own blog said Google Cloud would bring its A5X offering, built on Vera Rubin NVL72, to customers “next month”, despite a headline that spoke of general availability. Google Cloud said it plans to scale to as many as 960,000 Rubin GPUs across multiple sites. Oracle has not published a comparable GPU count; its public figures are “multi-gigawatt” and “hundreds of thousands of GPUs” for network reach, which are not the same measure.

On timing, Oracle’s wording is the firmer. It says Vera Rubin NVL72 is generally available now, while Google described a launch still ahead. Neither company has said how many racks are actually installed, so “available” does not tell a buyer how much capacity can be reserved today.

The CPU is where Oracle’s message differs most. Nvidia’s January coverage, reported in our earlier article, described Vera Rubin as a six-chip platform whose flagship server holds 72 Rubin GPUs and 36 Vera CPUs, with Nvidia claiming roughly five times the AI computing power of the previous generation. Selling Vera as a separate OCI compute option goes a step beyond that rack-bound design, and the Google Cloud material Nvidia published does not mention a standalone CPU instance. Whether customers will pay for general-purpose Arm-style agent hosts from an AI cloud, rather than from their usual CPU suppliers, is not answered by the release.

The Intel Xeon 6 choice for the planned HGX Rubin NVL8 shows Nvidia still ships x86 hosts for customers who want them, alongside its own Vera processors.

What the announcement leaves open

Everything about performance comes from the companies. The release carries no benchmarks for Vera Rubin on OCI or for Vera CPUs against rival processors, and no pricing, regions or instance names. The five-times figure from January is Nvidia’s claim and has not been independently tested in what we have seen. The multi-gigawatt target is a plan without a timeline, and the HGX Rubin NVL8 system is described only as planned.

For developers, the practical step is to check OCI’s console and sales channels for Vera Rubin and Vera CPU shapes, since the release does not say where in Oracle’s regions they are offered. Teams running agent fleets can compare the separate scaling of CPU and GPU in OCI Kubernetes Engine against their current setups once pricing appears.