NVIDIA and Oracle used Oracle AI World on October 27 to put NVIDIA's newest data center hardware into general availability on Oracle Cloud Infrastructure. The companies said Vera Rubin NVL72 rack-scale systems and NVIDIA Vera CPUs can now be ordered on OCI, and that Oracle plans to grow its Vera Rubin-based Superclusters to multi-gigawatt scale to meet demand for training, inference and agent workloads.
The GPU side follows a familiar pattern. Each NVL72 rack links 72 GPUs over NVLink, and OCI Superclusters tie racks together with Oracle's Acceleron networking, built on NVIDIA Spectrum-X Ethernet with Spectrum-6 switches and BlueField-4 DPUs, in campuses that Oracle says can reach gigawatt-class power. Oracle repeats NVIDIA's figures that Vera Rubin offers up to 30 times the inference throughput per watt and four times the training performance of Blackwell; those are vendor numbers. The expansion builds on the GB200 and GB300 systems OCI already runs.
The CPU announcement is the more unusual part. Vera is NVIDIA's first fully custom processor, with 88 Olympus cores per chip and memory bandwidth that NVIDIA puts at as much as 1.2 TB/s, and Oracle describes itself as the first hyperscaler to offer it. OCI instances pair two Vera chips for 176 cores, alongside 1 TB of LPDDR5X, 7.68 TB of NVMe drives on the host and 200 Gbps networking, sold as bare metal or flexible virtual machines. Oracle says an initial 150,000 Vera CPUs will be deployed across its regions, although its product blog says the instances arrive in select regions during the current quarter.
The pitch for Vera is about agents rather than model math. Agents spend much of their time running tools, testing code and fetching data between model calls, and slow CPUs leave expensive GPUs idle. Customers can scale CPU and GPU capacity separately through OCI Kubernetes Engine, and Acceleron adds hardware isolation for running untrusted, model-written code. OpenAI is cited as an early Vera customer: its compute chief says the company is moving Vera into production on OCI for its fast-growing agentic sandbox workloads.
Oracle also plans HGX Rubin NVL8, an x86 server with eight Rubin GPUs on NVLink, two 64-core Intel Xeon 6 processors, 4 TB of DDR5 and 6.4 Tbps of RDMA networking, for customers who want Rubin in a smaller box compatible with existing x86 software. No date was given for that configuration or for the multi-gigawatt expansion.
The release cites OpenAI's use of an OCI Supercluster in Abilene, Texas, with more than 100,000 Blackwell GPUs to train and serve GPT-6 Astra, and an OpenAI engineering leader says Vera Rubin on OCI will help it keep pace as models grow. The news landed on the same day Oracle and Google said Google TPUs would come to OCI in 2027, part of a broader push by Oracle to offer several accelerator families. Earlier in October, Google Cloud announced its own Vera Rubin NVL72 instances, and the sandbox company Modal said it would run agent sandboxes on Vera CPUs.
Several things remain unclear. Neither company said how many Vera Rubin racks are live today, how quickly OCI will reach multi-gigawatt scale or what the new instances cost. Performance gains are NVIDIA's own figures, and the statements about OpenAI come from quotes supplied in the companies' release rather than independent reporting.