Oracle customers can now access NVIDIA’s Vera Rubin NVL72 systems and Vera CPUs through Oracle Cloud Infrastructure, the companies said on October 27, 2026. Their announcement pairs a planned multi-gigawatt expansion with a different efficiency pitch: giving AI agents enough CPU capacity to execute tools and code without leaving GPUs waiting between model calls.
Two kinds of computing, two kinds of work
The collaboration, announced at Oracle AI World in Las Vegas, covers both sides of an agent’s workload. GPUs handle the intensive computing involved in training models and generating responses. CPUs perform surrounding tasks, including retrieving information, running tools and testing code. Slow execution of those steps can hold up an agent even between model responses, the companies said.
Oracle’s pitch for Vera CPUs rests on two features: strong performance on individual processing threads and high memory bandwidth. The first is intended to accelerate individual tool calls and code execution. The second keeps processor cores supplied with data when agents run many tasks at once. The companies say that combination can reduce GPU idle time; they did not provide a quantified improvement in this announcement.
Customers can scale CPU and GPU resources independently through OCI Kubernetes Engine, Oracle’s service for managing software workloads. That lets them adjust each type of computing to the work at hand rather than increase both together. Oracle also says its Acceleron infrastructure provides hardware-based isolation for workloads running untrusted, AI-generated code.
We are bringing Vera into production on OCI to support these workloads at scale.
Uday Ruddarraju, OpenAI chief technology officer of compute, on agentic sandbox workloads
A 72-GPU rack—and a smaller option
For the largest model workloads, the companies plan to take OCI Superclusters using Vera Rubin NVL72 systems to multi-gigawatt scale. Each NVL72 rack connects 72 GPUs through NVIDIA’s NVLink networking. Oracle says its network can connect hundreds of thousands of GPUs, combining the rack systems with gigawatt-class power and cooling.
The network combines Oracle Acceleron with NVIDIA Spectrum-X Ethernet technology, Spectrum-6 switches and BlueField-4 data processing units. This planned expansion builds on Oracle’s existing GB200 and GB300 NVL72 infrastructure. NVIDIA says the shared CUDA software foundation helps customers support changing models and workloads across those hardware generations.
Not every planned configuration is rack-sized. Oracle also intends to offer HGX Rubin NVL8, an x86-based server with eight Rubin GPUs connected through NVLink. The company positions it as a smaller option for training, fine-tuning and generating model responses, while retaining compatibility with x86-based applications.
The planned eight-GPU server
- Two Intel Xeon 6 CPUs, each with 64 cores.
- 4 TB of DDR5 memory.
Available hardware is not the same as completed capacity
The immediate change is general availability of Vera Rubin NVL72 and Vera CPUs on OCI. The multi-gigawatt Supercluster expansion and eight-GPU offering remain future plans. Oracle’s product disclaimer says development, release timing and pricing can change, and that the outlined direction is not a delivery commitment.
OpenAI supplies a concrete example of the existing infrastructure’s use. The announcement says it trained and served GPT-6 Astra on an OCI Supercluster in Abilene with more than 100,000 NVIDIA Blackwell GPUs. That deployment used Blackwell hardware, not the newly available Rubin systems. OpenAI’s Vera CPU production work targets a different part of the stack: the environments where agents execute code and run tools.
Reader comments
Newest comments first. Replies stay oldest first.