Nvidia and Oracle announced plans on 27 October to expand Oracle Cloud Infrastructure (OCI) Superclusters with Vera Rubin NVL72 systems to multi-gigawatt scale, alongside the general availability of the systems and Vera CPUs on OCI, according to Nvidia’s press release. The announcement came at Oracle AI World in Las Vegas, StreetInsider reported.
Key points
- Oracle and Nvidia plan to take OCI Superclusters using Vera Rubin NVL72 systems to multi-gigawatt scale.
- Oracle plans an initial deployment of 150,000 Vera CPUs across its regions.
- Nvidia claims up to 80% faster agentic CPU performance than leading x86 CPUs, based on platform benchmarks rather than tests of OCI instances.
- OpenAI says it is bringing Vera CPUs into production on OCI for agent workloads.
Oracle plans more NVL72 capacity
The proposed expansion builds on OCI’s existing GB200 and GB300 NVL72 systems. Nvidia says an NVL72 rack connects 72 GPUs through NVLink, while OCI’s network can extend across hundreds of thousands of GPUs. The companies are pitching the larger Superclusters for model training, inference and applications in which AI agents carry out tasks between model calls.
Oracle also plans to offer HGX Rubin NVL8, a smaller configuration with eight Rubin GPUs connected through NVLink. Nvidia says the planned x86-based server will have two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory and 6.4 Tbps of high-speed remote direct memory access networking, which moves data between machines without routing it through their CPUs. The company positions NVL8 for training, fine-tuning and inference.
Nvidia says customers can scale CPU and GPU resources independently rather than buying both in a fixed balance. Agents can spend time executing tools, testing code and retrieving data between model calls, leaving GPUs waiting for CPU work to finish. OCI Kubernetes Engine supports separate scaling of those resources, according to Nvidia.
OpenAI uses more than 100,000 Blackwell GPUs
OpenAI has trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 Nvidia Blackwell GPUs in Abilene, Texas, Nvidia said. That deployment uses Blackwell hardware, the generation already running on OCI, rather than the newly available Vera Rubin systems.
OpenAI is also among the customers with early access to Vera CPUs. Uday Ruddarraju, OpenAI’s chief technology officer of compute, said the company is bringing Vera into production on OCI to support agent workloads that execute code and run tools. His statement concerns the CPU capacity used between model calls, a different part of the workload from the GPUs used to train and serve GPT-6 Astra.
In a separate measure of enterprise demand, monthly surveys by VentureBeat Intelligence in July, August and September found that 79% of 218 respondents reported GPU utilisation of 50% or below. Of that group, 70% planned platform changes. The findings were self-reported rather than measurements of hardware use.
Oracle plans 150,000 Vera CPUs
Oracle announced an initial deployment of 150,000 Vera CPUs across its cloud regions, Unite.AI reported. Each processor has 88 Arm-compatible Nvidia Olympus cores, 176 threads and up to 1.2 TB/s of memory bandwidth. Oracle described itself as the first hyperscaler to bring Vera CPUs to market.
Nvidia claims up to 80% faster performance on agent-related CPU workloads than leading x86 processors. Oracle said that figure comes from Nvidia platform benchmarks, rather than tests of OCI instances, and that performance varies with workload and configuration, according to Unite.AI.
The CPUs are offered through Bare Metal servers and Flex Virtual Machine instances, allowing customers to run CPU services separately from GPU infrastructure. Nvidia says OCI’s Acceleron networking also provides hardware-based isolation for workloads executing AI-generated code.
Oracle plans to make the Vera CPU instances available in select regions in the fourth quarter of 2026, Unite.AI reported.
