NVIDIA and Oracle scale Vera Rubin to multi-gigawatt on OCI
NVIDIA and Oracle have announced plans to extend Oracle Cloud Infrastructure Superclusters with NVIDIA Vera Rubin NVL72 systems to multi-gigawatt capacity, as demand for AI training, inference and agentic workloads continues to accelerate. Alongside the capacity expansion, the two companies confirmed general availability of NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPUs on OCI, deepening a partnership that already covers earlier Blackwell-generation infrastructure.
The announcement was made at Oracle AI World in Las Vegas on 27 October 2026. OpenAI is the headline named customer, having trained and served GPT-6 Astra on an OCI Supercluster running more than 100,000 NVIDIA Blackwell GPUs at a facility in Abilene, Texas.
What is being deployed
OCI Superclusters combine NVIDIA NVL72 rack-scale systems with Oracle Acceleron networking, NVIDIA Spectrum-X Ethernet switching featuring Spectrum-6 switches, and BlueField-4 DPUs. Each rack connects 72 GPUs through NVLink scale-up networking, and OCI's multiplanar fabric is said to extend coherently across hundreds of thousands of GPUs. Oracle also plans to introduce an NVIDIA HGX Rubin NVL8 option: eight Rubin GPUs in a standard x86 server configuration, paired with two 64-core Intel Xeon 6 CPUs, 4 TB of DDR5 memory and 6.4 Tbps of RDMA networking. The HGX NVL8 form factor is designed for customers that want next-generation GPU capability without committing to full rack-scale deployments.
The Vera CPU, distinct from the Vera Rubin GPU, targets the CPU-intensive tasks that AI agents run between model calls: tool execution, code testing and data retrieval. OCI Kubernetes Engine allows GPU and CPU capacity to be scaled independently, so agents are not left waiting on GPU cycles for work that a high-throughput CPU can handle more efficiently. Oracle Acceleron adds hardware-based isolation for workloads executing untrusted, AI-generated code, which is increasingly relevant as autonomous coding agents move into production.
Sreekanth Pothanis, engineering leader for compute infrastructure at OpenAI, said the company needs "infrastructure that can keep pace" as its frontier models advance, citing Vera Rubin on OCI as meeting that bar. Uday Ruddarraju, OpenAI's chief technology officer of compute, added that Vera CPUs deliver the "loaded performance needed to execute code and run tools across our rapidly growing agentic sandbox workloads."
Market context
The multi-gigawatt ambition puts Oracle squarely in competition with the hyperscaler GPU clouds operated by Amazon Web Services, Microsoft Azure and Google Cloud, each of which is pursuing similar rack-scale, high-bandwidth-memory configurations based on their own NVIDIA allocations and, increasingly, custom silicon. Oracle's strategy is differentiated by its exclusive focus on compute density and its long-standing NVIDIA partnership, rather than an integrated cloud application stack. The company has also been more willing than some peers to publish raw cluster sizes and customer names, which provides useful reference points for enterprise buyers evaluating total cost at scale.
Power at multi-gigawatt scale raises significant infrastructure questions. A single gigawatt of sustained load represents roughly the output of a large gas-fired power station. Data centre developers and hyperscalers are under growing scrutiny from grid operators and regulators in both the US and Europe over the pace at which AI workloads are absorbing new generation capacity. Neither Oracle nor NVIDIA detailed the grid arrangements or power-purchase agreements underpinning the planned expansion in this release.
Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, described the goal as making AI factories "productive, durable and fungible," pointing to CUDA software continuity as a mechanism for protecting customer investment across GPU generations. For enterprise buyers, that backward compatibility argument is meaningful: it reduces the risk of stranded infrastructure investment as model architectures evolve. The next milestones to watch are firm availability dates for HGX Rubin NVL8 and public capacity figures for the multi-gigawatt Supercluster buildout.