Oracle and Nvidia have announced plans to expand Oracle Cloud Infrastructure (OCI) Superclusters with Nvidia Vera Rubin NVL72 systems to multi-gigawatt scale. The announcement was made on Tuesday at Oracle AI World in Las Vegas, and it pairs a long-range expansion plan with systems customers can use now: Vera Rubin NVL72 racks and Nvidia Vera CPUs are now generally available on OCI.
The two companies say the expansion is driven by growing demand for AI training, inference and agentic applications — the autonomous software that carries out multi-step tasks on a user's behalf. Rather than treating graphics processors as the whole story, the expanded superclusters are built around both GPU-heavy model work and the CPU-heavy work that happens between model calls, such as executing tools, testing code and retrieving data.
What is going live now
The immediate milestone is general availability. Customers can now run workloads on Vera Rubin NVL72 systems and Vera CPUs through OCI, instead of waiting for a future hardware generation. The companies also plan to introduce Nvidia HGX Rubin NVL8 systems, a smaller configuration with eight Rubin GPUs in an x86-based server, though that remains a planned offering rather than a generally available one.
Each NVL72 rack connects 72 graphics processors through Nvidia's NVLink networking, allowing the GPUs inside a rack to work as a single large system. OCI's superclusters combine those racks with Oracle Acceleron networking, Nvidia Spectrum-X Ethernet networking with Spectrum-6 switches, and BlueField-4 data processing units, backed by gigawatt-class power and cooling. Oracle says its network design scales across hundreds of thousands of GPUs.
The expansion builds on OCI's existing Nvidia-powered infrastructure, including GB200 and GB300 NVL72 systems from the previous Blackwell generation. Both generations share Nvidia's CUDA software foundation.
Why the architecture matters for agents
Vera Rubin is Nvidia's successor to the Blackwell generation, and the company positions it as a platform for agentic AI as much as for model training. The Vera CPU is designed for the work agents do between model calls, while the Rubin GPUs handle the model itself. Customers can scale CPU and GPU resources independently, the companies say, which helps reduce idle capacity when a workload needs far more of one resource than the other.
Nvidia says Vera Rubin is designed to deliver substantially higher inference throughput per watt than Blackwell, and each Rubin GPU uses high-bandwidth HBM4 memory. Those are vendor claims, and real-world results will depend on the workload — but inference, the computing that produces answers to user requests, is now a large and growing share of AI demand alongside training.
What is still just a plan
The multi-gigawatt headline needs a note of caution. The companies have not disclosed a purchase price, a firm completion date for the full expansion, or an exact number of gigawatts. General availability of the NVL72 systems and Vera CPUs is real and immediate, but the multi-gigawatt build-out itself is a plan that will be delivered over time, not a data centre that switched on overnight.
Even so, the announcement lands at a telling moment: cloud providers are competing to host frontier models and agent workloads at scale, and customers are weighing networking, power, cooling and software maturity alongside raw chip performance. By putting Vera Rubin into production on OCI, Oracle and Nvidia are betting the next phase of AI growth will be won on infrastructure breadth — chips, CPUs, networking and power together — rather than on any single processor alone.
