NVIDIA Vera Rubin NVL72 systems and Vera CPUs are now generally available on Oracle Cloud Infrastructure, giving cloud customers access to GPU and CPU hardware aimed at different parts of AI workloads. The companies also plan to scale Rubin superclusters to multi-gigawatt capacity, a substantial expansion rather than a new chip launch.
NVIDIA Watch analysis
What happened
NVIDIA and Oracle announced general availability on 27 October. Their OCI offering pairs Rubin systems for training and inference with Vera CPUs for agent tasks such as executing tools, running code and retrieving data. Customers can scale CPU and GPU resources independently, according to the companies.
They also plan to add HGX Rubin NVL8, an eight-GPU system in an x86-based configuration, for training, fine-tuning and inference. The announcement gives no availability date for that planned system.
Why it matters
AI agents do more than call a model: they may need to run code, retrieve information and use tools between model calls. Putting CPU resources alongside GPUs, and allowing customers to scale them separately, could help operators match infrastructure to those different jobs instead of leaving one type of hardware waiting on the other.
This is a meaningful cloud availability change, but customers still need real workload results and pricing to judge the practical trade-offs. The multi-gigawatt expansion is a plan, not capacity already available to book.
Our read
The useful news is that Vera Rubin and Vera CPUs have moved from future infrastructure plans to general availability on OCI. The more interesting question is whether separately scaling the agent’s CPU work makes a measurable difference to cost and responsiveness. Hardware announcements are easy; efficient workloads are the bit that has to earn its keep.
What to watch
- OCI pricing, regions and customer access for the generally available systems.
- Measured performance for agent workloads using Vera CPUs alongside GPUs.
- Timing and specifications for the planned HGX Rubin NVL8 and multi-gigawatt expansion.
Discussion spark: For agent workloads, should cloud customers pay for separate CPU and GPU scaling if it improves utilisation, or favour simpler all-in-one capacity?
Sources and evidence
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.