Oracle Separates CPU and GPU Scaling for NVIDIA’s Vera Rubin Push

Oracle is making NVIDIA’s Vera Rubin infrastructure available. NVIDIA and Oracle announced the general availability of Vera Rubin NVL72 systems and standalone Vera CPUs on Oracle Cloud Infrastructure, with plans to expand their Superclusters to multi-gigawatt scale.

The October 27 announcement targets two different computing demands. Advanced models need substantial GPU power for training and execution, while agents need CPU capacity to run code, retrieve information and use software tools between model calls.

That distinction matters because customers may not need both resources at the same level. NVIDIA and Oracle argue that letting customers scale CPU and GPU capacity separately can reduce unused infrastructure — a practical response to a problem that does not improve just because the rack gets larger.

The utilization problem is not theoretical

VentureBeat Intelligence’s monthly Pulse surveys from July, August and September found that 79% of 218 respondents reported GPU utilization at 50% or below. Of that group, 70% still plans platform changes, suggesting that low use has not reduced the pressure to redesign AI infrastructure.

The surveys cover organizations with at least 100 employees and include consultants answering for clients. The findings are self-reported, not a hardware audit, so they show how organizations describe their infrastructure rather than measuring every GPU directly.

That limitation does not erase the underlying tension. Businesses can spend on large GPU systems for training and inference, then leave much of that capacity unused when workloads shift toward agents, code execution and data processing.

Vera adds a CPU layer to AI infrastructure

NVIDIA describes Vera as a processor for the work surrounding model execution. Its roles include coordinating agents, running code in isolated environments, processing data and supporting reinforcement learning.

Oracle says Vera is available through bare-metal servers and flexible virtual machines. The standalone offering lets customers deploy CPU services separately from GPU infrastructure, giving them a way to match each resource to the workload that needs it.

Oracle cites NVIDIA’s claim of up to 80% faster performance on agent-related CPU workloads compared with leading x86 processors. The company also states that these results come from NVIDIA platform benchmarks, not tests of the OCI instances themselves — a distinction that belongs in the headline-sized facts, not the footnotes.

The strategy also supports large agent deployments. NVIDIA’s Vera positioning includes workloads involving more than 1,000 domain-specific agents, where CPU demand can rise as agents execute code, retrieve information and use software tools between model calls.

OpenAI is bringing Vera into production on OCI to support these workloads at scale, according to Uday Ruddarraju, OpenAI’s chief technology officer of compute. “We are bringing Vera into production on OCI to support these workloads at scale,” Ruddarraju said.

Rubin systems keep the GPU path open

Alongside the available 72-GPU rack configuration, Oracle plans an eight-GPU Rubin NVL8 configuration using Intel Xeon processors. The system is intended to support training, fine-tuning and inference while maintaining compatibility with x86 applications.

That x86 compatibility gives Oracle a way to add Rubin systems without forcing every existing application into a new processor architecture. Oracle also says its infrastructure can shift between training and inference workloads and accommodate successive NVIDIA platforms.

Oracle’s Vera CPU post says availability will extend to selected regions this quarter. Together, the NVL72 systems, standalone Vera CPUs and planned NVL8 configuration give customers several ways to build AI infrastructure without treating every workload as a GPU problem.

The larger plan reaches beyond individual instances. NVIDIA and Oracle intend to expand their Superclusters to multi-gigawatt scale, pairing broad GPU capacity with CPU resources for the operational work around models and agents.

The pitch is not simply more computing power. It is better alignment between the kind of work a system performs and the hardware assigned to it. Given that 79% of surveyed respondents reported GPU utilization at 50% or below, that may be the more important upgrade.

Stay connected via Google News

Follow us for the latest AI News