Nvidia and Oracle expand Vera Rubin cloud infrastructure

Low GPU utilization tests the efficiency pitch

In VentureBeat Intelligence’s July, August and September surveys, 79% of 218 respondents said their GPU utilization was no higher than 50%. Of that group, 70% planned to make platform changes.

The survey covered organizations with at least 100 employees, including consultants responding on behalf of clients. Its results were self-reported, not based on a technical audit, and do not establish why GPUs were underused.

The cause matters. If GPUs sit idle while CPUs run AI-agent code, a faster CPU might help. If there is no suitable work for the GPUs, a faster CPU will not create demand. And if the constraint is a remote service, database access or application design, replacing the processor alone may not remove it.

Those are possibilities to investigate, not explanations established by the survey. For buyers, the test is whether a proposed configuration lowers the total cost of completing a task at acceptable quality, response time and reliability—not just whether its chip is faster.

Where Vera fits

Nvidia describes Vera as a processor for work around model execution: coordinating AI agents, running code in isolated environments, processing data and supporting reinforcement learning. A model can choose an action, but other software must carry it out and return the result. If that work becomes the bottleneck, improving the model itself may not speed up the full agent.

Oracle says Vera is available on dedicated bare-metal servers and flexible virtual machines. Customers can also deploy CPU services separately from GPU infrastructure.

Oracle cites Nvidia’s claim that Vera can be up to 80% faster than leading x86 processors on CPU workloads related to AI agents. But Oracle says the results came from Nvidia platform tests, not tests of OCI instances. Buyers should treat the figure as a vendor benchmark, not proof of an improvement in their own applications.

A useful comparison would run the same agent workflow on existing infrastructure and Vera, measuring successful task completion, end-to-end latency, CPU and GPU use, and total cost—including model API fees.

Two markets, different needs

OpenAI is the most prominent customer named. In a joint announcement, OpenAI’s chief infrastructure officer, Uday Ruddaraju, said the company is moving Vera into production on OCI for large-scale AI-agent sandbox workloads.

That use case speaks to foundation-model developers, but Oracle is also pitching the platform for data search, analytics, AI-agent coordination and GPU fleet management—work that can arise in enterprise AI systems.

The relevant distinction is the workload, not the size of the organization. A company handling large volumes of agent-generated code or data searches might find Vera useful without training a foundation model. A large company that mainly calls external models, by contrast, may have little reason to own a large GPU fleet.

Oracle’s offer addresses both kinds of demand, but through different products. The expanded Supercluster is aimed chiefly at resource-intensive model workloads. Standalone CPUs are a more targeted option for companies seeking to speed up the software around models.

I think the utilization figures make the CPU offer more interesting than the headline’s GPU expansion. But the survey does not show that CPU performance is the reason those GPUs sit idle, and the announcement does not provide prices or workload-specific performance data to estimate customer savings.

Available systems and planned capacity

Oracle’s NVL72 systems are available now. The company also plans a Rubin NVL8 configuration with eight GPUs and Intel Xeon processors for training, fine-tuning and inference, while retaining compatibility with x86 applications. NVL8 is planned, not yet announced as generally available.

Oracle says its infrastructure can switch between training and inference workloads and support later Nvidia platforms. It presents that flexibility as a way to use capacity as needs change. Its plans for infrastructure measured in several gigawatts describe the providers’ ambitions, not a recommended scale for an individual enterprise.

Availability also needs checking. The companies announced a launch, but Oracle’s Vera announcement says the processors will arrive in select regions this quarter. Buyers should verify which instance types, locations and capacity are actually available to them.

What I’d want to know is whether Vera improves the economics of a customer’s real workload once model API fees, delays across the full workflow and successful task completion are counted. If GPUs are waiting for CPU work, the configuration offers something concrete to test. If they lack useful tasks, more flexible infrastructure will not solve the underlying problem.