Nvidia and Oracle are expanding their AI infrastructure partnership with the general availability of Vera Rubin NVL72 systems and standalone Vera CPUs on Oracle Cloud Infrastructure, alongside plans to expand their Superclusters to multi-gigawatt scale.

The October 27 announcement addresses two different demands: the computing power required to train and run advanced models, and the CPU capacity needed for agents to execute code, retrieve information and use software tools between model calls.

The companies argue that allowing customers to scale CPU and GPU resources separately can help reduce unused capacity.

For enterprise buyers, that makes the announcement about more than access to Nvidia’s latest hardware: it raises the question of whether a different balance of infrastructure can produce more useful work from the capacity they pay for.

That question is particularly relevant given the latest findings from VentureBeat Intelligence, VentureBeat’s in-house research division, showing low enterprise GPU usage.

Enterprise utilization presents a test for the efficiency pitch

VentureBeat Intelligence's monthly Pulse surveys from July, August and September found 79% of 218 respondents reporting GPU utilization at 50% or below; 70% of that group nevertheless plans platform changes.

The research surveys organizations with at least 100 employees, including consultants answering for clients. These are self-reported findings, not a hardware audit.

For buyers assessing Nvidia and Oracle’s announcement, the implication is that access to more powerful chips should come with a diagnosis of what prevents existing resources from doing useful work.

A GPU waiting for a CPU to finish executing an agent’s code presents a different purchasing problem from a GPU with no queued work. Faster CPUs could help the former.

They would not create demand for the latter. Likewise, if the constraint lies in a remote service, database access or an application’s design, a processor upgrade alone may leave the underlying problem unresolved.

Those are possible explanations that an enterprise should test, rather than causes established by the utilization finding.

The distinction also separates the business cases for a foundation-model developer and an enterprise application team. A model developer with sustained training and inference demand may need additional capacity and faster supporting infrastructure. A business with lightly used GPUs should first establish which workloads would move to the proposed system and why they would run more economically there.

The appropriate test is therefore not simply whether a new chip is faster. It is whether the proposed configuration lowers the total cost of completing an acceptable task, at the response time and reliability the business requires.

Why CPUs matter to an AI agent

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

A model can decide which action to take, but other software must carry out that action and return the result. When that work becomes a bottleneck, improving model performance alone does not necessarily make the whole agent finish sooner.

Oracle’s technical announcement says Vera is available through bare-metal servers and flexible virtual machines. The standalone offering allows customers to deploy CPU services separately from GPU infrastructure.

Oracle cites Nvidia’s claim of up to 80% faster performance on agent-related CPU workloads compared with leading x86 processors. It also explicitly says those results come from Nvidia platform benchmarks, not tests of the OCI instances themselves. Buyers should treat that figure as a vendor benchmark rather than a demonstrated improvement for their applications.

For an enterprise considering a migration, a representative test would run the same agent workflow on its existing infrastructure and the proposed Vera configuration. It should measure successful task completion, delays across the full workflow, CPU and GPU consumption, and total cost—including any model API charges.

The customer is both AI labs and enterprises—with different needs

The strongest named customer example is OpenAI.

“We are bringing Vera into production on OCI to support these workloads at scale,” Uday Ruddarraju, OpenAI’s chief technology officer of compute, says in the joint announcement, referring to its agent sandbox workloads.

That example supports a clear role for foundation-model companies. But the potential customer base extends beyond organizations training their own frontier models.

Oracle identifies retrieval, analytics, agent coordination and GPU fleet management among the CPU platform’s uses. Those workloads can also exist inside enterprise AI deployments.

Our reading is that workload characteristics provide a better guide than company size alone. An enterprise running substantial volumes of agent code or retrieval operations could have a reason to evaluate Vera without training a foundation model. Conversely, a large company whose AI use consists mainly of calling externally hosted models may have little reason to acquire a large GPU allocation of its own.

The announcement therefore addresses both markets, but its components serve different needs. The Supercluster expansion most directly addresses exceptionally demanding model workloads. Standalone CPUs offer a more targeted option for organizations seeking to accelerate the software around those models.

A smaller Rubin configuration is still on the roadmap

Alongside the available NVL72 systems, Oracle plans an eight-GPU Rubin NVL8 configuration using Intel Xeon processors. It is intended to support training, fine-tuning and inference while maintaining compatibility with x86 applications.

That offers a smaller deployment option than the 72-GPU rack configuration, although the announcement describes NVL8 as planned rather than generally available.

Oracle also says its infrastructure can shift between training and inference workloads and accommodate successive Nvidia platforms. That is part of its argument that customers can keep using capacity as their requirements change.

For procurement teams, the distinction between available products and planned expansion matters. The multi-gigawatt buildout describes the providers’ infrastructure ambitions; it is not a minimum deployment requirement for an individual enterprise.

What buyers still need to establish

The announcements reviewed do not provide enough workload-specific pricing and performance evidence to calculate a customer’s expected savings.

Regional access also requires checking. Although the companies announce general availability, Oracle’s Vera CPU post says availability will extend to selected regions this quarter. Buyers should confirm the instance types, locations and capacity they can actually obtain.

For a team whose GPUs repeatedly wait on CPU work, Vera presents a concrete proposition to test. For a team with insufficient demand to occupy its existing GPUs, the first task is different: identify useful workloads, reconsider capacity commitments and establish whether a new platform improves the economics.

Nvidia and Oracle are offering additional infrastructure and a way to balance parts of an AI system more independently. The enterprise purchasing decision rests on demonstrating where that flexibility translates into completed work.