Standalone Nvidia Vera CPUs join Oracle’s multi-gigawatt deployment

By Rebecca Stone

Oracle and Nvidia announced general availability of standalone Vera CPUs on Oracle Cloud Infrastructure (OCI) Superclusters on Tuesday at Oracle AI World, marking Nvidia’s first standalone CPU push into hyperscale computing as Oracle brings up a multi-gigawatt deployment.

The launch covered both standalone Vera instances and Nvidia’s Vera Rubin NVL72 racks. It followed Google Cloud’s general-availability announcement for Vera Rubin the previous week, giving Nvidia two major hyperscalers. Oracle’s standalone option matters because customers no longer need a Rubin GPU to deploy Vera.

Vera instances were already under way at Nscale, CoreWeave and Nebius, among other neoclouds, where the chip fits GPU-heavy AI clusters. Oracle sends a different signal. Hyperscalers operate large fleets and sometimes rent capacity from neoclouds, but their CPU deployments remain primarily x86; OCI otherwise offers Ampere Altra for an Arm-based instance.

What does Vera change in hyperscale computing?

Vera will not, by itself, topple the x86 guard. Ian Buck, the Nvidia CUDA architect behind the company’s AI infrastructure buildout, said in March when Nvidia revealed the chip: “the world couldn’t be served by a single, 88-core SKU.”

The warning keeps Oracle’s launch in perspective, while Vera gives hyperscalers an opening to add another CPU supplier in future generations.

AMD and Intel still lead single-socket density, and AMD’s upcoming EPYC Venice processors are designed to pack up to 512 threads into one socket, a roadmap target that could widen that lead. Vera counters with a topology that can adapt to the workload.

How large is Vera’s memory-bandwidth lead?

Nvidia said Vera’s memory system was “capable of 1.2 TB/s of bandwidth.” That is a vendor capability claim, not a measured OCI result. Independent testing on a local Vera server nevertheless confirmed the comparative lead, although those tests did not run on OCI.

In the industry-standard Stream benchmark, Vera’s LPDDR5X-9600 memory, monolithic die and distributed mesh cache delivered more than twice the bandwidth available on a dual Intel Xeon 6980P system, even when the Xeon used MRDIMMs running at 8,800 MT/s. Triad, Stream’s most demanding operation, showed similar Vera leads in add, copy and scale.

OpenMP’s “spread” policy amplified the advantage by binding processes as far apart as possible. AMD and Intel centralize memory control and system-level cache, whereas Vera distributes them across a mesh across the die. Set correctly, that layout produced much higher memory bandwidth in the test.

The bandwidth can matter when an agentic loop queries a large database or works through extensive documentation. Python remains central because CrewAI and LangGraph use it, while Rust and TypeScript frameworks are emerging. Vera also beat CPUs with stronger on-paper specifications in independent Python testing; PyTorch, TensorFlow and other popular AI applications remain in that ecosystem.

Does Vera matter outside agentic AI?

Oracle’s standalone deployment widens Vera’s addressable workload beyond Nvidia’s agentic-AI pitch. Many agent jobs remain ordinary data-center jobs under high concurrency, and SQLite illustrates the point: database-query speed matters whether logic, a person or an AI agent wrote the query.

Vera’s second-generation Scalable Coherency Fabric, already visible in Stream, also appeared in the SQLite workload, where queries remained largely bound by data movement across the system. The result could make Vera preferable to x86 for selected general-purpose tasks, even without access to an Nvidia accelerator, but only production use will settle that claim.

Oracle is approaching its multi-gigawatt rollout cautiously, and the announcement did not specify Vera instance counts, shipments or revenue. It will also offer Nvidia’s HGX NVL8, which clusters eight Rubin GPUs and connects them to a host with two 64-core Intel Xeon 6 processors and 4 TB of DDR5 memory.

Tom’s Hardware senior CPU analyst Jake Roach conducted the separate benchmarking session. Full technical details and results will follow shortly in a Tom’s Hardware review. The evidence supports Vera’s usefulness in agentic loops and its early promise in some general-purpose work.

Oracle’s general availability suggests demand extending beyond Rubin, but customer workloads, adoption levels and pricing will determine whether standalone Vera becomes more than an opening. The next test is whether OCI can attract customers to run Vera without Rubin, what pricing it can sustain, and whether those gains translate into x86 share.