Nvidia's Vera CPU has begun shipping in multiple form factors, with the company confirming that dedicated Vera racks are available now ahead of the OCP Global Summit in San Jose.

The platform is Nvidia's first serious bid to challenge AMD and Intel's grip on the CPU market, with more than 20,000 cores per rack aimed at agentic workloads. Meta, CoreWeave, and Amazon Web Services are among the named early customers, with AWS billing the chip as a dedicated compute option for agentic jobs.

What does the Vera rack platform include?

Vera is built on Nvidia's MGX modular reference architecture. It ships as a discrete side-cabinet placed adjacent to an NVL72 rack. The system sits outside the core Vera Rubin platform, where CPU resources are reserved for orchestration.

Nvidia's argument for the separate cabinet rests on workload division. The CPUs already embedded in an NVL72 handle orchestration and management tasks, leaving too few cores free to run agentic jobs at scale. Splitting the agentic compute into a dedicated rack keeps both functions fed.

"If you get a Vera CPU rack, it's got over 20,000 cores available for agents to operate quickly and complete these tasks as fast as possible," Joe DeLaere, Nvidia's product lead for AI infrastructure solutions, told reporters.

What performance gains is Nvidia claiming?

Nvidia published two benchmark comparisons against its previous Grace generation:

• More than 2× performance on agentic workloads, measured using SPEC CPU 2026
• 1.8× performance on Terminal Bench, a test that runs real agent code in real containers

DeLaere tied the gains to Vera's high-bandwidth memory subsystem and high single-threaded performance. The combination accelerates the tool-calling and code-execution paths that often bottleneck agent systems.

"Because Vera CPU was designed specifically for agents with its high bandwidth and its high single-threaded performance, we see that it really accelerates the agentic tool calling and execution of code … and having those separate is really how we can achieve these tremendous agentic workload benchmarks," DeLaere said.

How does Vera fit into Nvidia's wider product roadmap?

Vera entered full production in spring 2025 and is set to replace Grace as Nvidia's flagship processor. The chip also marks a design shift for Nvidia. Earlier CPU work leaned on standard Arm cores; Vera uses a custom core developed largely in-house, while still implementing the Arm v9.2-A instruction set.

Three deployment options will be available at launch:

• Standalone Vera processor
• Dedicated Vera rack platform for agentic workloads
• CPU portion of the upcoming Vera Rubin Superchip

The Vera Rubin Superchip pairs Vera with Nvidia's next-generation Rubin GPU. It is scheduled for release in the second half of 2025.

Why is Nvidia pushing into standalone CPUs now?

Nvidia first announced plans for a standalone CPU business at the start of 2025, framing the move as a response to the AMD-Intel duopoly that has shaped x86 server purchasing for decades. Meta and CoreWeave were named as launch customers at the time.

The push coincides with a broader shake-up of the processor market. Arm, long a licensor of chip designs, has begun selling physical chips directly, betting that agentic and AI-heavy workloads demand purpose-built silicon. That move puts Arm in a more direct competitive stance with both Nvidia and its traditional licensees.

AWS has confirmed plans to offer Vera as an additional compute option for agentic workloads, alongside its existing Graviton Arm-based instances and Intel Xeon-based instance types.