Nvidia Vera CPU: Architecture, First Customers, and Its Role in the Rubin NVL72 AI Supercomputer

Nvidia’s Vera CPU is the company’s first custom processor designed specifically for modern AI data centers. Introduced alongside the Rubin GPU architecture, Vera serves as the CPU backbone of Nvidia’s next generation of rack‑scale AI systems built for what the company calls the era of agentic AI—systems where AI models run complex tool‑using workflows rather than simple prompt‑response inference.

The chip is tightly integrated into Nvidia’s new Vera Rubin platform, which combines CPUs, GPUs, networking, and data‑processing hardware into unified “AI factory” infrastructure for large‑scale training and inference workloads.

First AI Labs and Cloud Providers to Receive Vera CPUs

The first Vera CPU systems were delivered in May 2026 to several leading AI developers and cloud providers. Nvidia reported that the initial shipments went to:

• Anthropic in San Francisco
• OpenAI in Mission Bay, San Francisco
• SpaceXAI in Palo Alto
• Oracle Cloud Infrastructure (OCI) in Santa Clara

According to Nvidia’s blog and newsroom reports, these early units were hand‑delivered by Nvidia VP of hyperscale and HPC Ian Buck as part of initial partner deployments.

These organizations are among the first to experiment with the new CPU as part of next‑generation AI infrastructure built around Rubin GPUs.

What the Vera CPU Is Designed For

Unlike general‑purpose server processors, Vera is built specifically for AI infrastructure workloads that require heavy coordination between CPUs and GPUs.

Modern AI systems increasingly run complex workflows involving:

• Agent orchestration and planning loops
• Retrieval‑augmented generation and database queries
• Reinforcement learning and simulation environments
• Tool use, sandboxes, and multi‑step reasoning

These tasks require high single‑thread performance, fast memory access, and extremely fast communication with GPUs—areas Nvidia optimized Vera to handle.

Core Technical Architecture

Public specifications describe Vera as a custom Arm‑based data‑center CPU built with Nvidia‑designed cores and optimized memory bandwidth.

Key architectural features include:

• 88 custom Armv9.2 “Olympus” cores
• 176 threads via spatial multithreading
• High‑bandwidth memory system (~1.2 TB/s)
• Coherent NVLink‑C2C connection to GPUs at up to ~1.8 TB/s

These features allow the CPU to act as the memory and orchestration engine for AI systems while Rubin GPUs perform the intensive matrix and tensor operations used for training and inference.

The design continues Nvidia’s strategy of using custom Arm‑based silicon for AI infrastructure, building on the earlier Grace CPU used in the Grace Hopper platform.

Performance Claims for Agentic AI Workloads

Nvidia positions Vera as optimized for workloads that involve coordination between many AI components rather than pure GPU math.

Company benchmarks claim that Vera can deliver:

• Up to 2× efficiency compared with traditional rack‑scale CPUs
• Around 50% faster performance in some workloads
• Up to 50% faster agent sandbox execution
• Up to 3× faster enterprise data queries

These figures come from Nvidia’s own benchmarks and have limited independent validation so far, so they should be interpreted as vendor‑supplied claims rather than definitive comparisons.

How Vera Fits into the Rubin NVL72 Platform

The Vera CPU is not intended to run alone. Instead, it forms the CPU half of Nvidia’s Vera Rubin AI platform.

The flagship system is the Vera Rubin NVL72 rack, which integrates:

• 72 Rubin GPUs
• 36 Vera CPUs
• NVLink 6 switching fabric
• ConnectX‑9 SuperNIC networking
• BlueField‑4 data processing units (DPUs)
• Quantum‑X800 InfiniBand or Spectrum‑X Ethernet networking

Together, these components form a rack‑scale AI supercomputer designed to function as a single system optimized for training, inference, and agentic workloads.

Nvidia also described Vera CPU‑only racks containing up to 256 liquid‑cooled CPUs designed for orchestration and reinforcement‑learning workloads within large AI clusters.

Why the Vera CPU Matters for Nvidia

For years Nvidia dominated AI infrastructure through GPUs alone. But large‑scale AI systems increasingly depend on a full stack of compute, networking, storage, and orchestration hardware.

With Vera, Nvidia is expanding its role from GPU supplier to complete AI infrastructure provider.

Instead of relying on host CPUs from Intel Xeon or AMD EPYC, Nvidia can now provide the CPU that coordinates its GPU clusters. This lets the company tightly integrate:

• CPU scheduling
• GPU memory access
• networking fabric
• software frameworks like CUDA

The strategy reflects a broader industry shift toward rack‑scale AI systems designed as unified platforms, rather than clusters assembled from separate components.

The Competitive Context: AMD and Intel

The launch also positions Nvidia more directly in the server CPU market.

That market is currently dominated by x86 processors from Intel and AMD, though competition has intensified. For example, AMD’s EPYC chips captured more than 33% of server CPU unit share and about 46% of server x86 CPU revenue share in early 2026, according to Mercury Research data reported by Tom’s Hardware.

By introducing its own CPU architecture for AI systems, Nvidia aims to capture more of the data‑center bill of materials and reduce dependence on external CPU vendors.

The Bigger Picture: AI Infrastructure Becoming Vertical

The Vera CPU launch highlights a larger trend in AI computing: infrastructure is becoming vertically integrated.

Rather than selling standalone chips, Nvidia now packages CPUs, GPUs, networking, DPUs, and software as a complete AI factory platform. The Rubin generation is built around this philosophy, combining multiple custom chips into unified rack‑scale systems designed to power the next wave of AI models and autonomous agents.

If the architecture delivers the performance gains Nvidia claims, Vera could play a key role in shaping how future AI data centers are built—where CPUs, GPUs, networking, and software are designed together rather than assembled from independent components.