en.Wedoany.com Reported - NVIDIA officially announced on May 19 local time that the first systems featuring its Vera CPU, the company's first processor purpose-built for agentic AI, have been delivered. Ian Buck, NVIDIA's Vice President of Hyperscale and HPC, personally delivered the initial systems to four customers: last Friday, systems arrived at three of the world's top AI labs—Anthropic in San Francisco, OpenAI in Mission Bay, and SpaceXAI (formerly xAI) in Palo Alto; on Monday, another batch of systems was delivered to Oracle Cloud Infrastructure (OCI) in Santa Clara. This marks the official transition of the Vera CPU from the announcement phase to volume production and delivery.
The Vera CPU is NVIDIA's first fully custom-designed CPU, built specifically for agentic AI workloads. It features 88 custom NVIDIA "Olympus" cores, supporting 176 threads, and delivers a 50% improvement in single-core performance over the previous-generation Grace CPU under full load. In terms of memory, Vera is the world's first data center CPU to use LPDDR5X memory, achieving 1.2 TB/s of memory bandwidth via SOCAMM modules and supporting up to 1.5 TB of system memory—triple the capacity of Grace. For interconnect capabilities, Vera supports 1.8 TB/s of second-generation NVLink-C2C coherent memory interconnect, enabling it to form NVIDIA's next-generation Vera Rubin AI factory architecture alongside Rubin GPUs, BlueField-4 DPUs, ConnectX-9 SuperNICs, and Spectrum-X Ethernet switches.
At the GTC conference in San Jose this March, NVIDIA CEO Jensen Huang positioned the standalone Vera CPU as the company's next billion-dollar business. NVIDIA not only uses Vera as the host processor for Rubin GPUs in the Vera Rubin NVL72 rack-scale system—each NVL72 integrates 36 Vera CPUs and 72 Rubin GPUs—but will also sell it as a standalone product, directly targeting the data center CPU market.
During the delivery, Buck stated: "Agentic AI is creating a new CPU moment in AI factories—as models shift from answering questions to proactively executing tasks, Vera is specifically designed to support these massive-scale workloads." He further explained that when an AI model faces a problem, the answer is often not pre-computed; the model needs to actually generate Python code, invoke tools, and orchestrate tasks to arrive at the correct result. All of this is core CPU-level work, and it is the observation of this trend that has driven the surge in CPU demand.
Among the first customers to receive the systems, each has a distinct focus for Vera's application. James Bradbury, Head of Compute at Anthropic, said after receiving the system: "Scaling compute is a key accelerator for model growth, and we look forward to seeing Vera become an important part of the AI ecosystem in the agentic workload space." At OpenAI's Mission Bay headquarters, Sachin Katti, Head of Compute Infrastructure, personally received the system, where Buck also opened the chassis cover on-site to showcase the internal architecture. SpaceXAI's delivery was personally signed for by founder Elon Musk, who inquired in detail about the core count, memory layout, and cooling solution. SpaceXAI is evaluating Vera's performance in reinforcement learning workloads and agent-based simulation pipelines.
Oracle Cloud Infrastructure's commitment was the most explicit. Karan Batta, Head of OCI Product Management, stated: "Oracle Cloud plans to deploy hundreds of thousands of NVIDIA Vera CPUs starting in 2026. Agentic AI requires sustained performance at scale, and Vera's architecture is designed for high-throughput inference workloads, delivering the efficiency, density, and footprint that Oracle Cloud needs to power the next generation of enterprise AI." Oracle thus becomes the first cloud service provider to commit to hyperscale deployment of Vera.
The delivery of the Vera CPU comes at a critical juncture as the AI industry transitions from generative AI to agentic AI. Traditional generative AI primarily focuses on answering questions and generating content, whereas agentic AI needs to autonomously plan processes, invoke external tools, execute code, retrieve information, and complete multi-step tasks. This shift places entirely new demands on data center CPUs—requiring them to simultaneously handle diverse workloads such as intelligent sandboxing, tool calling, task orchestration, and long-context retrieval under high-concurrency, real-time task pressure. Vera is a new class of processor designed from the ground up with this reality in mind.
The Vera CPU is built on TSMC's 3nm process technology, utilizing 2.5D/3D advanced packaging, and has a shorter production cycle than the Rubin GPU, which is a key reason Vera could be delivered to customers first. The production ramp for the Vera Rubin platform is accelerating: Vera CPUs have completed initial deliveries, and Rubin GPUs are expected to enter high-volume production and shipment from the second half of this year into the third quarter. As NVIDIA takes commercial steps into the CPU market, its full-stack AI infrastructure blueprint—spanning GPUs, CPUs, DPUs, and network switching silicon—is rapidly taking shape.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
