bytevyte
bytevyte
Language
EN DE ES FR PT IT
ai-beats —

NVIDIA Ships First Vera CPUs to OpenAI and Anthropic for Agentic AI Workloads

agentic AI

NVIDIA has started shipping its first custom CPU, Vera, to major artificial intelligence laboratories including OpenAI and Anthropic. This hardware release is a shift in data center architecture, moving away from general-purpose processors toward silicon optimized for agentic AI workloads. The Vera CPU is designed to manage the complex orchestration, tool-calling, and reinforcement learning tasks that often create bottlenecks in traditional computing environments.

The new processor features 88 custom-designed Olympus cores per socket and provides 1.2 TB/s of memory bandwidth. According to NVIDIA, the architecture delivers twice the energy efficiency of existing infrastructure. While it can function as a standalone unit, the processor is primarily intended to serve as a host for Rubin GPUs via the NVLink-C2C interconnect. Initial recipients of the hardware also include SpaceXAI and Oracle Cloud Infrastructure.

Optimizing Infrastructure for Agentic AI

The introduction of the Vera CPU addresses the specific computational demands of autonomous agents. These systems require rapid data analytics and frequent interaction with external software tools, tasks that differ significantly from the massive parallel processing handled by GPUs. NVIDIA reports that enterprise data queries run up to three times faster on Vera compared to standard server CPUs, while agent sandboxes see a 50% performance increase.

Strategic deployment of this hardware is already scaling. Oracle Cloud Infrastructure has announced plans to integrate hundreds of thousands of Vera units into its data centers starting in 2026. This large-scale adoption suggests a transition toward specialized clusters where the CPU is no longer a secondary component but a specialized engine for logic and tool management.

Cost efficiency remains a primary driver for this hardware shift. When paired with the Vera Rubin NVL72 system, NVIDIA states that agentic AI inference costs can be reduced to one-tenth the price per token. By offloading orchestration tasks to a dedicated processor, AI developers can maximize the utilization of their GPU clusters, potentially shortening the training and deployment cycles for next-generation autonomous models.

While we strive for accuracy, bytevyte can make mistakes. Users are advised to verify all information independently. We accept no liability for errors or omissions.

Sources

Vera Arrives: NVIDIA’s First CPU Built for Agents Lands at Top AI Labs

NVIDIA CEO Jensen Huang at Dell Technologies World: ‘Demand Is Going Parabolic’

AI-generated image.

✔Human Verified

Share