Nvidia launches Vera, the first processor designed for AI agents

The adoption of Vera is transversal and involves leading players in different fields. From finance, with the NYSE aiming to optimize the latency of its infrastructure, to research laboratories such as Anthropic and OpenAI, up to large hyperscalers such as Oracle Cloud Infrastructure and ByteDance. Major system manufacturers, including Dell, HPE, Lenovo and Supermicro, have also already integrated the technology into their hardware roadmaps. The common goal is to support a new generation of AI capable of not only processing responses, but executing code autonomously, using external tools and evaluating the results of its actions.

From a technical point of view, the heart of the system is represented by the Olympus cores, designed to efficiently manage Python runtimes, code execution in a sandbox environment and orchestration logic. With 88 cores, Spatial Multithreading technology, and an LPDDR5X memory subsystem capable of 1.2 TB/s of bandwidth, Vera is designed to eliminate computational bottlenecks that often slow down data pipelines. This allows the accelerators to be constantly powered, drastically improving the overall throughput of the so-called AI factories.

Integration into the existing ecosystem is ensured by the second generation NVLink-C2C interconnect technology, which enables coherent communication with GPUs at up to 1.8 TB/s. Additionally, the Vera BlueField-4 STX processor expands capabilities to include network and storage acceleration, as well as integrated security features. With availability expected starting next autumn, the system is proposed as the first concrete alternative to x86 architectures for the standalone CPU server segment, offering solutions both in liquid-cooled configurations for large clusters and in air-cooled systems for enterprise implementations.