Nvidia Unveils Vera CPU Rack Specification for Agentic AI

What happened

On Oct. 12, Nvidia unveiled its Vera CPU Rack specification, introducing a server rack design centered on central processing units (CPUs) for the first time, optimized for agentic artificial intelligence (AI) workloads. According to Nvidia, agentic AI relies heavily on CPU computing; under its architecture, the GPU-based Vera Rubin NVL72 rack handles traditional large language model (LLM) training and inference, while the Vera CPU Rack is designed to process agentic AI workloads. The Vera CPU Rack is based on Nvidia's third-generation MGX architecture; MGX is a modular reference specification for data center servers that Nvidia provides to server manufacturers, with partners including Dell, Supermicro and Quanta using it to design and manufacture their own server products. The rack integrates computing, networking, power and cooling systems into a common modular platform. Nvidia also said it submitted designs for the Vera Rubin rack and drawer-style compute trays based on MGX 1.2, an updated version of its third-generation MGX specification, to the Open Compute Project's (OCP) Open DC Initiative, intended to help server manufacturers develop next-generation AI factory infrastructure based on Vera Rubin and bring it to market more quickly. MGX 1.2 is compatible with 19-inch server rack equipment, supports up to 250 kilowatts (kW) per rack, and introduces rack-level liquid cooling; Nvidia said standardized interconnects, reconfigurable input/output (I/O) slots and fanless compute trays reduce assembly and maintenance time.

Impact and signal

Affected components and suppliers

Source and time

Source: The Elec · 10/14/2026 · View original source