NVIDIA's Vera CPU is a purpose-built data center processor targeting AI factory workloads, particularly reinforcement learning post-training and agentic inference pipelines. It features 88 custom Olympus cores (NVIDIA's first fully custom data center CPU core), 1.2 TB/s LPDDR5X memory bandwidth, and a second-generation Scalable Coherency Fabric delivering 3.4 TB/s bisection bandwidth. Key claims include up to 1.5x agentic sandbox performance versus competitive x86 platforms and 2x performance-per-watt at rack scale. The Vera CPU Rack supports over 22,500 concurrent sandboxes per rack. Vera-based platforms span standalone CPU racks, single/dual-socket servers, and integrated Vera Rubin NVL72 GPU systems. OEM availability from Cisco, Dell, HPE, Lenovo, and Supermicro is expected in H2 2026.
Table of contents
The post-training realityPerformance across the AI factoryAgentic environments by the rackVera platforms and configurationsPlatform availabilitySort: