Nvidia unveiled the Vera CPU at GTC 2026-the first chip purpose-built for agentic AI workloads and reinforcement learning.
As AI shifts from chatbots to autonomous agents that reason, code, and use third-party tools, infrastructure demands have evolved. GPUs handle model inference and training; CPUs now manage orchestration, data validation, and tool coordination-tasks where Vera excels.
Vera delivers 50% higher performance and double the efficiency of x86 CPUs for these workloads. It features 88 custom Olympus cores with Spatial Multithreading and a low-power LPDDR5X memory subsystem delivering 1.2 TB/s bandwidth-twice the throughput of general-purpose CPUs at half the power.
The chip powers Nvidia’s new Vera CPU rack system-256 liquid-cooled chips supporting over 22,500 concurrent CPU environments-and is tightly integrated into the NVL72 platform: 72 Rubin GPUs and 36 Vera CPUs linked via NVLink 6, enabling 1.8 TB/s coherent CPU-GPU bandwidth.
Early testing shows 5.5x lower latency on Apache Kafka workloads. Hyperscalers Oracle, Meta, and Alibaba, plus neoclouds CoreWeave and Nebius, are first adopters. Dell, HPE, Supermicro, and Lenovo will launch Vera-based servers in H2 2026.
Vera is in full production now.