Nvidia brings Vera CPU to AI servers

Nvidia is extending its AI hardware push beyond GPUs with Vera, a data center CPU built for agentic AI and reinforcement learning. The company said Vera chips were delivered to Anthropic, OpenAI, and SpaceX in June 2026, and OpenAI plans to begin deploying the processors in large quantities starting this quarter. Oracle is the only major cloud provider currently named among Vera’s listed partners.

Vera uses 88 Nvidia-designed Olympus cores, supports Armv9.2, and uses spatial multithreading to enable 176 total threads. The CPU supports up to 1.5 TB of LPDDR5X memory, up to 1.2 TB/s of memory bandwidth, a 3.4 TB/s scalable coherency fabric, and 1.8 TB/s of NVLink-C2C bandwidth, with power consumption ranging from 250 watts to 450 watts. Nvidia says Vera delivers 50% better performance for AI agent workloads than x86 processors, with an emphasis on single-core speed, memory bandwidth, and low latency.

The launch puts Nvidia against AMD and Intel in server CPUs as AMD reached 33.2% of x86 server CPU unit share in the first quarter (Q1) 2026, while Intel kept 66.8% of unit shipments. Nvidia estimates the server CPU opportunity could reach $200 billion USD, and Wolfe Research expects Vera to average roughly $5K per chip and reach 1.3M shipments this year (2026). Vera will be sold independently, in single and dual-socket configurations, and as part of a liquid-cooled rack with up to 256 CPUs, as well as Vera Rubin systems where the NVL72 configuration pairs 36 Vera CPUs with 72 Rubin GPUs.