Nvidia launches Vera AI-agent processor and ships it to AWS

Ian Buck, vice president of hyperscale and HPC at Nvidia, physically delivered the first Vera CPU server and the Vera Rubin GPU to Willem Visser, vice president of EC2, and Supreeth Sheshadri, vice president of hardware engineering at EC2, at the AWS data center in Seattle. The handover marks the start of practical deployment of Nvidia’s first custom-built processor for agentic AI workloads.

Vera is equipped with 88 Olympus cores, a self-designed Nvidia architecture, and memory bandwidth of 1.2 TB/s. According to the company, it delivers up to 1.8× improvement in single-core performance on agentic AI workloads. The numbers come from the manufacturer and not from independent tests; Nvidia has not yet published benchmark results such as SPEC or MLPerf that would allow direct comparison with AMD EPYC or Intel Xeon processors. Buck explained that the shift from models that answer to models that act—writing Python, running simulations, performing long-context retrieval—creates CPU load that did not exist before, and architectures focused on core density were not designed for it.

Karan Batta, who leads product management at Oracle Cloud Infrastructure, said OCI plans to deploy hundreds of thousands of Vera processors beginning in 2026. He added that the architecture is designed for high-throughput, continuous reasoning workloads, offering efficiency, density and footprint that match OCI customers’ scaling needs. Gary Miller, head of customer and partner success at OCI, joined a tour of Oracle’s excellence center where the system was evaluated against global customer workloads.

Alongside the Vera handover, Nvidia and AWS announced an expansion of their 16-year collaboration, including plans for an additional 2 million Nvidia GPUs and joint work to bring Vera-based infrastructure to AWS. The Seattle delivery is the latest step in a rapid rollout: Vera systems had previously arrived at OCI and at three leading AI labs, Anthropic, OpenAI and SpaceXAI. The original blog was posted on 18 May 2026 and updated on 27 August 2026 with details of the AWS delivery.

Nvidia has not disclosed clock speeds, TDP, unit pricing or broad commercial availability. No performance metrics under production conditions or real-world customer loads have been released, only the vendor’s claims about self-defined “agentic AI” workloads. Until independent benchmarks and early deployment reports appear, Vera remains an architecturally interesting promise rather than a proven fielded product.