NVIDIA Ships Vera CPU to AWS, OCI and Three AI Labs

NVIDIA began shipping its Vera CPU at scale on August 27, 2026, with Vice President of Hyperscale and HPC Ian Buck hand-delivering the first systems to AWS, Oracle Cloud Infrastructure, Anthropic, OpenAI and SpaceXAI. AWS received its first Vera CPU server and Vera Rubin GPU in Seattle, part of a broader expansion of the 16-year NVIDIA-AWS collaboration that includes plans for 2 million additional NVIDIA GPUs and Vera-based infrastructure on AWS.

The Vera CPU is NVIDIA's first custom processor, built with 88 custom-designed Olympus cores and 1.2TB/s of memory bandwidth. NVIDIA claims up to 1.8x faster per-core performance on agentic AI workloads compared with traditional designs. The company positions Vera as a response to the shift from models that answer questions to agents that act, where CPU work includes tool calls, orchestration layers, long-context retrieval and sandbox execution.

First Deployments

Oracle Cloud Infrastructure is the first cloud provider to deploy Vera at hyperscale, with plans to deploy hundreds of thousands of Vera CPUs beginning in 2026, according to Karan Batta, who leads overall product management at OCI. The company's AI Customer Excellence Center hosted an unboxed Vera system for customer evaluation. Gary Miller, chief customer and partner success officer at OCI, said the team expects customers to customize and validate agentic AI workloads on the new hardware.

At SpaceXAI's Palo Alto offices, NVIDIA's team walked Elon Musk through the system interior, covering cores, memory layout and cooling. SpaceXAI is evaluating Vera for reinforcement learning workloads and agent-based simulation pipelines that drive its training stack.

OpenAI received its Vera server at Mission Bay headquarters, where Sachin Katti, head of compute infrastructure, took the handoff. Buck opened the system with a screwdriver to show the internals. Anthropic's James Bradbury, head of compute, received the first May delivery at the company's San Francisco offices. Bradbury said scaling compute is an important accelerant for model growth and that Vera looks promising for agentic workloads.

Architecture and Positioning

Vera is part of NVIDIA's codesign effort alongside the Rubin GPU, BlueField-4 DPU, Spectrum-X and MGX rack architecture. Beyond standalone CPU systems, Vera serves as the host processor for the Vera Rubin NVL72, pairing via second-generation NVLink-C2C with a pair of Rubin GPUs. In those systems, Vera and Rubin share a unified memory architecture designed to keep accelerated compute utilized. NVIDIA says Vera's cores and interconnect handle orchestration, control and data movement needed to feed GPUs at 2x the energy efficiency of traditional infrastructure.

For inference engineers, the arrival of Vera at multiple cloud providers signals a shift in how agentic workloads get served. The CPU is no longer just a host for GPU work; it is the execution engine for the growing share of agent tasks that never touch a GPU. The 1.2TB/s memory bandwidth and 88 cores target the concurrent, real-time demands of tool calling and long-context operations, where traditional core-density-focused designs fall short. As OCI deploys at hyperscale and AWS brings Vera online, the practical question becomes how these systems hold up under production agent traffic, not just in vendor benchmarks. The hand-delivered systems at Anthropic, OpenAI and SpaceXAI suggest early evaluation is underway at labs that define the agent workload frontier.