NVIDIA has begun shipping Vera systems to partners in the AI ecosystem, moving its custom central processor from the announcement stage to evaluation and deployment at cloud providers and model-development laboratories. Deliveries included AWS and Oracle Cloud Infrastructure, as well as Anthropic, OpenAI, and SpaceXAI. NVIDIA published the material on May 18, 2026, and updated it on August 27, 2026, explaining that Vera shipments are taking place at an increasing scale.
A Processor for Functions GPUs Cannot Perform Alone
NVIDIA presents Vera as its first custom central processor, built from the ground up for agentic AI workloads. The company explains that these workloads involve more than running a model on graphics processing units; they also include running isolated environments, calling tools, managing execution steps, retrieving information from long contexts, and moving data between system components.
According to NVIDIA's specifications, Vera includes 88 custom-designed Olympus cores and provides memory bandwidth of 1.2 terabytes per second. NVIDIA says the processor delivers up to 1.8 times the per-core performance in agentic AI workloads compared with the reference it uses, and also indicates that integration with accelerator units can provide twice the power efficiency of a traditional architecture. These figures come from the company, and the material does not provide details about the testing methodology or the competing systems used for comparison.
Who Received the Systems and What Is Being Evaluated?
AWS received the first Vera server and the first Vera Rubin GPU unit at its headquarters in Seattle, as part of an expansion announced by the two companies for their 16-year collaboration. The announced plans include adding two million NVIDIA GPU units and working to bring Vera-based infrastructure to AWS.
Oracle Cloud Infrastructure said it plans to deploy hundreds of thousands of Vera processors beginning in 2026, describing NVIDIA as the first cloud provider to deploy Vera at hyperscale. The material does not provide a detailed timeline or information about the cloud regions where these systems will be available.
SpaceXAI is evaluating the processor for reinforcement-learning workloads and agent-based simulation pipelines. Anthropic and OpenAI also inspected the systems as part of the initial deliveries, while Anthropic indicated that expanding compute capacity could accelerate model growth, without announcing independent test results in the material.
Vera's Position Within NVIDIA's Platform
NVIDIA does not present Vera merely as a standalone processor, but as part of an ecosystem that includes the Rubin GPU, the BlueField-4 DPU, Spectrum-X networking, and the MGX rack architecture. In the Vera Rubin NVL72 system, Vera operates as a host processor for a pair of Rubin GPU units and connects to them through the second generation of NVLink-C2C within a unified memory architecture.
In practice, NVIDIA is betting that the central processor will handle the control, orchestration, and data-transfer tasks that could limit the sustained utilization of GPUs. This matters to cloud operators and developers of agentic models because increasing the speed of these supporting tasks could affect response time and the operating cost of each task, not merely inference speed.
What Remains Unresolved?
The start of shipping demonstrates that Vera has moved into actual systems at major partners, but it does not by itself establish broad practical superiority across all agent workloads. The source relies primarily on NVIDIA's specifications and claims and does not mention server prices, commercial availability terms, or published comparison results from customers. OCI's announcement of a broad future deployment also remains an announced plan, while Vera's actual viability will require independent evaluation across different model types, tool-use workflows, and cloud architectures.