NVIDIA has introduced Vera, a CPU engineered specifically for the demands of agentic AI workloads. This marks a clear break from traditional processors built for general-purpose tasks.

Agentic AI agents don’t just generate text—they execute code, manage multi-step workflows, run sandboxed tools, and continually evaluate results. CPUs now form a critical part of the AI processing loop, not just feeding GPUs. Vera addresses this new reality by delivering 1.8 times faster task completion than standard x86 processors.

Built with 88 custom Olympus cores, Vera emphasizes fast per-core performance under heavy concurrency. It supports up to 1.2 terabytes per second of LPDDR5X memory bandwidth, ensuring that data keeps flowing without bottlenecks. This architecture sustains high throughput on branch-heavy, memory-sensitive code typical in AI orchestration and reinforcement learning.

Unlike older CPU designs that chased core count and cloud virtualization economics, Vera prioritizes tokens per dollar—measuring AI output rather than raw core numbers. Each AI agent runs sequential steps that depend on the CPU’s speed, making per-core performance critical. Vera’s neural branch predictors and deep out-of-order scheduling achieve up to 50% higher instructions per cycle compared to prior NVIDIA CPUs.

This CPU is at the heart of NVIDIA’s broader AI factory vision. Vera powers standalone servers and forms the host CPU in the Vera Rubin platform, tightly integrated with Rubin GPUs through NVLink-C2C. This unifies memory between CPU and GPU, doubling energy efficiency while pushing total throughput to new heights.

The Vera ecosystem now includes BlueField-4 STX, a platform blending Vera CPUs with AI-native storage acceleration and in-silicon security. This addresses data movement and storage bottlenecks, enabling secure, high-performance AI pipelines that handle constant data retrieval and update operations.

Adoption is already underway among top AI labs and hyperscalers. Anthropic, OpenAI, SpaceXAI, Oracle Cloud Infrastructure, ByteDance, and CoreWeave are testing or deploying Vera to scale agentic workloads. Major OEMs like Dell, HPE, Lenovo, and Supermicro are building systems with Vera CPUs, marking the first standard CPU option beyond x86 for AI-centric data centers.

In real-world terms, Vera reduces latency in multi-step AI reasoning, speeds up reinforcement learning evaluations, and increases the throughput of agentic tasks. This can save millions in token computing costs for large-scale deployments running tens or hundreds of thousands of agents simultaneously.

Industry benchmarks confirm Vera’s lead in AI-critical tasks like code compilation, Python execution, Java workloads, and database processing. These functions form the backbone of modern AI factories, where CPU speed directly impacts agent interactivity and efficiency.

NVIDIA’s push with Vera signals a strategic shift from GPU-only dominance to full-stack AI infrastructure innovation. The company now offers a coordinated hardware stack—CPU, GPU, networking, storage, and system orchestration—optimized for autonomous AI agents and their complex, multi-step workflows.

This evolution reflects a fundamental change in AI infrastructure. Future supercomputers will look less like traditional HPC clusters and more like dynamic AI factories. They must juggle thousands of concurrent agents, persistent memory states, fast context switching, and continuous data orchestration. Vera is designed for precisely that challenge.

By redefining CPU performance for agentic AI, NVIDIA stakes a claim in the next frontier of AI computing—where efficiency and speed in CPU-bound orchestration tasks are just as vital as raw GPU power. The era of AI agents demands processors built not just to compute, but to think and act at scale.