CoreWeave to Run NVIDIA Vera CPU Bare Metal at Rack Scale

CoreWeave said on September 30, 2026, that it will expand its compute portfolio with the NVIDIA Vera CPU, which it described as the first CPU designed for AI agents, and will run the processor bare metal at rack scale for agentic AI and reinforcement learning workloads.

The news was shared at Fully Connected, CoreWeave’s AI cloud conference in San Francisco, which the company said brings together more than 4,500 customers, partners, developers, and AI leaders. Fully Connected 2026 runs September 29 through October 1, 2026, at Moscone South, 747 Howard Street.

The CPU Work Behind Agentic AI

CoreWeave described agentic AI as operating through a continuous loop of run, observe, curate, improve, and evaluate. Within that loop, GPUs handle model training and reasoning, while CPUs run the work around each step: isolated sandboxes, reinforcement learning environments, tool calls, code execution, and data pipelines.

CoreWeave called demand for data preprocessing pipelines and reinforcement learning training episodes bursty and difficult to predict: a single step in the loop can call for thousands of environments for an hour, then almost none until the next run. As post-training and agentic workloads grow, the company said, that CPU-intensive work increasingly sets the pace of the AI loop, and infrastructure needs both the density for that demand and the flexibility to scale as the demand changes.

The company said the Vera CPU runs the same way as the rest of its fleet on CoreWeave: bare metal, under the same platform, consumption models, and economics.

Rack-Scale Deployment and CoreWeave Sandboxes

CoreWeave said one critical performance measure for agentic AI is the number of isolated agent environments a team can run at the same time, and whether per-sandbox performance stays consistent as that count rises. Its rack-scale Vera deployment fits 128 CPUs and 11,264 cores into a single rack, with BlueField-4 DPUs and Spectrum-X Ethernet switching connecting the Vera nodes; the company said that provides capacity for more than 11,000 concurrent environments.

CoreWeave said its Sandboxes product deploys these agents with full hardware isolation, placed directly alongside the training jobs they support. In its testing, the company reported, agent sandbox startup times on the NVIDIA Vera CPU were more than 3x faster than on an x86 CPU.

“General-purpose infrastructure bottlenecks agentic AI; Vera is the first CPU explicitly designed to accelerate it,” Chen Goldberg, executive vice president of product and engineering at CoreWeave, said in the company’s announcement. Goldberg said CoreWeave’s platform natively supports Vera through products such as CoreWeave Sandboxes, letting teams immediately spin up thousands of isolated environments without added operational friction.

Ryan Shrout, president and general manager at Signal65, said a growing share of every agent cycle is CPU work, and that the share will only increase as post-training and agentic deployments scale. He said the surrounding platform determines whether teams can actually use that capacity, and that the operational difference comes from running the CPU work on bare metal alongside the training fleet under orchestration that already exists.

CoreWeave launched CoreWeave Sandboxes on May 14, 2026, as an execution layer that provides secure, isolated environments where AI researchers and platform teams run reinforcement learning, agent tool use, and model evaluation. It is available on a customer’s own CoreWeave infrastructure through CoreWeave Kubernetes Service or as a serverless runtime through Weights & Biases, with a Python SDK that includes session management, storage integration, and monitoring tools. CoreWeave states that by default each sandbox operates in a fully isolated virtual environment of its own, so a failure or runaway process in one cannot spill into another.

CoreWeave also pointed to its MLPerf benchmark results in inference and training and its position as the only AI cloud to earn the top Platinum ranking in SemiAnalysis ClusterMAX three times in a row.

NVIDIA Vera CPU Specifications

NVIDIA’s Vera CPU product page describes the processor as built for the CPU work behind agentic AI and reinforcement learning: code execution, tool use, sandboxing, analytics, data pipelines, and orchestration beyond the model. According to NVIDIA, Vera carries 88 custom Olympus cores, and its Spatial Multithreading creates 176 threads with partitioned core resources.

Memory bandwidth reaches up to 1.2 terabytes per second on LPDDR5X, with up to 1.5 TB of memory capacity, NVIDIA states. A second-generation NVIDIA Scalable Coherency Fabric links all 88 cores, cache, memory, IO, and NVLink-C2C across a single compute die with 3.4 TB/s of bisectional bandwidth, and NVLink-C2C provides up to 1.8 TB/s of coherent bandwidth between Vera CPUs and NVIDIA GPUs, according to the company.

NVIDIA says Vera runs the compilation, code analysis, and Python tool-chain workloads of an agent’s reasoning loop up to 1.8x faster than leading x86 CPUs, and that its LPDDR5X memory subsystem delivers 2x the bandwidth and 3x the bandwidth per core of leading x86 CPUs with DDR5. The product page notes that those relative performance figures are based on measured data, baselined to the latest-generation x86 CPU, and subject to change.

NVIDIA positions Vera both as a host CPU for accelerated systems, including the Vera Rubin NVL72 and HGX Vera Rubin NVL8 platforms, and as a standalone CPU for agentic AI, reinforcement learning, data processing, and analytics. The Vera CPU Rack, built on NVIDIA MGX, packs up to 256 Vera CPUs and runs over 22.5K concurrent environments, according to the company. Vera Rubin NVL72 combines 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, and BlueField-4 DPUs in one rack-scale platform.