Nutanix Plans NVIDIA Vera CPU Support to Run Agentic AI Workloads

Why agents put more weight on the CPU

Nutanix product management senior director Oscar Wahlberg writes that large language models still mostly depend on GPUs for inference. Agents, however, do more than consume tokens. According to the post, they autonomously execute multi-step business workflows, compile dynamic code, query enterprise databases and interact directly with line-of-business applications. All of these activities run on CPUs.

The post lists what an agent may need to do: execute code in a controlled environment, invoke APIs and enterprise tools, retrieve information from databases and business systems, coordinate steps across a workflow while applying policies and safety controls, communicate with other agents or services, and validate results before acting. Nutanix describes this as a continuous execution loop around the model. In its words, the GPU may generate the response, but the CPU coordinates much of what happens before, between and after inference.

Nutanix points to three areas: sandboxing and tool execution, integration with enterprise systems such as ERPs, CRMs, databases and analytics pipelines, and orchestration of data pipelines, model routing, retrieval-augmented generation (RAG) and safety guardrails. The company says that as enterprises scale from tens to tens of thousands of concurrent agent workflows, CPU performance becomes a critical factor in AI-factory efficiency, and a slow CPU layer turns the infrastructure around the GPU into a bottleneck.

Vera support starts with NKP Bare Metal

Nutanix describes NVIDIA Vera as designed for this changing workload profile, with a focus on high-performance CPU execution, power efficiency, and the demands of agentic AI and data processing. The company frames the goal as a balanced foundation in which CPU and GPU resources work together across the full agent execution path.

Support will begin with NKP Bare Metal for highly performance-sensitive AI-factory environments. Nutanix says it is also evaluating support for virtualized Nutanix environments. The stated aim is to let IT teams deploy, operate and govern Vera CPU systems with the same management platform, storage fabric and hypervisor stack they use across their hybrid multicloud footprint, instead of building fragmented bare-metal AI clusters that add management overhead.

The post carries a forward-looking statement notice. Nutanix says some features remain in development and will be offered on a when-and-if-available basis, and it gives no dates or pricing.

Security and control features Nutanix highlights

Because agents routinely generate and run code dynamically, Nutanix says running untrusted agentic code in bare-metal containers without a sandbox poses serious security risks. The post names four building blocks:

• VM-level sandboxing: Nutanix Cloud Platform can run Kubernetes on virtual machines, using the AHV hypervisor, Flow virtual networking and multi-tenant aware resource isolation.
• Sandboxing for containers: for customers who prefer NKP Bare Metal, Nutanix plans to integrate with NVIDIA OpenShell to provide secure, private agent sandboxing.
• DPU-accelerated Zero Trust networking: offloading Flow virtual networking and micro-segmentation to NVIDIA BlueField DPUs removes network security processing from the CPU, which Nutanix says preserves Vera performance for agent execution and tool calling.
• Central control plane: Nutanix Enterprise AI includes the Nutanix Agent Gateway, offering token cost visibility, rate limiting, access control for Model Context Protocol (MCP) servers, and integration with NVIDIA NeMo Guardrails, designed to help detect and mitigate prompt injection attacks.

What it means for enterprise IT

The post signals that Nutanix wants enterprises to treat agentic AI as a full-stack operations question covering models, CPUs, GPUs, storage, networking, runtime security, governance and operations, not only GPU capacity. For customers already running Nutanix, the pitch is to manage Vera-based agent environments with existing tools and processes.

Nutanix says pairing NVIDIA's CPU architecture with Nutanix Cloud Platform is intended to support predictable performance and a low-latency foundation for running advanced AI agents. The post offers no benchmarks, availability dates or pricing, so the practical impact will depend on what Nutanix and NVIDIA deliver.