Arm and NVIDIA are joining forces to build a trusted computing foundation for the era of AI-powered agents

From the CPUs powering intelligent agents to the infrastructure ensuring their secure operation, Arm is building a unified computing platform spanning both ends of the trust boundary in the era of intelligent agents

The migration of artificial intelligence (AI) infrastructure to the Arm architecture is accelerating. Today, Arm technology is deployed across the platforms of all major hyperscale cloud service providers. According to IDC data, the market for rack-scale servers based on the Arm architecture has now surpassed that of x86, becoming the mainstream platform in the field of accelerated computing.

As AI enters the era of intelligent agents, this shift is becoming increasingly significant. AI agents are not only capable of generating answers, but can also run continuously, invoke tools, access data and execute tasks autonomously. As the scale of these agents continues to expand, two fundamental computing requirements arise: the computational power needed to run the agents and that required to protect them.

Arm supports both of these computing requirements. On the one hand, high-performance CPUs can be used to run agent workloads; on the other, Arm-based infrastructure processors, such as Data Processing Units (DPUs), provide independent observability, isolation capabilities, policy enforcement and security protection.

NVIDIA’s recently launched NVIDIA Open Agent Safety Platform is a practical implementation of this architectural philosophy.

Arm Computing Platform Powers the Operation of Intelligent Agents

The CPU is at the heart of the AI infrastructure for intelligent agents. Intelligent agent runtimes, execution frameworks, orchestration systems, tool calls and various applications all rely on the CPU. As the capabilities of intelligent agents continue to improve, their runtime continues to increase, and they operate concurrently on an ever-larger scale, the demand for CPUs will continue to grow.

Currently, Arm provides the computational power for the majority of computing tasks in intelligent agents and cloud computing workloads. Whether it is the Arm AGI CPU, Amazon Graviton, Google Axion, Microsoft Cobalt or NVIDIA’s custom-developed Vera CPU, whilst these products each have their own distinctive design features, they all share Arm’s unified software ecosystem. NVIDIA OpenShell extends the secure execution capabilities of intelligent agents to this broad ecosystem of Arm CPUs.

OpenShell establishes a secure runtime boundary for intelligent agents, controlling the resources they access and the operations they perform. This means that, within the increasingly rich and diverse ecosystem of Arm cloud computing and AI infrastructure, a unified approach can be adopted to ensure the secure operation of intelligent agents.

Governance and Control of Agent Behaviour on the Arm Computing Platform

In addition to running agents, the agent infrastructure must also provide independent computing resources—offering observability, isolation capabilities and policy control—separate from the agents’ own runtime environments.

Arm-based DPUs and infrastructure processors provide this independent layer. By separating infrastructure services—such as networking, monitoring, isolation and security—from the host CPU, they establish independent control nodes that can always operate independently of the agents and applications they protect.

This architectural pattern has been extended to the broad Arm DPU ecosystem. As these infrastructure processors are also built on Arm CPU technology, developers and infrastructure providers can build security capabilities based on a unified computing architecture.

NVIDIA BlueField embodies this pattern within the NVIDIA Open Agent Security Platform. Powered by NVIDIA Grace, which features 64 Arm Neoverse V2 cores, BlueField-4 provides a separate infrastructure runtime environment outside the host. OpenShell works in conjunction with NVIDIA Sentry, deployed on BlueField, to extend policy enforcement capabilities from the agent runtime to this independent trust domain. Consequently, infrastructure and network security controls can operate continuously, independent of the agents themselves.

As a real-time monitoring mechanism, NVIDIA Sentry adds a further layer of protection, enabling continuous monitoring of agent behaviour and isolating agents should any unsafe behaviour be detected. These capabilities collectively demonstrate a more comprehensive system-level security architecture: the Arm CPU is responsible for hosting and executing agent workloads, whilst the Arm-based DPU is capable of independently observing, governing, isolating and protecting them.

Arm CPUs host and execute agent workloads, whilst Arm-based DPUs independently monitor, govern, isolate and protect them

Safeguarding the Scalable Development of Agent-Based AI

As agents become increasingly capable, with greater endurance and autonomy, the demand for computing power in agent-based AI will continue to grow. At the same time, security capabilities must scale in tandem with computing power, without the need to compromise between security and performance.

NVIDIA’s Open Agent Security Platform demonstrates this direction: agent workloads run on a variety of computing platforms within the Arm ecosystem, whilst a dedicated Arm-based DPU assists with monitoring, isolation and protection.

As intelligent agents expand from cloud infrastructure to edge devices and physical systems, these security boundaries must also extend to cover the diverse scenarios in which intelligent agents operate and execute tasks. As Arm increasingly becomes a core component of AI infrastructure, it is driving the entire ecosystem towards the synchronised scaling of both performance and security.

The era of intelligent agents is being built upon Arm, and Arm remains committed to laying a solid security foundation for this era from the very beginning.