Arm says its CPUs and DPUs secure both sides of agentic AI with NVIDIA

Two compute needs for AI agents

In a blog post published on September 28, 2026, Arm argues that AI infrastructure is increasingly moving to its architecture. According to the company, Arm technology is deployed across every major hyperscaler, and, citing IDC, Arm-based rack-scale servers have now overtaken x86 as the dominant platform for accelerated computing.

Arm says the agentic era raises the stakes. AI agents do not only generate answers. They run continuously, use tools, access data and take actions to complete tasks. As they scale, Arm describes two basic requirements: compute to run the agent and compute to protect it.

According to Arm, it covers both: high-performance CPUs for agentic workloads, and Arm-based infrastructure processors, such as DPUs, for independent observation, isolation, policy enforcement and security.

CPUs run the agent, OpenShell contains it

Arm says the CPU sits at the center of agentic AI infrastructure. Agent runtimes, harnesses, orchestration, tool execution and applications all depend on it, and the demands grow as agents become more capable and persistent and run concurrently at scale.

The company lists several CPUs that it says are architecturally distinct but compatible with the Arm software ecosystem:

• the Arm AGI CPU
• AWS Graviton
• Google Axion
• Microsoft Cobalt
• NVIDIA's custom Vera CPU

NVIDIA OpenShell, according to Arm, brings secure agent execution to this broad CPU ecosystem. It establishes a secure runtime boundary around the agent, governing what it can access and the actions it can take. Arm says this means the same basic approach to secure execution can extend across a growing range of Arm-based cloud and AI infrastructure.

DPUs as an independent control point

Arm says running the agent is only half the picture. Agentic infrastructure also needs independent compute that can observe, isolate and govern policies outside the environment where the agent runs. Arm-based DPUs and infrastructure processors provide that layer, according to the company.

By moving services such as networking, monitoring, isolation and security away from the host CPU, these processors create a separate control point that stays independent of the agents and applications it protects. Because they are built on Arm CPU technology, Arm says developers and infrastructure providers can build security capabilities on a common compute architecture.

The post uses NVIDIA BlueField as its example. BlueField-4, powered by NVIDIA Grace with 64 Arm Neoverse V2 cores, provides an independent infrastructure environment beyond the host. OpenShell integrates with NVIDIA Sentry on BlueField to extend policy enforcement from the agent runtime into this separate trust domain, so infrastructure and network controls can operate independently of the agent itself.

NVIDIA Sentry is described as a real-time watchdog that monitors agent behavior and allows an agent showing unsafe behavior to be quarantined.

What it means for scaling agents

Arm's central claim is that agentic AI will need more compute as agents become more capable, persistent and autonomous, and that security must scale alongside it without forcing a tradeoff between the two.

The company adds that as agents move from cloud infrastructure to edge devices and physical systems, these security boundaries will need to follow them wherever they run and act. Arm says it intends to help the ecosystem scale performance and security together.

The post is Arm's own account of the architecture and of NVIDIA's platform, and it does not give performance figures, pricing or availability details for the Open Agent Safety Platform.