Arm and NVIDIA: Building the trusted compute foundation for the agentic AI era
Arm and NVIDIA outline a security-focused AI infrastructure framework using Arm CPUs and DPUs to support scalable, agentic workloads while enforcing independent policy controls.
Arm technology now underpins accelerated computing in rack-scale servers, surpassing x86 as the dominant platform according to IDC, a shift critical for the emerging agentic AI era. AI agents require continuous operation, tool access, and task execution, creating dual compute needs: performance for running agents and security for protecting them. Arm provides both through high-performance CPUs for agent workloads and Arm-based DPUs for independent observation, isolation, and policy enforcement. NVIDIA’s Open Agent Safety Platform integrates Arm-compatible CPUs with BlueField-4 DPUs to demonstrate this architecture in practice.
The Open Agent Safety Platform places Arm CPUs at the center of agentic AI infrastructure, where agent runtimes, orchestration, and tool execution depend on compatible Arm-based solutions like AWS Graviton, Google Axion, and Microsoft Cobalt. NVIDIA’s OpenShell adds a secure runtime boundary around agents, governing access and actions across the Arm ecosystem. This approach ensures consistent security controls as agent workloads scale across diverse cloud and AI environments. Arm-based DPUs, such as those in NVIDIA BlueField, provide an independent layer for networking, monitoring, and isolation, reducing reliance on host CPUs.
NVIDIA Sentry, integrated with BlueField-4, monitors agent behavior in real time, enabling unsafe agents to be quarantined without disrupting the broader system. BlueField-4, powered by NVIDIA Grace with 64 Arm Neoverse V2 cores, offers a separate trust domain for infrastructure services, extending policy enforcement beyond the agent runtime. This model allows security controls to operate independently of agent workloads, addressing scalability challenges in agentic AI deployments. The architecture aligns with Arm’s expanding role in AI infrastructure across hyperscalers and edge devices.
As agentic AI expands from cloud to edge and physical systems, Arm’s compute foundation supports both performance and security without tradeoffs. The Open Agent Safety Platform exemplifies this balance, using Arm CPUs for agent execution and Arm-based DPUs for protection. Arm emphasizes its commitment to scaling secure agentic AI infrastructure, ensuring security measures evolve alongside computational demands across diverse environments.