Nvidia is evolving its data processing units from traditional infrastructure offload devices to critical elements that enforce security and control across complex AI operations. This move targets safer deployment of agentic AI systems that interact dynamically within enterprise AI environments.
- Scale-in infrastructure extends security into AI factory operations.
- BlueField DPUs central to controlling agentic AI activity.
- OpenShell runtime enforces agent permissions independently.
Market signal
Nvidia is signaling a strategic expansion of its infrastructure portfolio through the concept of 'scale-in,' designed to address emerging needs driven by agentic AI workloads. Unlike traditional infrastructure categories focused on networking between and outside compute resources, scale-in aims to govern AI agents that perform multiple interconnected tasks by enforcing security policies within the infrastructure itself.
This innovation responds to growing concerns around AI safety and operational trust in enterprise AI deployments. Nvidia’s approach leverages its BlueField data processing units (DPUs) combined with the DOCA software stack to create a separate control plane that monitors and restricts agent access to data and services while minimizing overhead on core GPU and CPU compute resources.
Operator impact
For operators of AI infrastructure, Nvidia’s scale-in introduces a new layer of security and operational control to manage agentic AI’s complex behaviors. By deploying BlueField DPUs with OpenShell's sandboxed agent runtime, operators can define and strictly enforce what resources agents may access and what actions they can execute without relying solely on application-level controls.
This helps mitigate risks associated with unmonitored autonomous AI activity, enabling more predictable and secure AI workflows. It reduces the system burden on primary GPUs and CPUs, offloading security processing to dedicated hardware, thereby improving overall efficiency and potentially simplifying compliance with internal and external AI safety regulations.
What to watch next
Market participants should closely observe Nvidia’s ecosystem adoption of scale-in infrastructure, including uptake by AI platform builders and integration with frameworks such as Codex, Claude Code, Pi, and Hermes. The competitive landscape for DPU software development and adoption will also be critical as rivals may seek to develop similar agent control capabilities.
Additionally, the evolution of OpenShell and its formal policy enforcement mechanisms will be key indicators of how effectively Nvidia can address real-world security challenges posed by agentic AI systems. Watching Nvidia’s collaboration with cloud and enterprise AI customers will provide insight into how scale-in infrastructure influences broader AI factory architecture standards and operational practices.