Nvidia announced Scale-In, a new pillar for its AI networking stack. The system uses the BlueField‑4 data processing unit and DOCA software to keep agentic AI workloads sandboxed and watched in real time. Nvidia says the design adds tenant isolation, out‑of‑band telemetry and hardware‑based enforcement, so agents can’t stray from their assigned tasks. The company also rolled out an Open Agent Safety Platform and OpenShell runtime earlier this year to support the new architecture.
Why it matters
Enterprises running multiple autonomous AI models can rely on hardware‑level checks to prevent rogue behavior, reducing risk of data leaks or misuse.
