NVIDIA Unveils OpenShell Reference Architecture

What happened

NVIDIA announced on its official blog the OpenShell Reference Architecture, unveiled at GTC Berlin and built on the OpenShell secure runtime software. The architecture aims to give organizations full-stack control when deploying AI agents: agents do what they are meant to do without moving beyond the boundaries people set.

Key facts

1. Three layers: NVIDIA Vera Rubin NVL72 handles model inference; Vera CPUs run OpenShell for agent harnesses and tool use; NVIDIA BlueField-4 DPUs run the NVIDIA Sentry security framework for in-silicon enforcement. Each plays a distinct role and they work together.
2. In-silicon enforcement: Sentry on BlueField-4 DPUs performs out-of-band monitoring plus in-silicon policy enforcement — intercepting out-of-bounds behaviour at the chip level, outside the host operating system.
3. Cloud partner support: AI cloud providers including CoreWeave, Nebius, Nscale, Lambda, Crusoe, DigitalOcean, Vultr, G42, IREN, Fireworks and Baseten announced support for the architecture.
4. Launch context: The official blog framed the release under “sovereign” European AI clouds, “open” models and libraries, and “safe” agent boundaries — aimed at developers connecting agents to tools and business systems.