By Hamza Faisal

Nvidia wants your AI agents to be powerful — and locked in a glass box. At GTC Berlin on Tuesday, CEO Jensen Huang unveiled the NVIDIA OpenShell Reference Architecture, a full-stack hardware-and-software blueprint for running enterprise AI agents inside enforced boundaries. "AI capability and safety must advance together," Huang said. "As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture."

The architecture builds on NVIDIA's OpenShell secure runtime software and extends it into industrial-scale infrastructure for agent development and operations. The slogan Nvidia is pushing: agents can think on Rubin, act on Vera and be governed on BlueField.

Think on Rubin, act on Vera, governed on BlueField

The stack pairs three silicon pillars. Vera Rubin infrastructure handles model inference and enterprise workloads, with Vera Rubin NVL72 systems delivering 30x higher throughput per megawatt for agentic workloads. Vera CPUs — which Nvidia says deliver 2x the performance of x86 processors — run agent harnesses and tool use.

The enforcer is the BlueField-4 DPU, running the NVIDIA Sentry unified security framework. Sentry provides continuous out-of-band monitoring of agent behavior and in-silicon policy enforcement: if an agent attempts to move outside its software boundary, Sentry quarantines and stops it in milliseconds. Nvidia says the combination boosts agent capacity by up to 20%. The architecture also stretches to the edge across retail, healthcare, manufacturing, robotics and telecoms via RTX Spark, DGX Spark and DGX Station systems.

Rivals become partners — everyone signs on

The partner list reads like the entire industry deciding safety is now Nvidia's layer. Model builders including Anthropic, OpenAI, Google, Meta, Cohere, Perplexity and SpaceX Super Intelligence are working with OpenShell. Cybersecurity leaders CrowdStrike, Palo Alto Networks, Cisco, Fortinet, 1Password, Check Point, F5, Okta, TrendAI and Zscaler are integrating support into their agent security products.

Enterprise software giants SAP, ServiceNow, IBM, Siemens, Dassault Systèmes, Cadence and Synopsys are building agent platforms on it, while Dell, HPE, Lenovo, Cisco and Supermicro ship OpenShell-ready infrastructure. Cloud providers from CoreWeave and Crusoe to Nebius, Nscale, Lambda and Oracle Cloud Infrastructure are deploying it, and inference providers Baseten, Fireworks and Together AI are wiring it into their pipelines. Even Europe's BBVA and Santander are exploring it as a governance framework. The reference architecture is available now.

ServiceNow Arc shows what it looks like in production

The clearest real-world proof came the same day: ServiceNow announced general availability of ServiceNow Arc, a personal long-running AI agent whose actions are governed at runtime by the ServiceNow Super Intelligence Control Tower and enforced by NVIDIA OpenShell. ServiceNow's new AI Passport authority layer carries policies down to OpenShell and Sentry, so an enterprise can revoke an AI agent's access across software and hardware within seconds.

Coming a day after OpenAI disclosed two incidents where its own agents weakened internal safety monitoring, Nvidia's timing is impeccable. The message to the industry: agents get their freedom, but the walls are silicon.

By Hamza Faisal