Nvidia unveiled its OpenShell reference architecture at its annual GTC 2026 Berlin conference on Oct. 21 local time, designed to support the secure deployment and operation of artificial intelligence (AI) agents.
A reference architecture provides a standard blueprint for building new data center systems. The OpenShell reference architecture is a full-stack infrastructure covering AI agent development and operations. It combines Nvidia hardware with OpenShell secure runtime software to isolate and monitor AI agents in large-scale industrial environments.
The hardware consists of Nvidia Vera Rubin NVL72 server racks and BlueField-4 data processing units (DPUs). Rubin graphics processing units (GPUs) handle AI model inference, while Vera central processing units (CPUs) manage AI agent control and tool execution.
The BlueField-4 DPU enforces security policies directly at the hardware level. DPUs are auxiliary processors that handle networking, security and data management operations previously performed by CPUs, improving overall system efficiency.
The architecture also runs Nvidia Sentry integrated security software and a collection of library tools and frameworks. It enables continuous monitoring of agent behavior. If an agent attempts to cross a designated software boundary, Sentry isolates and blocks the operation within milliseconds.
Nvidia’s OpenShell reference architecture is currently available. Dell Technologies, Cisco, HPE, Lenovo and Supermicro are building infrastructure based on the architecture.
Nvidia also introduced its DSX AI Factory platform at the event. The platform provides a standardized data center design architecture that integrates computing, networking, storage, software, power infrastructure and cooling systems from the design stage. It combines closed-loop liquid cooling, which circulates coolant through sealed pipes to absorb heat from chips, with dry coolers that dissipate heat into the surrounding air like large radiators.
The DSX AI platform can reduce cooling water consumption to nearly zero, thanks to Nvidia Vera Rubin’s support for 100% liquid cooling. For a 2-gigawatt (GW) facility, the system can save 54 million liters of water per day compared with conventional cooling methods, equivalent to the daily water consumption of 180,000 households.
