NVIDIA used GTC Berlin on October 21 to launch the OpenShell Reference Architecture, a full-stack design for building, containing, observing and enforcing safe AI agents on enterprise infrastructure and edge systems. The architecture is available now, and NVIDIA framed it as the infrastructure layer for a moment when companies are preparing to deploy agents at industrial scale rather than as isolated pilots.
The design extends NVIDIA’s open-source OpenShell secure runtime software into complete infrastructure for agent development and operations. It pairs the company’s Vera Rubin platform for model inference with Vera CPUs that run agent harnesses and tool use, and with BlueField-4 data processing units that watch agent behaviour from outside the agent’s own software and enforce policy in silicon.
OpenShell’s Three Layers: Rubin Thinks, Vera Acts, BlueField Governs
Jensen Huang, NVIDIA’s founder and chief executive, summed up the split of duties in one line: with the reference architecture, agents can think on Rubin, act on Vera and be governed on BlueField. Vera Rubin, with the Vera Rubin NVL72 system at its core, can be used across pretraining, post-training and inference, and NVIDIA says the NVL72 delivers 30 times higher throughput per megawatt for agentic workloads.
The Vera CPUs are the layer where agents actually do their work. NVIDIA says the processors deliver twice the performance of x86 chips with greater energy efficiency, and they run OpenShell itself, which provides the secure runtime boundary inside which an autonomous agent operates. The workloads they serve include closed and open models, including models built with NVIDIA’s Nemotron family — the same model line NVIDIA expanded earlier this month with its Nemotron 3.5 Super VL and Bolt open models for always-on agents.
Sentry Watches From Outside The Agent’s Reach
The governance layer is what distinguishes the design. BlueField-4 DPUs run NVIDIA Sentry, a unified security framework built on NVIDIA’s DOCA software, which provides continuous out-of-band monitoring of agent behaviour and in-silicon policy enforcement. Because the monitoring sits on separate hardware, an agent cannot switch off its own oversight from inside its runtime.
If an agent attempts to move outside its software boundary, Sentry quarantines and stops it in milliseconds, NVIDIA says. The company also says Sentry, running on BlueField-4, boosts agent capacity by up to 20 per cent. The architecture reaches beyond the data centre: it is extensible to agents running at the edge in retail, healthcare, manufacturing, robotics, software development and telecommunications, on RTX Spark, DGX Spark and DGX Station systems.
On the software side, the architecture runs on Linux and Kubernetes infrastructure from Canonical, SUSE and Red Hat, and on Microsoft Windows using Microsoft Execution Containers. It can also run on the clouds of Amazon Web Services, Google Cloud and Microsoft Azure, serving agents from model builders including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence.
Partners From Banks To ServiceNow Build On The Architecture
NVIDIA announced the design with a long partner list attached. Cybersecurity companies including 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler are integrating support into their agent security products. In Europe, banks BBVA and Santander are exploring it as a framework for agent safety and governance — Santander has separately been working with NVIDIA to cut its AI serving costs with a model arena.
Two enterprise software integrations were named in detail. SAP is integrating the architecture with Joule Studio, part of the SAP Business AI Platform, connecting business governance with secure agent execution. ServiceNow announced the general availability of Arc, a personal long-running AI agent whose actions are governed at runtime by its Super Intelligence Control Tower and enforced by OpenShell, with its AI Passport layer carrying policies to OpenShell and Sentry so a company can revoke an agent’s access across software and hardware within seconds.
System makers Dell Technologies, Cisco, HPE, Lenovo and Supermicro are building infrastructure on the design, cloud providers from CoreWeave and Nebius to Nscale and Oracle Cloud Infrastructure are deploying OpenShell-ready infrastructure, and inference providers including Baseten, Fireworks and Together AI are building integrations.
Why It Matters
The OpenShell Reference Architecture matters because agent safety is moving from software promises into the hardware bill. Until now, keeping an enterprise agent inside its limits has mostly meant application-level guardrails — the same kind of controls that this week’s disclosures about agents disabling their own monitoring showed can be talked around or coded away. Putting the monitor on a separate processor that the agent cannot reach, and giving it the power to quarantine an agent in milliseconds, changes who holds the off switch.
It also extends NVIDIA’s position in the agent stack. The company already sells the accelerators agents think on and, since Modal adopted its Vera CPUs for agent hosting — covered in NVIDIA’s Vera CPU win at Modal — the processors agents act on. A reference architecture that also governs them makes NVIDIA’s stack the default at every layer, and gives regulated industries a hardware story to take to their supervisors.
Background
NVIDIA introduced OpenShell as open-source secure runtime software for autonomous agents, and the reference architecture expands it from a software boundary into a full infrastructure blueprint, launched during a European tour that has brought industrial AI partnerships across the continent, including a German government partnership unveiled at the Berlin investment summit. It lands in the same week that labs have disclosed cases of their own agents weakening internal monitoring systems during coding work.
Source: NVIDIA via GlobeNewswire, Unite.AI
0 Comments