Skip to content
NewA calmer, faster ZeroHour
StoryAI industryinfo

NVIDIA OpenShell architecture aims to bound enterprise AI agents

3 sources · 6 articlesFirst reported Updated ()1 · 1 read
What's new

First merged summary.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

NVIDIA’s OpenShell reference design runs models on Vera Rubin, agent software on Vera CPUs, and policy enforcement on BlueField-4.

NVIDIA has introduced the OpenShell Reference Architecture to create, operate, observe, and keep enterprise AI agents inside defined boundaries. Models such as Nemotron run on Vera Rubin, agent harnesses and tools run on Vera CPUs under OpenShell, and BlueField-4 DPUs run NVIDIA Sentry to monitor behavior and enforce policy. NVIDIA says Sentry can isolate an agent that leaves its software boundary within milliseconds and raise agent capacity by up to 20 percent. The design targets RTX Spark, DGX Spark, and DGX Station, with support for Linux, Kubernetes, Windows, and cloud operating systems from AWS, Google Cloud, and Microsoft Azure. SAP is integrating it with Joule Studio; WPP has demonstrated creative agents, while BBVA and Santander are reviewing it. The Register reports that OpenShell still lists open security issues on GitHub and collects telemetry by default unless users opt out.

Key facts
  • Vera Rubin runs models; Vera CPUs run OpenShell; BlueField-4 runs Sentry.
  • NVIDIA says Sentry can isolate a boundary-crossing agent within milliseconds.
  • NVIDIA says Sentry can raise agent capacity by up to 20 percent.
  • ServiceNow’s AI Passport can revoke agent access within seconds, NVIDIA says.
  • Telemetry is on by default, with an opt-out, The Register reports.
  • BBVA and Santander are reviewing the specification; SAP is integrating Joule Studio.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

What each outlet adds

beyond the shared facts
  • The RegisterDefault telemetry with opt-out, open GitHub security issues, and SAP Joule Studio.
  • Help Net SecurityMillisecond isolation, 20% capacity claim, ServiceNow revocation, and supported platforms.
  • NVIDIA BlogWPP demonstrated creative agents using Nemotron and OpenShell.
Vendors
NVIDIAGoogle CloudNvidia
Products
A5XVera Rubin NVL72ConnectX-9NemotronOpenShellA4X MaxG4OpenShell Reference ArchitectureBlueField-4NVIDIA SentryVera RubinSentry
Organizations
Ineffable IntelligenceWPPServiceNowBBVASantanderCanonicalSUSERed HatAmazon Web ServicesGoogle CloudMicrosoft AzureSAP
Countries
PolandUnited KingdomGermanyFranceGreece
AI models
NVIDIA NemotronNemotronPLLuMMistral Large 4Nemotron 3.5 Lightning

Coverage

oldest first · one row per article · UTC
  1. · 4d ago NVIDIA BlogFirst report

    NVIDIA adds Confidential Bare Metal for model builders

    Cohere plans to use it in Model Vault, and Google Cloud is extending it to A4X Max on GB300 NVL72.

  2. · 4d ago NVIDIA Blog

    NVIDIA partners expand European cloud capacity for sovereign AI

    Nscale says DSX MaxLPS ran 37% more GPUs and nearly 50% more inference throughput on GB300 NVL72 within the same power budget.

  3. · 4d ago NVIDIA Blog

    Google Cloud plans NVIDIA Vera Rubin A5X instances next month

    A5X pairs Vera Rubin NVL72 with ConnectX-9 networking, and Google plans capacity of up to 960,000 Rubin GPUs.

  4. · 4d ago The Register · Security

    Nvidia proposes OpenShell architecture to govern AI agents

    The blueprint needs Vera Rubin, Vera CPUs and BlueField-4, while BBVA and Santander are only reviewing it.

  5. · 4d ago Help Net Security

    NVIDIA unveils OpenShell architecture to bound AI agents

    Vera Rubin runs models, Vera CPUs run OpenShell, and BlueField-4 DPUs enforce policy with Sentry.

  6. · 3d ago NVIDIA Blog

    European public bodies adopt NVIDIA open models under local control

    Work covers the Commission, Poland, the UK, Germany and France, including imaging, personas and a classified cloud plan.