NVIDIA on Wednesday, Oct. 21, 2026, at GTC Berlin announced the NVIDIA OpenShell Reference Architecture, a full-stack design for building, containing, observing and enforcing safe AI agents on enterprise infrastructure and at the edge. The company said the architecture is available now.
The design sits on NVIDIA OpenShell secure runtime software and pairs three hardware layers the company named in the release: NVIDIA Vera Rubin infrastructure for model inference (including closed and open models built with NVIDIA Nemotron), NVIDIA Vera CPUs for agent harnesses and tool use, and NVIDIA BlueField-4 DPUs running NVIDIA Sentry for continuous out-of-band monitoring and in-silicon policy enforcement. Edge coverage extends to NVIDIA RTX Spark, DGX Spark and DGX Station systems.
WebProNews covered the earlier open-source OpenShell runtime and Sentry hardware watchdog on Sept. 29, 2026. The new reference architecture extends that OpenShell runtime into full-stack infrastructure for agent development and operations, according to NVIDIA.
What the stack claims to do
In the release, NVIDIA said Vera Rubin NVL72 systems deliver 30x higher throughput per megawatt for agentic workloads, Vera CPUs deliver 2x the performance of x86 processors with greater energy efficiency while running OpenShell, and BlueField with Sentry can quarantine an agent that moves outside its software boundary in milliseconds while boosting agent capacity by up to 20%. Those figures come from NVIDIA; independent benchmarks were not cited in the announcement.
“AI capability and safety must advance together,” said Jensen Huang, founder and CEO of NVIDIA. “As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture. NVIDIA created OpenShell to give agents the freedom to do useful work within clearly defined boundaries — contained, monitored and under control. With the OpenShell Reference Architecture, agents can think on Rubin, act on Vera and be governed on BlueField.”
Huang made the same case in his GTC Berlin keynote, according to live notes on the NVIDIA Blog, where he described an agent as a model plus harness, sandbox, knowledge and tools, and said an agent’s privileges should be minimal and monitored. “Every single time an agent is launched, the privileges are different,” Huang said.
Who is wiring into it
NVIDIA said the architecture runs on Linux and Kubernetes stacks from Canonical, SUSE and Red Hat, and on Microsoft Windows using Microsoft Execution Containers — the MXC path WebProNews previously covered when Microsoft put policy enforcement around Windows agents. Cloud operating systems listed include Amazon Web Services, Google Cloud and Microsoft Azure.
NVIDIA said enterprises can run agents on OpenShell infrastructure from model builders including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence. Cybersecurity vendors listed as integrating support include 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler. System makers Dell Technologies, Cisco, HPE, Lenovo and Supermicro are among those building infrastructure to the design, and cloud providers named include CoreWeave, Crusoe, Lambda, Nebius, Nscale, Oracle Cloud Infrastructure and Vultr.
On the application side, NVIDIA said SAP is integrating the architecture with Joule Studio runtime on the SAP Business AI Platform, and that Palantir is working with NVIDIA to deploy secure agents in production. Banks BBVA and Santander were named as exploring the architecture for agent safety and governance. NVIDIA’s release also said 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.
The OpenShell Reference Architecture is available now, according to the Oct. 21 release.
from WebProNews https://ift.tt/yHAtGS3
No comments:
Post a Comment