NVIDIA says its OpenShell Reference Architecture, announced on 21 October 2026 at GTC Berlin, is now available. It is a blueprint for building, containing, observing and enforcing safe AI agents on enterprise infrastructure and edge systems. Its headline numbers, 30x and 2x, are NVIDIA’s own claims.
The short answer
- What it pairs: Vera Rubin inference, Vera CPUs that run the agents, and BlueField-4 DPUs that run a monitoring framework called Sentry.
- Availability: NVIDIA says the reference architecture is now available, with documentation linked on its site.
- Evidence: the performance and security figures come from a press release. No independent benchmarks were found.
- Partners: most partner work is described as in progress.
Which layers does NVIDIA describe?
| Layer | Role according to NVIDIA |
|---|---|
| Vera Rubin (NVL72) | Model inference, with closed and open models built with Nemotron |
| Vera CPUs | Run agent harnesses and run OpenShell, the secure runtime boundary for autonomous agents |
| BlueField-4 DPUs | Run NVIDIA Sentry, a unified security framework for out-of-band monitoring of agent behavior and in-silicon policy enforcement, built on NVIDIA DOCA |
| NVIDIA OpenShell | Open source secure runtime software |
| Edge | RTX Spark, DGX Spark and DGX Station systems for retail, healthcare, manufacturing, robotics, software development and telecommunications |
The idea behind the layout: the agent runs inside a software boundary, and a separate chip watches it from outside. If the agent’s own software is compromised, the watcher is still out of its reach.
What does NVIDIA claim about performance and safety?
- Vera Rubin NVL72 delivers 30x higher throughput per megawatt for agentic workloads.
- Vera CPUs deliver 2x the performance of x86 processors, with greater energy efficiency.
- If an agent tries to move outside its software boundary, Sentry quarantines and stops it “in milliseconds”.
- Sentry running on BlueField-4 boosts agent capacity by up to 20%.
These are the company’s own figures in a press release, and no independent benchmarks were found.
Who is involved?
The release names these groups of companies:
- Model builders: including Anthropic, Cohere, Google, Meta, OpenAI and Perplexity.
- Cloud and operating-system vendors: including AWS, Google Cloud, Microsoft Azure, Canonical, Red Hat and SUSE.
- Security companies: including 1Password, Check Point, CrowdStrike, Cisco, Okta, Palo Alto Networks and Zscaler.
- System makers and cloud providers: Dell, HPE, Lenovo and Supermicro; CoreWeave, Nebius and Oracle Cloud Infrastructure.
The release describes most partner work as in progress. Partners “are integrating,” “are working with” or “are building,” and the banks BBVA and Santander are “exploring” it. ServiceNow announced general availability of ServiceNow Arc, a personal long-running agent governed by OpenShell.
Is it available now?
The release says the reference architecture is now available. Its disclaimer says many products and features remain in various stages and will be offered on a when-and-if-available basis, without naming which ones. Read the documentation to see which parts you can deploy.
Who should act, and what should you do next?
Teams planning to run autonomous agents on their own infrastructure should read this first. Teams that only call hosted AI services can skip it.
- Read NVIDIA’s documentation to see what you can deploy today.
- Treat the 30x, 2x and 20% figures as company claims until independent tests appear.
- We suggest asking each partner whether its work is available or still in progress.
Questions and answers
Is OpenShell a product I can install today?
NVIDIA describes OpenShell as open source secure runtime software and says the reference architecture is available. Its documentation shows what you can deploy.
What does Sentry do?
Per NVIDIA, it runs on BlueField-4 DPUs, monitors agent behavior out of band and enforces policy in silicon. It quarantines out-of-bounds agents.
Are the 30x and 2x figures independent?
No. They come from NVIDIA’s own press release.
Which companies are building on it?
Model builders, cloud vendors and security firms, among them Anthropic, OpenAI, AWS, CrowdStrike and Palo Alto Networks. Most of that work is described as in progress.
Terms in plain language
- Agent: an AI system that takes steps and uses tools on its own.
- DPU: a chip that handles networking and security tasks so the main processor does not have to.
- Out-of-band monitoring: watching a system from outside it, so the monitored software cannot easily switch the watcher off.
- Runtime: the software environment in which a program runs.
- Quarantine: isolate something so it cannot affect the rest of the system.
Sources
Facts read on 22 October 2026.
