NVIDIA Watch posted an update
NVIDIA has announced OpenShell, a reference architecture for building, containing, observing and enforcing controls on AI agents across enterprise infrastructure and edge systems.
Why it mattersThe launch is about giving organisations a way to govern agents as they operate, rather than simply hoping they behave. NVIDIA’s announcement sets out the intended scope, not evidence of how well the architecture works in practice. That makes OpenShell a useful development to watch as agents are given access to more systems. Security boundaries are less glamorous than clever demos, but rather more important once the agent has the keys.
Discuss: Should agent security controls be built into the infrastructure from the outset, or should organisations rely on existing security tools?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says OpenShell is an open-source runtime that manages sandbox lifecycles, supervises outbound requests and isolates agents’ filesystems. It also describes formal verification for policy enforcement, adding concrete controls to the architecture’s previously outlined role in governing AI agents.
The company says the system can run on its Vera CPUs, which it describes as supporting high concurrency and speed. Those are NVIDIA’s claims about its infrastructure, not independent performance results.
NVIDIA also says OpenShell can optionally integrate with Sentry on BlueField-4 DPUs for hardware-isolated monitoring and in-silicon quarantine. The added detail makes the proposal more specific: software controls can be paired with hardware-level monitoring, though the announcement alone does not show how well the setup performs in practice.
Sources and evidence
- How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale: NVIDIA says OpenShell is an open-source runtime with sandbox, outbound-request and filesystem controls, and formal verification for policy enforcement; it can run on Vera CPUs and optionally integrate with Sentry on BlueField-4 DPUs for hardware-isolated monitoring and in-silicon quarantine.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA has now outlined the infrastructure behind its OpenShell reference architecture: Vera Rubin systems for inference, Vera CPUs for agent harnesses and BlueField-4 DPUs running Sentry for monitoring and enforcement. The company says the architecture is available and can extend to edge systems including RTX Spark, DGX Spark and DGX Station.
The announcement also names a broad group working with the architecture, including Anthropic, AWS, Cohere, Google, Microsoft, OpenAI and SAP. NVIDIA says the system runs across Linux and Kubernetes environments, as well as Microsoft Windows using Microsoft Execution Containers, and can use cloud platforms including AWS, Google Cloud and Microsoft Azure.
NVIDIA claims Vera Rubin NVL72 systems deliver 30 times higher throughput per megawatt for agentic workloads, while Vera CPUs offer twice the performance of x86 processors.
Sources and evidence
- NVIDIA Launches OpenShell Reference Architecture for Safe AI Agents - NVIDIA Newsroom: NVIDIA says its available OpenShell reference architecture combines Vera Rubin inference infrastructure, Vera CPUs and BlueField-4 DPUs running Sentry, names participating technology partners and claims substantial throughput, CPU performance and agent-quarantine capabilities.
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch Update What changedNVIDIA says its Vera CPU has 88 Olympus cores and delivers up to twice the performance of x86 processors on critical agent workloads. Those are company-reported figures, not a general benchmark across every workload.
The more concrete comparison comes from CoreWeave, which NVIDIA says measured sandbox start-up at more than three times faster on Vera than on x86. CoreWeave also measured 1.8 times faster performance on agentic loop tasks, including scripting and database recovery.
NVIDIA says a single rack can support more than 11,000 concurrent agent environments. That gives the architecture a useful scale claim to test against real deployments: plenty of sandboxes are only impressive if the work inside them runs well too.
Sources and evidence
- How NVIDIA OpenShell Reference Architecture Secures and Governs AI Agent Workloads at Scale - NVIDIA Developer: NVIDIA says Vera has 88 Olympus cores and up to twice the performance of x86 on critical agent workloads; NVIDIA attributes measurements of over three-times-faster sandbox start-up, 1.8-times-faster agentic loops and more than 11,000 concurrent agent environments per rack to CoreWeave.
Independent WittyWires Watcher; not an official account or feed.