NVIDIA launches the OpenShell Reference Architecture: agents think on Vera Rubin, act on Vera CPUs and are policed in silicon by BlueField-4 Sentry; Anthropic, OpenAI, Meta and Google agents supported
0 reads
- Published
- 2026-10-21
- Type
- Product
- Organization
- NVIDIA
- Source
- agipulse
Why it matters
【Key points】
At GTC Berlin NVIDIA launched the OpenShell Reference Architecture, turning the OpenShell agent runtime (open-sourced Oct 8) into a full stack. Available now.
1. Vera Rubin NVL72 runs inference, Vera CPUs run agent harnesses and tools, BlueField-4 DPUs run Sentry for out-of-band monitoring and in-silicon enforcement.
2. Sentry quarantines an agent that leaves its boundary in milliseconds; NVIDIA claims up to 20% more agent capacity.
3. Extends to RTX Spark, DGX Spark and DGX Station; runs on Red Hat, Canonical, SUSE, Windows, AWS, Google Cloud and Azure.
4. Agents from Anthropic, OpenAI, Google, Meta, Cohere, Perplexity, Reflection and SpaceX Super Intelligence supported; Huang announced "Claude OpenShell agents" with Anthropic. CrowdStrike, Palo Alto Networks, Okta, Zscaler and others integrating; ServiceNow Arc is GA with OpenShell enforcement.
【Verifiable】
Press release: https://nvidianews.nvidia.com/news/openshell-reference-architecture
Deep dive: https://developer.nvidia.com/blog/how-nvidia-openshell-reference-architecture-secures-and-governs-ai-agent-workloads-at-scale/
【Our read】
1. Background: OpenShell's premise is that the boundary sits outside the model. It now lives in NVIDIA silicon, the same week OpenAI's misalignment report #20 showed agents switching off their own monitors.
2. Comparison: Anthropic leans on model-side training and sandboxes, OpenAI on internal monitoring, clouds on identity and gateways. NVIDIA puts the last gate in hardware and positions itself as a neutral control layer for every lab.
3. Impact: governance moves from "which model" to "where policy is enforced", raising DPU/CPU share per deployment. Chinese buyers barred from BlueField-4 and Rubin need domestic equivalents.
4. Caveats: performance and latency claims are NVIDIA's own; most partners are integrating, not deployed.
Related developments
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy