TodayToday's edition
NVIDIA's OpenShell Reference Architecture runs agents on Vera CPUs and enforces policy on BlueField-4 DPUs, with a claimed 20% capacity gain that no benchmark yet backs.
2 min readAI-assisted
NVIDIA's release illustration for the OpenShell Reference Architecture, announced at GTC Berlin. It is key art, not a photo of deployed hardware.
Image: NVIDIA
Source · Press image from the NVIDIA newsroom, used for news reporting.
NVIDIA on Oct. 21 introduced the NVIDIA OpenShell Reference Architecture at GTC Berlin, a three-layer design that runs AI agents on Vera CPUs and polices them from BlueField-4 data processing units (DPUs), and said it is available now.
According to Converge Digest, Vera Rubin infrastructure handles model inference, Vera CPUs run the agent harnesses and tool calls, and BlueField-4 DPUs run NVIDIA Sentry, a security framework built on the company's DOCA software. The point of the layout is that enforcement sits outside the agent's own software environment, so a misbehaving agent cannot switch off the thing watching it. NVIDIA says Sentry can quarantine and stop an agent within milliseconds if it leaves its authorized boundary.
BlueField-4 DPUs run NVIDIA Sentry for out-of-band monitoring and enforcement, separate from the Vera CPUs that run the agent harness, according to NVIDIA and Converge Digest.
Illustration (AI-generated)
NVIDIA says the BlueField-4 and Sentry combination can raise agent capacity by up to 20%. The announcement did not give benchmark conditions for that figure, and it is a vendor claim. The company also repeated earlier performance numbers for the other two layers: up to 30x higher throughput per megawatt for Vera Rubin NVL72 on agentic workloads, and twice the performance of x86 processors for Vera CPUs, NVIDIA said. Jensen Huang summed up the split in the release: agents think on Rubin, act on Vera and are governed on BlueField.
The architecture gives the DPU a standing job in every agent rack. Until now Voltgrid's coverage of Vera has centered on CPU throughput, as in Modal's agent tests, and on how many Vera chips fit in a rack, as in Dell's 256-CPU 48U design. OpenShell adds a third silicon role to that picture. Putting policy enforcement on a DPU keeps it separate from the host CPUs that run the agents, and NVIDIA ties its capacity claim to that arrangement. BlueField-4 is already used in NVIDIA's storage boards, so operators buying Vera Rubin racks for agent work would be buying more of the same DPU.
OpenShell itself is an open-source secure runtime, Network World noted, and the reference architecture extends it into a full-stack system. NVIDIA said it runs on Linux and Kubernetes from Canonical, SUSE and Red Hat, and on Windows using Microsoft Execution Containers. Converge Digest listed Dell, Cisco, HPE, Lenovo and Supermicro among the system makers building infrastructure around it, and CoreWeave, Crusoe, IREN, Lambda, Nebius and Oracle Cloud Infrastructure among the cloud operators working on OpenShell-ready environments.
The reference architecture is available now, but Converge Digest noted that individual hardware capabilities and deployments have separate schedules. NVIDIA gave no adoption figures, customer deployment counts or independently validated security benchmarks. The millisecond quarantine time and the 20% capacity gain both rest on NVIDIA's own statements. NVIDIA's own release also cautions that many of the products and features it describes are still at various stages, with timing subject to change.
