NVIDIA OpenShell reference architecture adds a governance stack for safe agents
Agent governance is the change, and the OpenShell reference architecture assigns inference to Vera Rubin, harness work to Vera CPUs, and enforcement to BlueField-4.
| Vera Rubin NVL72 | Vera CPUs | BlueField-4 | |
|---|---|---|---|
| configuration | NVL72 | ||
| throughput per megawatt | 30x higher | ||
| performance vs x86 | 2x | ||
| agent capacity boost | up to 20% | ||
| stop time | milliseconds |
NVIDIA frames the stack as a boundary model for enterprise buyers. It must accept the model before comparing cost or capacity against another vendor.
The 30x throughput-per-megawatt figure for Vera Rubin NVL72 systems is the number that matters to the fab. It shifts the sales argument from raw tokens to containment overhead.
Absolute throughput, dollar cost per agent, rack power draw, and benchmark methodology remain unmeasured. The reference architecture is described as available, but no SKU, price, or warranty is named.
The reference architecture changes the sales argument from tokens to containment overhead, but the 30x claim lacks price, SKU, rack power draw, and benchmark method.
After nvidianews.nvidia.com. We did not report this. The pictures, if any, are theirs. Also filed by NVIDIA Newsroom.
