NVIDIA's flagship edge AI platform ships without the instrumentation needed to measure how much energy it actually uses.
Researchers auditing the ASUS Ascent GX10, one of several GB10-based desktop AI systems now on sale from ASUS, Dell, HP, MSI, Acer, and Gigabyte, found the platform exposes exactly one energy metric: instantaneous GPU wattage via NVIDIA's NVML library. CPU energy counters, power-rail monitors, and standard management interfaces like IPMI and SCMI powercap are all absent from supported software interfaces. That matters because CPU-side processing accounts for up to 90.6% of total latency and 44% of total dynamic energy in agentic workloads, the multi-step, tool-calling tasks these boxes are designed to run. The kicker: the MediaTek firmware already computes per-rail energy internally via an undocumented ACPI interface, but NVIDIA says it has "no plans to expose CPU rail information."
If you can't measure it, you can't manage it, and the energy bill for agentic AI is steeper than most benchmarks suggest. The same research group previously found that orchestration overhead pushes energy consumption 4.33x higher per successful task than a linear baseline, hitting 7.63x for multi-step reasoning tasks. Deploying these systems at edge scale without per-process energy attribution means organizations are building AI infrastructure with no clear view of its power footprint.
The researchers did find one exception: the Acer Veriton GN100 has active CPU energy accumulators, which gave them a calibration reference for an interim workaround. They've also sketched a standards-track path via SCMI powercap. Whether the rest of the industry treats that as a roadmap or a curiosity will say a lot about how seriously the sustainable-AI pitch is meant to be taken.