Enterprises deploying agentic AI on edge devices face a hidden cost: the inability to measure per-process CPU energy consumption on NVIDIA's latest GB10-based systems. A systematic energy-observability audit of the ASUS Ascent GX10 (GB10 SoC) reported in an arXiv paper by Panigrahy, Deepak, Tyagi, and Aakash found that the platform exposes no CPU energy counter, no INA power-rail monitor, no IPMI/BMC, and no SCMI powercap protocol through any supported software interface. The only on-device energy telemetry is instantaneous GPU power via NVML. This blind spot is critical as agentic AI workloads—multi-step orchestration with tool calls and retries—are being targeted for edge deployment, with NVIDIA, Dell, HP, ASUS, MSI, Acer, and Gigabyte all shipping GB10-based desktop AI systems in 2026.
The Energy Cost of Orchestration
The researchers previously demonstrated that orchestration structure dominates agentic energy cost: workflows consume 4.33x more energy per successful goal than linear baselines, and OOI reaches 7.63x for multi-step reasoning tasks. Separately, Raj et al. show that CPU-side processing accounts for up to 90.6% of total latency and 44% of total dynamic energy in agentic workloads. Without per-process CPU energy attribution, optimizing deployments for sustainability is impossible.
What the Audit Revealed
The audit uncovered a stark contrast between available and missing telemetry:
| Measurement Capability | Available on GB10? | Details |
|---|---|---|
| GPU instantaneous power | Yes | Via NVML |
| CPU energy counter | No | Not exposed |
| INA power-rail monitor | No | Not available |
| IPMI/BMC | No | Not present |
| SCMI powercap protocol | No | Not supported |
| Per-process energy attribution (like x86 RAPL) | No | Not reproducible through supported interfaces |
On x86 platforms, Running Average Power Limit (RAPL) enables per-process energy tracking; the GB10 lacks any equivalent.
The Hidden Data: MediaTek Firmware
A striking discovery: the MediaTek firmware already computes per-rail energy internally via an undocumented ACPI interface (SPBM). However, NVIDIA stated there are "no plans to expose CPU rail information," according to the paper. This means the hardware is capable but the vendor has locked out access.
Implications for Enterprise Edge AI
For CTOs and technology procurement leaders evaluating edge AI hardware for supply chain automation—such as warehouse robots or real-time logistics orchestration—the lack of CPU energy telemetry undermines efforts to meet carbon-reduction targets and optimize total cost of ownership. Without per-process energy data, organizations cannot attribute energy consumption to specific AI tasks, making it difficult to compare efficiency across workflows or hardware platforms.
A Path Forward
The researchers propose an interim calibration bridge for per-domain energy decomposition, confirmed functional on the Acer Veriton GN100 where CPU energy accumulators are live. They also identify a standards-track path via SCMI powercap. Their findings motivate the low-carbon computing community to demand energy observability as a first-class hardware requirement. For now, enterprises deploying GB10-based systems must rely on GPU-only metrics, leaving a significant blind spot in their energy management.
