Agentic AI: A Power-Hungry Path to the Edge

In 2026, the AI landscape is buzzing with agentic workloads heading to edge setups. Heavyweights like NVIDIA, Dell, HP, ASUS, MSI, Acer, and Gigabyte are all rolling out GB10-based desktop AI systems. But there's a snag. These agentic AI workloads aren't just about smart orchestration and reliable execution. They're energy hogs, consuming 4.33 times more energy per goal than their linear counterparts. Multi-step reasoning tasks stretch that figure to a staggering 7.63 times.

CPU Bottlenecks and Energy Drains

Rajat et al. have highlighted a glaring issue: CPU-side processing is the bottleneck, grabbing up to 90.6% of total latency and 44% of dynamic energy in these workloads. If we're pushing AI to the edge, why are we still stumbling over the CPU's inefficiencies? Where's the innovation in breaking this bottleneck?

ASUS Ascent GX10: A Missed Opportunity

Taking a deep dive into the ASUS Ascent GX10, packed with a GB10 SoC, reveals a stark reality. The platform's energy observability is practically nonexistent. No CPU energy counters, no power-rail monitors, and forget about IPMI/BMC or SCMI powercap protocols. What you do get is a limited view through NVML for instantaneous GPU power. It's like driving blindfolded.

The bigger bombshell comes from MediaTek. Their firmware already computes per-rail energy internally via a mysterious ACPI interface, SPBM. Yet, NVIDIA's stance is that CPU rail information won't be exposed. If the AI can hold a wallet, who's going to write the risk model?

Demanding Accountability

What's needed is a formalized hardware requirements spec for energy-attributed AI. Right now, ASUS's setup can't reproduce per-process energy attribution like x86 platforms can with RAPL. Our path forward? An interim solution through external DC metering and GPU subtraction. This isn't just a technical challenge, it's an industry-wide call to action. Energy observability must become a non-negotiable hardware requirement.

The intersection is real. Ninety percent of the projects aren't. Without verifiable energy telemetry, how can we claim to be building low-carbon, efficient AI systems? Slapping a model on a GPU rental isn't a convergence thesis. It's time the low-carbon computing community demands transparency and accountability.