GPU power draw doesn't scale linearly with utilisation. It ramps up steeply from 0–30% utilisation, then more gradually from 30–100% — so a GPU sitting at low utilisation still burns a large share of its peak power for a small share of its peak output. That "idle floor" shows up three ways here:
Power tab shows the actual watts wasted just from GPUs being powered on, doing no useful work. Cost tab converts that into an annual electricity dollar figure. Scenarios tab reframes it as a hardware question — how many physical GPUs you'd actually need, at your current utilisation, to match the throughput fewer GPUs would deliver at a higher target utilisation. That gap is capital sitting idle, not just an electricity line item.
The chassis breakdown goes one level deeper: GPU TDP is only part of total system power. CPUs, NVSwitch fabric, memory, networking, storage, cooling, and PSU conversion losses add further overhead that's largely fixed per chassis — so under-utilising GPUs means under-utilising everything built around them too.
GPU TDP is only part of total system power. Select a chassis to see how CPUs, NVSwitch fabric, memory, networking, and PSU losses stack up — and how the total fluctuates with GPU utilisation.