How does Windows monitors/calculates GPU utilization?

I’ve been coding in CUDA recently, and I launched 2^31 kernels onto my GPU to test out. The program took around 63 seconds to complete, but my GPU utilization (seen in Windows Task Manager) didn’t get over 10%, all utilization coming from Discord. Why is that so? My best guess is that it does not display (active CUDA Cores / number of CUDA Cores), but the PCIe bandwith utilized (my program does almost no RAM exchange).