“When performing inference tasks, GPUs are typically only 5% to 7% utilized, meaning 93% to 95% of their potential is wasted.”