“For many AI companies, GPU utilization for inference workloads is often under 15%.”