A 33% increase in GPU utilization resulted solely from changing the order of jobs in a compute cluster. Researchers at Hugging Face found that naive scheduling creates bottlenecks that leave hardware idle. This finding proves that software orchestration often limits performance more than raw silicon. Engineers should prioritize scheduling logic to maximize existing hardware efficiency.