A GPU runs hot while utilization falls and jobs slow down. What should be investigated?

1. Only the inference API authentication policy
2. Thermal limits and the cooling path
3. Only the cluster namespace quota names
4. Only model prediction accuracy