AI training tasks have relatively high temporal flexibility (a delay of several hours typically does not affect the final result), whereas inference tasks are extremely sensitive to latency.