we keep hearing the bottleneck for AI is compute, that there aren't enough GPUs, that everyone's fighting for H100s and B200s. so I went and actually measured what one of ours was doing during a training job. it was idle most of the time. not slow. idle. doing a quick burst of work, then sitting there waiting for the next batch of data to arrive, over and over. the expensive part (the GPU) spent most of its life waiting on the cheap part (moving data to it). green is the GPU doing work, ora...