That makes sense. I'm assuming that the majority of GPUs are not running scientific workloads, instead deep learning and matrix ops for finance.