Exactly, always best to rely on your hardware, we need to collect/add more data from self hosted models on different gpus/clouds to compare