until its cheaper to train and infer than 100k gpu data centers...i doubt it will ever compete.