“Aidan Gomez predicts that large-scale AI model training, which currently uses tens to hundreds of thousands of GPUs, could scale to millions of GPUs within a couple of years.”