That quote left me with the same question. Something about decent amount of ram on one board perhaps? That’s advantageous for training but less so for inference?