Comment by Twirrim

Even the large cloud AI services are focusing on this too, because it drives down the average "cost per query", or whatever you want to call it. For inference, arguably more even than training, the smaller and more efficient they can get it, the better their bottom line.