It's quickly pointed out that he's not counting the training of models, producing all the GPUs, energy spent on scraping, the increased storage needs from scraping the whole internet, etc.
It is fair to compare inference to streaming. Both are done by the end user.