The end result to all that wasteful training is that it doesn't work or it gets tricked or confused by junk input. Proving to be a waste of energy, time, money and the re-training, fine-tuning process has to be done repeatedly if there is an error, or everyday on new data.
Either way, it is pointless to outline a comment as 'This is a false dilemma' without countering my claims as I have both recognised the wastefulness of decentralized blockchains like Bitcoin, Bitcoin Cash, Doge, that are much worse for their intended purpose (payments) etc and using tons of data centers for Deep Learning / Machine Learning training, re-training, fine-tuning, etc everyday on user data which benefits surveillance capitalism, spyware uses etc which that is also worse for everyone.