The graphics card shortages proves my point here.
Oh man, do you think if the popular coins start moving to proof of stake, used graphics cards will flood onto eBay and the like and we'll get to snap them up cheap for some machine learning?
Floating point math errors similar to the recent Titan V issues could potentially mess with your machine learning.
Regardless, if it has been mining for some time, the lifetime of the card is greatly shortened, if the operator did not take precautions such as running undervoltage and designing a good thermal solution. Most people will not have done these things!
Anecdotally, I've been running 12 RX480's to mine ethereum since AMD launched the cards some year(s) ago. They're still just as fast as the day I got them. What in your opinion would make the cards fail at this point?
If anything im inclined to argue that reduced thermal cycling will result in these GPU's lasting longer than cards that get torched to 100% to game for an hour every day and then sit at idle the other 23 hours.