It's a win-win-win for Amazon if they destroy the source in the process of scanning it. They get the AI training data, and it's cheaper for them to destroy the book in the process. Destroying the book ensures that it will be more difficult for competitors to scan the same content, thereby increasing the value of the data they've scanned.

Lastly, there's a misguided belief among some that it's somehow less of a copyright violation if the source is destroyed, but it's a violation either way unless the entity doing the scanning has permission from the copyright holder to make the copy.