Except that the "sole purpose is to have teraflops of computing power" applies to neither quantum computing nor to machine learning in any meaningful way.

The upside of quantum computing is not that it can compute more, but that it needs to compute less for a very specific set of mathematical operations.

Machine learning is not enabling large amounts of computing power, it is enabled by large amounts of computing power.

There are indeed ciphers and cryptographic hashes which quantum computing might deprecate, but SHA-256 is not one of them for lack of a quantum algorithm that would offer an advantage over classical computation. Maybe we'll find one in the future, but for now that would be pure speculation and absolutely unrelated to anything machine learning.

Applying machine learning to essentially break cryptographic functions by understanding input-output correlations is a cute idea but pointless. If there's any correlation between input and output, than it's not a proper cryptographic function.