To provide a much tighter bound, you reach "atoms in the universe" levels of collision once your hash inputs are 260-270 bits bigger than the hash output. No need for even a single kilobyte.
Almost a "Graham's number" level of understatement.
Imagine something like finding the seed for a pseudorandom number generator where the numbers generated form the byte pattern for libgen.
That's similar to the cryptominer hashing where people just have rigs enumerating through possible values to generate a hash with the appropriate number of leading zeros. It's just that libgen would be little bit longer than a hash and also have some ones interspersed with the zeros.
There is no way of traversing hash collisions. It would be the hash of some string of cryptocurrency jarble that also happens to be the hash of some encoding of the libgen database and you would have no way of knowing this unless you specifically hashed all of libgen. In this case you know all of libgen already anyway, unless you forgot the information that you hashed.
That might be an easier solution than trying to achieve minimally sane copyright reform.