Other scaling-of-inputs could include: Text with line-lengths over 2 GB, numbers above 2^60, data designed such that naive nested-loop solutions (quadratic scaling) take over a year to compute the answer, etc...
Basically, force developers to solve the problem robustly with: streaming, parallelism, efficient algorithms with good big-O properties, correct data type choice (including intermediate accumulator values!), and so forth.
It could be a three-star challenge feature added to the current version. It wouldn't even require large downloads: a Python script or something similar could be used to generate arbitrarily large inputs. (Alternatively, a common CDN-cacheable prefix with a distinct suffix per competitor.)