He wants it to be one base pair per row, and also to express the multi-possibility rows in the source data. So he needs 4 bits per row for that, and he spreads them out rather than bunch up on one side.
That's not entirely true. There are also bits associated with ambiguous combinations of single base pairs, so there is actually 4 bits of information if all ambiguous combinations are possible. But it'd be much closer to 2 if you took into account how rare these are. Still, had he done exactly the same encoding without those column gaps between I wouldn't have criticized.