True, but in order to do that you have to know what the new random variations are. Output from existing sequencers, as well as the current crop of algorithms, a

True, but in order to do that you have to know what the new random variations are. Output from existing sequencers, as well as the current crop of algorithms, are not good enough to do that with high confidence. Thus, people tend to keep everything. You can get some gain by doing reference-based compression, but you're still looking at on the order of 100s of gb per sample.