That's pretty much what I was thinking, reading this. First off, a frequency measurement -- "once in a billion years", 3.16887646 × 10^-17 hertz (thank you, Google calculator) can't really be used as a probability, as they are trying to do.

Then, even if you assign that nonsense statement some sort of charitable reading, like "one bit has a 50% chance of spontaneously changing state over a one billion year period", and infer that every bit has a 1/10^9 chance of flipping every year, then the odds that 8×10^9 bits do not change at all over the course of a year are (1-10^-9)^(8×10^9), or about 0.034%.

Granted, it'll still take a long time for that whole gigabyte of data to be completely scrambled, but I think we'll still be using error correction and occasional refresh, even with a technology this reliable.