They derive the "1 in 1 million" chance of your vote altering the outcome by assuming the distribution of votes is given by a uniform distribution on the possible vote outcomes.
I have no idea where this assumption comes from. To me the more natural one is to assume a binomial distribution - everyone else has a certain chance P of voting for Obama. Then do the normal approximation, and you realize the odds of your vote mattering are vastly smaller (i.e., exp(- delta^2) rather than 1/delta).
I agree that the uniform distribution is fishy. I could object to the binomial assumption too; it seems kind of question-begging. My point was just that the thesis "your vote doesn't matter" is not at all self-evident.
That's nice sounding cynicism, but it's not obvious that the math actually comes out that way. Consider: http://www.overcomingbias.com/2008/01/yes-it-can-be-r.html