(Achtung: I'm mostly thinking aloud here and haven't extensively explored these ideas. Just something I thought I'd throw out there.)

Stretches of DNA can get themselves replicated in one of two ways: They can help an organismal vehicle reproduce more effectively, or they can hijack an organism's transcriptional and replicative machinery to get themselves copied. In the latter case, if they travel autonomously they're called viruses and if they quietly hitch a ride in the DNA of some host organism they're called selfish DNA. Replicating inert pieces of selfish DNA takes up resources, but the cost of the marginal copy is low enough that quite a few can accumulate -- over 40% of the human genome, for instance, is estimated to be made up of retrotransposons.

But as long as the cost is non-zero, there'll still be a kind of fitness optimum for any bit of selfish DNA -- one that was too adept at getting copies of itself spliced into the genome would eventually get to the point where the host organism would start taking a fitness hit, and one that wasn't adept enough would get outcompeted by ones that did a better job.

How can selfish DNA elements be said to compete, if they're neutral by assumption and hence will just be blown around randomly by drift? While it's true that at any single locus the dynamics of selfish DNA elements will follow straightforward drift dynamics, retrotranspositional elements are special because identical copies of them can exist at many loci. To quantify this, at any single locus a string of selfish DNA's odds of sweeping are 1/N, and that goes for all its competitors too. But its expected score in the multi-locus game will be k*(1/N), where k is the number of loci where copies of it are present. It now becomes a game of "whoever has the most loci wins": if you're a retrotransposon with 500 identical copies of yourself on a genome, you've got an absolute advantage in the replication game over a different one that only has 100 copies of itself.

What retrotransposons are competing for in this case is the total excess carrying capacity of a population's genome -- i.e. the finite amount of room to expand your numbers before your hosts start getting hurt -- and this will vary from species to species. Note that all other things equal, a bit of selfish DNA that gets in early in the game will tend to profit more than a similar one that comes in late. But also note that all other things may not be equal: if population size varies over the course of the game, then the expected payoff for a new piece of selfish DNA that enters the game will be greater when N is lower, since its odds of sweeping to fixation are better.

But what happens when the carrying capacity has been filled? Since the dynamics at all the loci will be governed by mutation and drift, the equilibrium in this game is that nobody really wins because in the long run very few (none?) of the loci will be identical by state. [1] But in the long run we are all dead, and mutation+drift acts slowly enough that it would take a very long time to reach this equilibrium. A string of selfish DNA that controlled many loci would "outlast" one that only controlled a few, since even if another sweep occurred by chance at one of its loci it'd still be dominating a whole lot more.

In the mean time, the circumstances could change: Mutations at other loci or changes in the organism's environment could increase the carrying capacity of the genome (creating more room for expansion), or conversely the organism could evolve some mechanism for recognizing and snipping out (or suppressing replication of) the most common lineages of selfish DNA, clearing out room for an uncommon variant to take up the newly-freed space. [2] In this latter case we could hypothetically see a discontinuous sort of negative frequency-dependent selection, with variants that were "too successful" getting periodically wiped out.

I just thought this up as a way of demonstrating how a form of natural selection among replicators can occur even conditions where the loci involved are all effectively "neutral" from the organismal point of view. I have no idea whether something this has actually happened, but it seems quite possible in principle, unless I've gotten something terribly wrong somewhere.

[1] There is a way to escape this endgame, however: A bit of selfish DNA could mutate into something useful for its organismal vehicle, thereby biasing its own odds of being replicated. But in that case it's nolonger playing the drift game and can be modeled by the standard single-locus selective equations that we all know and love.

[2] Remember that the assumption here is that the marginal locus is neutral; but if the same string of selfish DNA was present in a large amount of loci, the total fitness hit to the organism could in principle make it worthwhile to evolve some mechanisms to keep the parasitic DNA under control.