Indeed, abundance is irrelevant. "Coverage" is key. Coverage is the number of times that a part of the sequence is represented. Having more sequenced DNA increases the coverage, but a more important parameter is diversity. Whether it is inter- or intra-species diversity, having different sequences means you are less likely to run into that sequence again.
The most interesting metagenomic projects have been the low-diversity ones, where coverage is high enough to recreate the microorganisms there. In Banfield's pioneering work on the microbes composing biofilms growing on the acidic drainage from abandoned mines (aka "acid mine drainage"), near-complete genome sequences were obtained for the two most abundant constituents (of six total, I believe). Contrast this with Venter's simultaneously-published Sargasso Sea metagenome (sequencing the microbes caught in filters from deep sea water, erroneously purported to have low-diversity due to low mixing of waters): most sequences were never encountered a second time, after a ridiculous amount of sequencing.
To do a good metagenome study, you have to pick the environment cleverly. Getting the genome sequence of enrichments of unculturable microbes that are environmentally relevant (eg Annamox, Mark Strous), or the termite hindgut (Hugenholtz/Leadbetter).
As sequencing costs go down and throughput goes up, metagenomics is going to become more prevalent and inexpensive. LARGE LARGE LARGE datasets will be generated. While this scale of data has never posed a problem for you Slashdotting types, it becomes a matter of "what is the scientific question?" What are you looking for? Interesting things will be turned up by metagenomics, but few will ever thoroughly mine their metagenomic data for interesting information. On the applied realm, this may actually be a moot point. "Functional metagenomics" is already a normal strategy for drug discovery (cloning random bits of environmental DNA into a model organism and performing a clever screen)
