The idea itself is interesting anyway. Even if there are more keys than 2^96. Total addresses can be (base58) 33-34 2^199 904798310844700775313327215140493940623303545877497699631104. Moreover, for 1 digital address, 2 characters each.
It turns out they can fit entirely into spaces.
904798310844700775313327215140493940623303545877497699631104 (base58) 33-34 2^199
1606938044258990275541962092341162602522202993782792835301376 2^200
3213876088517980551083924184682325205044405987565585670602752 2^201
...
1645504557321206042154969182557350504982735865633579863348609024 2^210
1766847064778384329583297500742918515827483896875618958121606201292619776 2^240
1809251394333065553493296640760748560207343510400633813116524750123642650624 2^250
If we generate all addresses from 2^200 and 2^201, that is addresses must be repeated anyway.
The question is how much will be in the lower ranges 2^1-2^160 (and even with the balance).
Well, if you factor in hash160, then all addresses will be in the lower ranges of 2^1 to 2^160. Moreover, all addresses should be found in the 160 bit range. If you factor in compressed and uncompressed, maybe you can cut it down to 2^159 range (2^159 * 2 (for comp and uncomp keys) = 2^160) who knows really; waiting for that first collision of sorts.