generalization may or may not lead to interesting conclusions. for example using only low number of classes its possible to compress information in such tiny package we can deliver it quickly through network, and then do local classification search for large number of events.
consider:
using 10 classes = 10 bits uncompressed data per row
100 classes = 100 bits uncompressed data
calculating from current gzipped compression ratio which is reduction to about 2.3 % of original size, 100 bits uncompressed would result to 2.3 bits compressed on average. for 500M events thats 145 MB.