The key lies in the reduction of message traffic with intra-ultrapeer qrp & dynamic querying.

At the moment most of the ultrapeers are simply overloaded and they have to drop some of the messages they receive. The result is, that a query that might have reached some 50,000 ultrapeers (= about 1,000,000 unique leafs) will not even reach 3,000 because it will be dropped before that. So you see, in theory, tcp searches could easily search more than three times as many hosts as there currently are on the network. Practically your queries will be dropped so often that you can only search a fraction of the existing network.

The explanation is simple:
* There are far too many requeries, - even LimeWire clustering can't shield against all of them.
* Pong traffic requires lots of bandwidth, although it does serve a purpose.
* Searches for popular content return easily more than a thousand results and many results are probably already dropped before they reach you in that case.
* LimeWire limits the bandwidth allocated for gnutella traffic very strictly.

The hope is, that LimeWire 3.0 will change it all and make gnutella as good as Fasttrack and I am confident that it will manage to improve the network significantly, so it cannot be compared to the old network anymore.