maximum HTTP connections

Hello all

Facing the problem of designing a cache system, a very
basic question arises:

Knowing the maximum requests/time_unit that the system will have,
how many machines does it take?

I think the main limitation for a squid/linux is the maximum
HTTP/TCP connections suported. So, is 100 permanent connections
a reasonable number? How about 1,000? Or 10,000 ?

Thanks,
Joao Pagaime

--
PS: level 4 switching with Alteon. Current traffic ~ 16Mbps. 
Machines ~ 512MB RAM, 3 x 9GB SCSI, on Linux or FreeBSD.
Tests already made indicate about 300 permanent HTTP connections.