First work out how the different system counters look with different
numbers of virtual users, servers, and so on. Plot these on a graph,
then extrapolate like you would in simple mathematics.
http://en.wikipedia.org/wiki/Extrapolation
This is only useful for transaction response times and system
counters, it *cannot* show where you will reach hard limits within the
infrastructure. It is a useful technique when showing that a system is
unlikely to work, but it can't provide any assurance that it will
work! To extrapolate with any degree of accuracy you need to
understand system performance completely, like when you have
statistics from production and can compare them to a smaller test
environment.
So be very, very careful when doing this.