Technically we would say that the statistical multiplexing effects are greater.
The General principal is that the larger the sample the smaller the expected variability (in percentage terms).
In simple language lets take the example given. Suppose each user provides .075 e traffic then 100 users provide 75 erlangs while 2000 users provide 150 erlangs. If there were an infinite number of out going trunks we would say that on average 75 or 150 of them would be used. If we make the usual assumptions then the number of outgoing trunks in use is governed by a Poisson distribution.
We can determine that in
Case 1 (1000 users) : Prob more than 80 trunks used = 0.2589 Case 2 (2000 users) : Prob more than 160 trunks used = 0.2071
In both casse the percentage above the mean is the same but in the second the same degree of variability is less likely.
Hope this helps.
N.B. Blocking or sizing estimates based on Poisson distribution (or the normal approximation) are quite reasonable and give results close to those from ErlangB. Useful if Erlang-B tables not available. to it)