Since i and even 10000+i are very small against 36^20=1.33674945388437e+31, even for i=10^6,
(36^20-10000-i)/(36^20-i)
=
(1-36^(-20)*(10000+i))/(1-36^(-20)*i)
is approximatively
1-36^(-20)*(10000+i-i)-36^(-40)*((10000+i)*i+i*i)
so the product for 10000 trials is
(1-36^(-20)*10000)^10000-(1-36^(-20)*10000)^9999*36^(-40)*10000^3
and the first term is about
exp(-36^(-20)*10^8) approx 1-7.48083342838978e-24,
the second is smaller 36^(-20)*10^8*(36^(-20)*10^4)
, and thus very small against the difference of 1 and the first term.
So the probability to hit an existing name is about 7.5*10^(-24)
.
For one million trials, nothing fundamental changes, only the power goes from 10^4 to 10^6 in
(1-36^(-20)*10000)^(10^6) approx 1-7.5*10^(-22)
so the probability to hit an existing filename is 100 times greater, but still essentially zero.
(enter Dr. Evil mode "one hundred billion ...")