Scot Reader gave it a shot using tobacco production numbers and came up with up to 370 million total produced. (on page 26:
Inside T206)
Let's say one rounded down to 300 million because of other subjects printed, then only half a percent surviving (my number) puts the total at 1.5 million, then you divide by 500 – this is all so BotEC in nature, anyways – to get an average of 3,000 for each. (Again, Scot has a better discussion.)
I'd love to develop a model that estimates propensity to be graded. There must be a somewhat similar distribution of grades for each across all cards (hopefully, but we could account for people taking special care of their Mathewsons over the decades). Using cards of varying desirability (via price), if we could compare the difference of the right of the distribution to the differences in the center and left of the distribution, we might be able to get an idea of what remains. (Refine it from there to account for regrading.)