You cannot equate full expenses per year here with observations uploaded in that year, because the substantial computational running expenses of 45%+ish comes from both AWS photos grant + Google maps grants + iNat own servers costs wherever they are + other legal and such donated services - so that makes some expenses spread across full accumulated observations with storage, processing and handling services on platform from start instead of just per year and scaled with that observation usage over time.
and the staff salary value shouldnt be justified only for current year observations as the software is about capital value created in this year and is assets whose value extends over multiple years.
so yes, its tricky to separate strictly but if we make amortized assumptions, maybe we can assign something like c x N_t + m x S_t + B_t + AC_t = approx. economic costs per year t.
where c = marginal cost in per observation generation, N_t = number of observations in that year t, m = average annual maintenance cost per stored observation, S_t = all observations stored from iNaturalist start, B_t = staff benefits like travel, pension, insurance,… and any other overhead costs, AC_t = amortised per year staff costs considering salaries and capital value generation
AC_t is tricky to understand but conceptually something like sum (a x staff_t-i/L) over i = 0 to L - 1; basically we take staff salary in year t and spread it across L years based on capital generation factor a and how many years we think is the lifetime for those software and infrastructure capital - maybe 8 years average for iNat.
Now, your question of per observation cost would be technically and marginally low with these parameters and average per observation cost per year expense wont be causal.
we currently dont have numbers for these, well a picture that is deemed unidentifiable in X year can be identifiable in X+b years from other community efforts, field sampling and new literature, new clues found from other iNat photos of relevant things and exclusion choices, … although I agree there will always be something that is too low quality but we cant find all that set manually (DQA flags are rarely used and used even when they are not strictly un-indentifiable always and per individual identifier views) and any computational model that scores this observation quality can also be faulty for lot of reasons in taxonomy. I would say truly almost zero-value observations have yearly costs in few thousands of dollars as of now but since it is not measured nor platform sampled statistically, it is ignored for active measure actions such as better onboaring?
Pictures of friends can be fun expendable cost that my guess could be less than 1000$ and is overkill to create new restrictions, on duplications = computationally de-duplicable someday if there comes intent (here is relevant feature request on forum) maybe only when iNat running costs become too much and those photo grant providers pressure such actions,