In some applications, e.g. modelling something defined by PDE, you might need to solve a very large linear system of equations Ax=b, where the vector b has 100,000,000 elements, and A is a 100,000,000 by 100,000,000 element matrix. So for a naive dense matrix representation of A using 1 byte per entry you'd need around 9 petabytes of storage. In practise, for many interesting systems of equations that encode "local"…
And what happens if indeed the huge matrix elements are all non-zero? Like, let's say, satellite scan data for a given country when you want to spy on their underground systems (think North Korea facilities)? Wouldn't storing that data as COO would actually triple the amount of memory?
Presumable the designer(s) of such a system will be able to know in advance whether the resulting matrix will be sparse or not and choose their encoding appropriately. FWIW, for a lot of practical applications the raw sensor data would be non-sparse but it would be transformed/filtered almost immediately into a more space-efficient representation. For example in the case you mentioned (unless you think the entire country is completely underdug by tunnels) most of the raw data can be deleted immediately with no loss of signal as there is no underground system underneath that particular spot of land.