You are proposing that a reduced dimensional projection of a large dataset cannot approach this limit?

I.e. expose the underlying low rank of nearly any huge sparse data matrix with an SVD or NMF. Enable fast recovery with a shitty (CS-wise) hash function. Recover most of the information about an observation's neighbors in a fraction of the time taken by many other approaches.

What's popular for ML benchmarking these days is not necessarily the same as what's needed for a specific application. It's a useful proof to keep in mind before prematurely optimizing with overly complicated approaches.