In fairness, it's not space for zeros which is allocated, but data, i, p (each 10% of total), and AVL tree. E.g. 40K x 40K matrix of type double would require ~12.8 GB of dense storage, but ~7.5 GB (checked "Commit size") of sparse COO storage. I agree it's unexpectedly very much. If I change density to 20%, then value goes up to ~15 GB. Already more than for dense matrix. Well, for my initial purposes of "very large 64-bit indexes" actual population would be very low. Sorry to continue after discussion seemingly was supposed to end :)