A novel algorithm, the compressed block decomposition, is presented for highly accelerated direct (noniterative) method-of-moment capacitance extraction. The algorithm is based on a blockwise subdivision of the method-of-moment potential coefficient matrix. Matrix subblocks corresponding to distant subregions of the problem geometry are not calculated directly but approximated in a compressed form. Subsequently, the matrix is decomposed using an algorithm that preserves the compression. The efficiency of the method is demonstrated on a common benchmark problem - a 6 times 6 bus crossing. The numerical cost of the algorithm is shown, both theoretically and numerically, to scale with N log3N and the storage space with N log2N.