This paper studies the capacity limits of multi-input multiple-output (MIMO)-reconfigurable intelligent surface (RIS) communication systems when the number of transmit antennas, Nt, receive antennas, Nr, or RIS elements, M , is asymptotically large. We derive a novel asymptotic approximation and upper bound of the MIMO-RIS capacity formulated in a simple enough manner to provide meaningful insights that can then be translated into practical optimization guidelines. Our study reveals that the MIMO-RIS capacity can be decomposed into a line-of-sight (LOS) and a non-LOS (NLOS) part in the large asymptotic Nt, Nr, or M regime. We then also prove that the LOS and NLOS parts of the capacity can further be decomposed into the product between an array gain, i.e., related to the RIS amplitude response, and a directionality gain, i.e., related to the RIS phase response, when M is large. The array gain tends to be larger than the directionality gain, and the latter vanishes in certain conditions. These insights have clear implications for RIS optimization and lead to some simple practical guidelines. For instance, in the large Nt or Nr regime (in the presence of intra-element coupling at the RIS) as well as the large M regime (when √ NtNr is sufficiently large compared to M), the capacity of MIMO-RIS can be optimized without the need for complex channel state information (CSI) estimation since it mainly depends on the RIS amplitude response. As such, it can be optimized by using a simple fixed phase allocation. Whereas in the large Nt or Nr regime and in the absence of intra-element coupling at the RIS, the best strategy for improving the MIMO-RIS capacity is to optimize the RIS phase solely based on the LOS path. This can be easily done via any classic MIMO beamforming method.