The reason why the VRM is tightly integrated with the asic is that there are extremely stringent constraints on the power integrity connecting these two parts. To first order, a decent GPU draws about 1000 amps of current, so even a single milliohm of resistance would absorb the entire voltage supply. Not just the ripple tolerance, the entire power rail! This is before we even get into the ludicrous requirements on transient behavior, inductance, etc, of the connection.

They do make GPUs with a socket for the asic (typically only used in test labs) so it is theoretically possible that you could for example... buy a 2070 carrier and then get nvidia to sell you a 2080 asic to socket into the carrier. But generally the technologies used in GPUs change every generation (memory technology, number of memory channels, display technologies, etc) so you probabably wouldn't be able to keep your carrier more than 1 maybe 2 generations tops.