If you're old enough to remember the GTX 260 with a 448-bit bus, you might be old enough to remember the 128-bit 6600 GT, or even the 128-bit Geforce4 Ti4800.

The GTX 260 was kind of an outlier in the grand scheme of things. The fact that the HD 4870 could beat a GTX 260, with literally half the die size and bus width. This should be proof enough that the GT200 was an extremely bloated design.

The reason for Fermi cutting down the memory bus size was because of GDDR5's increased signalling requirements making PCB routing more difficult. Even the bigger GF100 went down to a 384-bit bus.

The GTX 960 was occasionally slower than the GTX 760, so the memory bus cutback in that scenario did actually matter. Maxwell relied on larger caches to reduce memory traffic, not delta color compression (which was added with Pascal).

Turing was on an old node when it launched (TSMC 16nm with reticle limit increase), that's why the memory bus width was kept up despite adding GDDR6. Ampere was similarly on an old and unused node (literally nobody else wanted Samsung 8nm at the time). The same argument could be said for Blackwell, but there's a huge silicon shortage.