This is likely a function of yields of the process. The better the yields, less the need to bin chips by what works. Likely the M4 yields are quite good at this point. Just a WAG.
"The 14-core M3 Max only enables 24 out of the 32 controllers, therefore it has 300 GB/sec vs. the 400 GB/sec for all models of the M1 and M2 Max, while the 16-core M3 Max has the same 400 GB/sec as the prior M1 and M2 Max models."