It ignores the fact that for big blocks to be happening, there must already be 75% of hashing power signalling support for it. When that happens the first big block can be mined, before that bigger blocks are not mined.

So now you have 75% of the network working on a chain that supports bigger blocks and mining bigger blocks, whilst 25% try to mine a chain that only contains <1MB blocks.

It doesn't take a genius to figure out which chain will *very quickly* becomes longest.