Because of memory bandwidth limitations. There are EIGHT memory channels in M1 Max. To pull off that amount of bandwidth you would need to do a lot of crazy engineering and get not that much in return, because this approach doesn't scale well beyond M1 Max performance. You can't just add more and more memory channels. Even if they slap two chips together and use memory controllers of both, total bandwidth still would be less than in one 3090. And it will, most likely, loose to 3090 in all tasks other than video encoding. And you wouldn't be able to put a lot of these chips together, too much overhead. But on the other hand you can put 4x3090 into one PC. At the end it is just cheaper to use dedicated GPU past certain level of performance, especially if you don't make your own hardware AND software.