Apple Silicons architecture is better for running huge AI models but much worse for just about anything else that you'd want to run on a GPU, bandwidth is far more important in most other applications.

That's not even close, the M4 Max 12C has less than a third of the 5090s memory throughput and the 10C version has less than a quarter. The M4 Ultra should trade blows with the 4090 but it'll still fall well short of the 5090.