1-2 RTX5090 will be better value than Macs because they have the memory bandwidth for somewhat fast local inference.
and they will become obsolete much slower if at all compared to when the mac will be barely usable once macos stop supporting it in 8 years or so. although maybe by that time there will be local models that you could run straight on the same mac that will reverse engineer it for you and write linux drivers
If you don't want to wait for prefill, you're going to want a CUDA dGPU system.