The answer depends on what you plan to do with it. Do you need to do fine tuning on a smaller model and need the highest inference performance with smaller mod

The answer depends on what you plan to do with it. Do you need to do fine tuning on a smaller model and need the highest inference performance with smaller models? Are you planning to use it as a lab to learn how to work with tools that are used in Big Tech (i.e. CUDA)? Or do you just want to do slow inference on super huge models (e.g Grok)? Personally, I chose the Nvidia route because as a backend engineer, Macs aren’t seriously used in datacenters. The same frameworks I use to develop on a 3090 are transferable to massive, infiniband-connected clusters with TB of VRAM.