Depends on what you’re doing with the AI. Macs are one of the most affordable ways to get a bunch of VRAM with current pricing insanity, but prefill is much slower than a regular CPU/GPU setup. You definitely want an M5 (massively faster prefill than M4) but it’s still not going to beat an Nvidia GPU.
For reference, I have a base M5 Pro (15C CPU/16C GPU) MacBook Pro with 48GB RAM, and a box at work with a 7950X, 96GB RAM (holy hell was I lucky we bought that before current prices), and two 5060 Ti 16GBs. The M5 Pro runs Qwen3.8-27B fine (my own mixed-bit - 4, 6, 8 bit - GPTQ quant), ~400-600 tok/s prefill and 34-60 tok/s decode depending on workload and context length. I run Strata on the Ryzen box, it’s at… I forget what prefill, I believe ~1500-2000+, and 50-70 tok/s decode (or higher, close to 100 sometimes), running a 4-bit quant of Qwen3.8-Flash-Next, a 125B MoE. The real hope for now is that the bubble pops sooner rather than later and the market floods with defunct datacenter parts so us peasants can purchase RAM and GPUs again.