The question this user is asking is not really about specs. It is about whether the hardware you own should define the limits of the work you want to do. When 70% of your projects involve fine-tuning pretrained models or building custom pipelines, you are not chasing raw compute bragging rights. You are chasing efficiency, memory bandwidth, and the ability to iterate quickly without babysitting a rig. The user already knows a Pro 6000 would crush every task they have, but the cost barrier is real, and that tension is exactly where most serious practitioners live.
What stands out here is the willingness to consider a Mac for serious model training, not because it is the obvious choice, but because the workload has shifted. Pretrained models are the foundation of modern ML work, and they are memory-hungry. VRAM matters more than core count when you are loading large image or video models and trying to push them through custom pipelines. Apple's MLX framework is not a gimmick, and dismissing it outright because it is not CUDA would be a mistake. The fact that this user is asking for real-world experience with an M5 Max tells you they are not looking for hype. They are looking for evidence that the tool can hold up under pressure.
The practical takeaway is this: your workflow should dictate your hardware, not the other way around. If most of your time is spent fine-tuning and adapting existing models, then a machine with unified memory and strong inference capabilities might serve you better than a workstation that excels at raw training but sits idle during long experimentation cycles. The user's instinct to question the Mac is healthy, but so is their openness to it. They are not asking whether Apple can beat NVIDIA in a benchmark. They are asking whether it can carry the load without breaking the bank.
The real answer is to stop optimizing for the 30% of work that is training from scratch and start optimizing for the 70% that is iterative, experimental, and memory-bound. If that means a Mac with a massive unified memory pool, so be it. If it means a used workstation with a big GPU, that works too. But the days of assuming you need a $10,000 workstation to do serious ML are over. The user's dilemma is not a sign of indecision. It is a sign that the tools have finally caught up to the work, and the choice is no longer about capability. It is about fit.