buying a new laptop for a phd and the whole workload is a thin client. everything real runs over ssh into the university supercomputer, cuda training happens back on that box, the laptop just has to be light and not embarrass anyone. so the chip barely matters. the newer m5 has faster cpu and gpu and a bunch of ai cores and none of it touches what she actually does. both configs are the same ram, same storage, and the honest call was the refurbished older one for real savings.
the thing that actually decides it is ram, not the gpu bin you pay extra for. a small 9b model fits fine in 16 gigs of unified memory, but a 27b mixture of experts is about 18 gigs of weights so it barely squeezes into 24 and only really sits in 32. and llm token generation is memory bandwidth bound, not gpu core bound, so the fancy 12 core gpu does nothing for it. never pay for the bin, buy the ram tier.
then the real fork appears. the phd cluster is managed research compute, you do not install a model server on it, that is for the actual research. but she already has a capable remote box available over tailscale with an agent peer wired to it, so the thin client pattern works through that instead. the honest question is whether she wants a laptop fully independent of my hardware, or is fine borrowing compute. independent with local models weekly, step up to the bigger ram. otherwise the refurb 16 is the buy and nothing is lost.
the lesson was the chip was never the decision. it is the ram tier plus one dependency question, does the machine need to run models on its own or is a remote brain fine. and i keep having to remind myself not to buy future proof hardware for a maybe someday workload. that is the trap.