helped someone decide between two computer science phd offers, both for the same research direction, llm inference and systems. on paper the comparison looked lopsided. one school sits in a major tech corridor with a bigger faculty and a better ranking, the other is a solid regional program. ranking says the first one, and ranking would have been the wrong way to decide.
the funded offer at the lower ranked school comes with a direct advisor whose group works on exactly the niche, caching and inference for language models, with an industry collaboration and real hardware access. that's the hireable topic and the fit is live, not theoretical. at the better ranked school the closest advisor to that niche is at capacity and nobody else owns that exact space. a good offer with a mentorship gamble loses to a good offer with an actual match.
the funding math is a trap too. the nominal stipend at the better ranked school is higher, but the cost of living is several times worse, so the disposable income is far lower. unless you can commute from home, in which case the rent disappears and the higher nominal number suddenly wins. location doesn't nudge this decision, it flips it.
the strongest fit turned out to be a professor people skip because he's a compiler and gpu systems person rather than an ml person, part time at a big lab working on the tensor compiler that everything else runs on top of. for someone who already has the models side, that's the layer below where the hiring money is. the takeaway generalizes. phd programs get ranked by department, but you don't join a department, you join a person. advisor match, funding, and cost of living beat the name on the building, every time.