A $45,000 stipend in San Francisco and a $32,000 stipend in Urbana-Champaign might be the same offer. Nobody tells you that at the visit weekend, because the visit weekend is a sales pitch with free lunch.
Look, I get the pull of the big nominal number. An offer letter with a big figure feels like winning. And then rent shows up. Rent has a way of re-reading your offer letter for you, line by line, and correcting your enthusiasm.
So here is the only comparison that matters: stipend minus rent. Not the stipend alone. Not the ranking someone posted on a forum in 2019. Stipend, minus what housing actually costs near that campus, minus fees, plus summer money if it exists. Everything else is marketing.
The pattern behind the big numbers is not mysterious. Private universities in coastal cities pay the most, because they have to, and programs in expensive tech hubs add a premium that rarely covers the premium they are compensating for. A ten percent bump against a housing market that costs twice as much is arithmetic, not generosity. Do the subtraction yourself. Trust the calculator over the brochure.
Then there is the part of the offer that is not a number at all. Some programs guarantee five years. Some guarantee one year and cross their fingers. I would take a lower stipend with five years in writing over a flashy number that turns into a TA scramble in year three. Ask, explicitly, what happens if your advisor's grant dries up. Watch the pause before they answer. That pause is information.
Benefits do the same quiet work. Health insurance, fee waivers, summer funding: on a grad budget these are thousands of dollars wearing a different name. Two identical stipends can be very different paychecks once you add up the rest.
Compare all of it at once. That is the whole point of phdstipend.fyi, which keeps per-program figures across 798 US programs: stipend, guaranteed years, benefits, adjusted for local costs. Stipend numbers mean nothing without cost of living, and this is where you see them honestly.