TBH I'm not sure exactly what the hardware entails, I just work on the same floor as a couple labs that use the fMRI facilities so I know how much people stress over booking time on the machines. They somewhat recently added a new one that has stronger magnets but the older ones are still frequently booked solid too.
From a neuroscience perspective I do think the good fMRI labs have improved their analyses over time. One thing increasing in popularity that has been helpful IMO is scanning the same participant across multiple days. It's a pain in the ass to get people to come back so it has generally been the norm to scan each participant in a study only once. But more powerful analysis techniques can be performed with these multi-scan datasets, as heterogeneity between and within individuals can be better accounted for.
This review focuses on the default network, but it mentions how the discovery that the default network is actually multiple networks that were being blurred together was enabled by the multi-scan approach. It also gives a decent general gist of some of the things that have been reliably reported in fMRI research versus some things that are lacking, but mainly just from the perspective of questions about the default network: https://pubmed.ncbi.nlm.nih.gov/31492945/
I also like the following paper as a deep dive into the scans of a single person, because he splits apart some seemingly small things like having coffee/breakfast in the morning and shows the effect on scan results. I also think a multi-modal approach is really helpful, so I appreciate the other datatypes included. Plus it's kind of funny the PI wrote a paper entirely about himself: https://www.nature.com/articles/ncomms9885
So yeah I think there is some self reflection and progress being made in the field in a meta sense, it just unfortunately can be drowned out (especially for people in other fields or even subfields of neuro) by the sheer popularity of shitting out fMRI papers.