So a big part of what's hard in HPC (really depends on what you do in the area. I mainly just use the machines to run stuff. You don't write software on the machines) is getting things to install (and run as expected). The people that maintain the machines will generally have your basic stuff installed (things like GCC, clang, your intel compilers, python, hdf5, etc) and you can load different versions using modules. Anything not there you need to do a local build of (your software and anything that your software depends on that isn't a super common library). Building these packages can be a real pain because architecture is so different (some people work on things called Data Parallel Primitives that try to make these easier. One common software that uses this is VTKm. They us a lot of template metaprogramming, which results in crazy errors that are hard to read but many times works more easily). Basically take the frustration you have for building things from source and multiply that by at least 2.
So basically what I've heard about Taihu is that installing any packages and writing code for the system is even more of a pain (like adds months to your work (probably an exaggeration)) to write for. Libraries that are commonly used aren't supported and so many things need to be written by hand (including all the optimization). So they go (from my knowledge) fairly unused. Hence calling them Linpack machines (someone got linpack running and submitted test results. Doesn't say how utilized the machine is).
But this could also be people in the labs playing up lab pride too. So take everything with a grain of salt. I want to try to be fair here. (US labs do have lots of internationals at them, not just western countries, plenty of Chinese nationals. They just want smart people).
As for performance, it depends on a lot of things (it is actually REALLY hard to measure). Linpack is just a good baseline and has been highly optimized. But it doesn't tell you how fast your software is going to be because frankly your software could be completely unoptimized. It is highly dependent.
That said, having the fastest/biggest HPC computer doesn't mean too much. No one is running across the entire machine at one time. It is generally a lot of different programs and groups working on the same machine. There are benefits though. It is easier to have a single big machine than many smaller machines and it does enable you to do the science you want to do if you do end up needing a huge capacity. More bang for your buck. And for only a few hundred million dollars too, not a bad price.