Even clustering does not alleviate these problems, it merely spreads out the problem over multiple systems.
When a PC becomes capable of performing in the terafops range and mainframes have not advanced, your argument will have merit.
Wishing that PCs are faster than mainframes won't make it so.
Never did, but they are faster than mainframes of just recent memory. Again, most companies do not and could not use such power. Reliability seems to be the dominate criteria right now for purchasers, and clustering gives them that. Distributed clustering is even better. My last large mainframe installation was a dozen 370 systems clustered across the globe. While an IPL was required of at least one system each day, we were off-line only 14 minutes the entire year, on average. That level of up-time is had also with clustering on Intel boxes at a price 1/10th that of a mainframe. While the mainframe might be faster, that level of power is not a norm.
Anything that needs terraflops of power will most likely be a parallel process and can be clustered. Even mainframes use multiple processors, they are just in the same box.