If the connection isn't perfect, which can be really hard to determine when you have to check 400 connections all of which are sandwiched between two boards (so, X-ray), gotta bake it back into shape. If balls fail you need templates and fresh balls to re-ball the chips. Oh also routing sucks when you have to tie signal lines around inside a matrix of several hundred vias.

Unless the footprint and pinout and protocol for successive generations of chips is the same (which is highly unlikely), building a board with sockets would be great for adding expense and slightly increasing the ease of manufacturing but not necessarily make a better end product. And it still doesn't solve the problem of me not wanting to design a couple-hundred-amp multiphase regulator and worry about high-power-density cooling concerns when, in my opinion, that design concept is just about the worst available. Remember, we like simple and efficient and the inherent reliability simplicity and efficiency help bring to the table.

The only way to make it worse worse is to greatly increase the required power without substantially increasing the die size available for cooling - so, like Hashfast chip instead of Spondoolies chip. The only reason SP's Rockerbox gear works as well as it does (which, by the way, it doesn't really work for crap if your ambient is over about 30C) is because the chips themselves are built like tanks and handle 120C. As much as I may conceptually dislike SP gear, their chip design design guys are quite skilled and I'm impressed at what their machines can do given the limitations their design concept imposes on efficiency.