I have been thinking about the struggle to build large computing systems. The major challenges of the current age of computing are power, cooling and scalability. The machine room and energy costs are dwarfing everything else. This concern has even spawned the creation of a new ranking of supercomputers, the Green 500, which ranks machines by compute power per watt rather than simple compute power.
Imagine a single computing node that is self-contained and sealed. Each has its own storage and set of processing cores, a solar panel of some sort for power (many advances have been made in solar power recently), some kind of power storage (maybe even capacitors). For cooling, it could use the solid state heat pumps from Cool Chips and/or circulating fluids. Each of these nodes could be connected to its four nearest neighbors by some network — a 2D mesh is not the best network one can design, but it is infinitely growable. Alternatively, one could make a hexagonal mesh.
Imagine this computer, allowed to grow in some wide flat country, stretching for miles in one all powerful matrix :). It would, in theory, be infiinitely scalable and require no machine room at all. You would probably mount it on some kind of rack a few feet above the ground so that people (or robots) could travel beneath it and fix or replace nodes. The system administrators would be like farmers, harvesting the compute cycles to feed the hungry scientists of the world.