I think that instead of going a rather expensive FPGA or custom ASIC path, the future mining machine could be built as custom designed "graphics" card. Using a several RV870 ATI GPU chips in single box.
Of course, designing and later manufacturing a GPU card is not something one could do in a garage, but I think that standard PCIe card producer could do that. For example, there is a lot of very small companies manufacturing 1x PCIe cards for CAN bus, I/O switching, GHz analog sampling, etc. Compared to motherboards, these Radeon 5870 cards have much simpler design with less layers.
My idea is that 4 or 8 GPU chips would be wrapped into large water-cooled copper block. All these would share a 1x PCIe card slot (using extender cable). And eat a lot of Watts, of course. Probably using industrial strength power supply (e.g. 200 Amps at 12V). Also, there will be needed a custom, special built of OpenCL drivers for this beast.
You could combine several blocks in paralel and let the water pump through it. That warm water could be used in house for showers, heating in cold days or even to regain some energy back.