I don't see why people say a 512+ bit memory bus is needed/wanted. What you want is
memory bandwidth, not memory bus width. Memory bus does nothing, is only the number of bits that can be read/write at a given clock tick, so you can increase memory bandwidth (data per time unit) either increasing the bus or the memory clock speed.
Now, memory bandwidth is one of those things that can make bottlenecks if it's too low (GPU don't get enough data to work and wastes time) but don't afect possitively if it's higher than needed (why would I want to have more data availabilty than the data I can process?), AKA, don't directly increase performance or image quality, it removes bottlenecks that impede good performance under given cirtumstances of high memory useage. I've always found much more interesting and spectacular the changes that really afect the pure processing power...
Even so, if more mem bandwidth is needed to feed the processing capabilities of the GPU under the 3d/gaming scene in the target market segment, you can go by the route of doubling (or increasing without doubling using odd quantities of VRAM) memory interface/bus, or using faster memory. Former is the complex, expensive, less flexible way to do it. Lots of additional transistors in the chip, complex PCB, dependant on the quantity of memory being used... now you have GDDR5, 2x faster than GDDR3/4, probably cheaper than 2x memory bus width, less power hungry than GDDR3/4... the way to go should always be 1st use the faster memory which is affordable (specially if it even leads to a power consumption reduction) 2nd if that's not enough (and only if that's not enough), increase memory interface.
I wonder how many of that "billion transistors" of the GT200 are wasted in that 512 bit bus, instead of using GDDR5 to achieve the same performance (better said, to rise the memory-gpu comunication bottleneck to the same high enough point), with a cheaper, cooler, less power hungry vga, all because they know that releasing a high-end vga with "only" 256 bit will hurt their image (thus sales) because all of these people that thinks 256 bit is "low-end".
Since the cost of the chips depends on size (area) of it, and this one depends on the number of transistors and the size (nm) of these, it costs the same a chip with X transistors if they are used to make shader processors, tmus, rops, memory controllers, whatever it wants to be. So, give me a graphics card with GDDR5, half the memory bus bits, and use all the extra transistors to increase shaders or tmus. I don't need to know that I've a 512 bit bus to feel better...
Anyway, lets see what nVidia and AMD/ATi give us this time...