This used to be called optical computing. A former co-worker of mine had done his doctoral work at Oxford on precisely this subject. This would have been back in the 1960s, if I'm not mistaken. At the time, it wasn't possible to create optical components on the same scale and power as electronic, but they showed it was possible at least in principle. Whether it's now practical, I don't really know.
The current supercomputers all use fiber optic interconnects. The major types are known as Infiniband and Tofu. Well, they use both copper and fiber but they already use tech like this, just not on the actual processor die. They don't really need that though, processors in super computers are already highly interconnected.
There are also some research labs working with laser-based computation that uses light interference to do calculations. They shine lasers in one end of a non-linear optical material and can get calculations out the other end. Non-linear optics is magic mixed with stupidly complex math. You can imagine making a NAND gate would be rather simple and so you could make any other gate from those simple NANDs.
As for light buses in future chips. I highly doubt that we'll see that by 2018 in consumer chips. Reason being that right now, signal propagation speed on copper (and aluminum) is effectively infinite in a chip. The delay is in the logic gates themselves. So by speeding up data transfer IBM really hasn't done anything super useful if you still have use the slow gates that we have today. It might be useful if you're accessing main memory but I doubt that we'll really make main memory fast enough to warrant that sort of tech.
Think of it this way, right now the slowest part of the chip by far is the logic. Feeding more information into that logic isn't going to make it go faster. I think you'll see supercomputers going for FPGA coprocessors and in-memory operations before you see them going optical. Actually, I know what Argonne National Lab is planning to build in 2018 and it is FPGA based I think popSci is forgetting that the supercomputers planned for 2018 have already been designed and are more or less just spending the next 5 years physically being built and funded.