My understanding is that sha256 asics are very simple and basically work in parallel*, so a 130nm design can easily be scaled to a 40nm design by adding more cores.
*I might be wrong about that, but they are connected in some simple way where you can add more cores and not have to tinker too much to make them all work.
This means that for the same sha256 asic design, you should have a performance evolution more or less equivalent to the square of the size reduction from one node to the next (physics permitting - 28nm doing some funky things AFAIK, but 40nm should still be safe).
Actually, the performance increase is even greater than size reduction, because the clock speed is also usually increased.
BUT on the other hand, the production cost for the same die area is MUCH higher for smaller nodes. NRE costs are also MUCH higher. Of course, the overall cost per GH/s is still lower, but not
that much lower.
P.S. Also, developing smaller node design is not only more expensive, but takes much more time. For example, at 130nm you can do a full-chip physical simulation, while for 28nm simulating the whole chip would take eternity, so they usually simulate only small parts of the chip, which increases probability of errors and degradations. The software toolchains are also much more complex and expensive. That's why 130nm chip can only take a few weeks to develop, while 28nm takes several months at best. KnC managed to shortcut this by using a standard-cell design (all other existing Bitcoin ASICs are full-custom), but this means that future 28nm chips will be both more cost-efficient and more energy-efficient than KnC.