If you are doing a software implementation, the MixColumns would take the longest because you have to do a multiply after pulling from memory and then putting it back. Generally speaking, in software you would use lookup tables for the naive approach, so everything that is not cached would be expensive. I say this is naive because it is susceptible to timing attacks, such as in this paper. The sbox will almost always be a lookup table in the "homework" problem type of implementation.
If you want to make an actual AES implementation in software, you must first understand how AES can be affected by timing and how caches do not run in constant time. You will like explicitly calculate everything.
If you have hardware, we look at expense a Joules, and that's not what anyone outside of device physics talks about. There's answers on this site where I talk about the power required for different choices. From the "software" point of view, if you have AES-NI instructions, there's no difference in cost of the steps because they are all abstracted for you.