A lot of cost is on training it as well. Which you need if you want to "build your own claude". If you run only the inferences with an open model then ya it's directly correlated to how fast you want the responses to come in.