The answer depends on how your data is distributed, and what kind of error you can accept. If the distribution is close to uniform, and you care about having an absolute error of, say, $1000, data does not necessarily need scaling, or you could use standard scaling (min to max).
In any other case, you may want to apply a non linear transformation to your target variable. For instance, if the observations are evenly distributed among 0.01, 0.1, 1, 10, 100, etc., or most specifically when you care about relative error (a certain percentage of the target value) I would recommend to use a log transformation. This is because an error of 10% will always represent 0.1 unit after being log-transformed, whatever the target variable value is.
Most cases, if not all, will lie in the second case.