The argument was made above that you know more about model structure and therefore should be able to beat the solvers in scaling. In most cases the user can do simultaneous scaling of multiple rows and columns that can greatly narrow coefficient disparity. To see why lets return to the example,
In this model the variable X1 is a sale variable that disposes of the products produced by X2 and X3 as added up in the first constraint. Thus, to preserve integrity of units, I scale X1 and the first equation by the same factor.
This problem knowledge gives us a leg up and is why the user can usually do better. However solver-scaling procedures should also be used so to get the benefit of double scaling.