I think this is a really key point. I would even go as far to say that most people likely “don’t need” Julia. Most users in the general realm of scientific programming likely are stringing together results from various packages. And for those people, sure if there’s already a package in Python, why switch?
The problem comes when there isn’t a package. The next group of users are those who wrote some simple libraries in Python/MATLAB/R. This is the group that gets buffed when trying to make Cython/MEX work, finds out that PyPy doesn’t work with some libraries they want to use, and tries to use Numba but runs into difficulties the moment they have “non-trivial codes” that aren’t just a benchmark loop and involve some objects. These are the people who find Julia and go “yes, this is what I needed”.
Overtime, these converts make(/made) some really interesting packages in Julia, and that gives the incentive for the “package user” group to switch. But I think that it’s really important to understand that the vast majority of users of the language will probably not know what “broadcast fusion” is or that it occurs, that in-place operations / mutating functions are the “fast way” to write array functions, etc.
And that’s okay! It just means that Julia, because it solves the two language problem, also has a two audience problem, and one should evangelize Julia by appropriately discussing Julia to the different audiences. Some people want to hear about compiler details and type systems for writing libraries. Some people want to see cool examples of solving difficult optimization problems in JuMP, complex simulations using DifferentialEquations, and easily train large neural nets using MXNet. Both of these groups are Julia’s (potential and current) audience.
(I have now written too much so I will stop 
)