Even more simply, fft(I(n), dims=1)fft(I(n), 1)
What’s slowing down the comprehension approach is recomputing the exponential for every element. There are various ways to speed it up, but I suspect that brevity is the priority here.
Of course, computing the matrix explicitly, as opposed to using an FFT as a linear operator, is usually unnecessary. I’m curious to know what the OP’s application is.