The difference between the high quality math guys and the high quality distributed jocks at that level is much larger than I think most people think. I come from the engineer side, and it's really bizarre to hear people w/ Ph.D.s in machine learning or physics give talks where they will mess up what is really a junior level distributed programming concept. Happened twice at Strata that I saw. Mind you, that's for BASIC things, not someone trying to right a highly concurrent computational system across hundreds of nodes. No lie, I had a conversation with a machine learn jock that asked me point blank when we couldn't make Matlab go faster.

On the flip side, the distributed engineers might even be worse at the math than the data jocks are at programming. A middling concept for a machine learning guy, like say a Bayes error, is like trying to write concurrent state machine in assembly.

tl;dr - the pool of people that are both elite math jocks and elite system jocks is vanishingly small, as in I would be shocked if that number was out of the triple digits world wide.

Julia has a lot of promise in that it gives the math guys familiar syntax, but appears to have the underlying capability to be very fast on shared memory architecture. It's already got the handles to call out to python / C / Fortran bindings.