I came here to say something similar, but with a different industry. None of the people/groups I know using matlab have moved to pytorch, or even gnu octave which is a much closer fit.

Because its not really about matlab so much has the huge collection of library/extensions for matlab or its close relationship to simulink. Decades of industry specific modeling/simulation/tooling built into peoples workflows are hard to change.

If you look at the "applications" section here https://www.mathworks.com/products.html?s_tid=gn_ps pytorch really only touches on a couple of them AFAIK.

Deep engineering tooling is one of the places where opensource really hasn't yet made inroads. From the lack of a good solidworks/cad toolchain to the modeling/simulation space. Even EDA remains deeply dependent on proprietary tooling despite the recent success in a number of places. So while board level design with kicad may start eating into the Altium space now that its beginning to get RF/etc capabilities. I don't see it biting into the cadence/synopsis/mentor space anytime in the near future if ever due to the secrecy of the foundries.

I think a large part of that is there is a lot of breath in this space, and the customer base is generally quite small.

So I think a good opensource PCB layout/modeling tool, and a decent parametric 3d cad program will happen a lot sooner.