I’m not familiar enough with neural-fortran by @milancurcic or fiats by @rouson but I’d say easy-to-use automatic differentiation and a package implementing the standard optimizers used for training such networks. In a nutshell, I’d say it rather is that the ecosystem is still lacking some building blocks rather than an actual limitation of the language.
A good playground may be to port as much as possible of nanochat in Fortran.