I think you may overstate the difference. Creating exporters to different common neural network models isn't rocket science. Several exporter exist for Julia. E.g. for ONNX, XLA, Mjolnir and NAS.
And keep in mind Julia is just 2 years old. A lot can change in short time.
Yes Julia is obviously a niche today, but so are all brand new languages. An important aspect missing from the comparison with Julia to Fortran and Matlab is that the latter two are destined to remain islands for all future. Matlab has had since 1984 to conquer general computing. Fortran has had since 1957. If they haven't made it by now, they aren't going to make it the next 10 years either.
Julia in contrast was released in 2018. And with is powerful features and high performance it would be very odd indeed if it remain a narrow niche in 10 years as well.
Julia is excelling in the very areas which are hot for future computing, specifically data science and machine learning. Python is primarily a player in this space because of libraries. There isn't much in the language which makes it well suited for machine learning.
However Julia does in fact have a featureset which matches the needs of this domain very well.