It's honestly a treat for personal use if you have the mind of a tinkerer, but it's a difficult proposition to sustain in business (you basically end up vendoring yourself if big enough).
Nix has a very different proposition that solves from the "inside" a problem that is currently usually solved from the "outside" using e.g. Ansible in ops/infra, or Vagrant in dev.
In my perspective, the outside/black-box solution is ultimately brute-force, plain and simple. It works because it scales well, because copying data is cheap enough (re. deduplication, delta sync, etc) You hit limits when systems grow older and complexity creeps in no matter what however, it's an approach that requires a blank slate every now and then.
But the inside package approach is elegant in that it's absolute, it's not conjuring a black box that "should just work" if Ubuntu18.04-387lts-32-1.989-2a and my_fav_package.1.1.1.9.3.5-f work fine together, this time around. Sure, there is testing but we're back to the popularity/support limit.
In the end, in a world where some of the stack is basically nailed and can be modularized with clear forever-true expectations of I/O (the content, not the hardware), then Nix eventually prevails— but we really have to evolve "LTS" for what it means (like we'd do in construction, electricity, plumbing...), and not a half-trendy windmill / rat race. But we have to think of those systems that could transfer almost-as-is well from now to 2050 or more, not 2025-2030. It's not impossible, it's what COBOL did, does as we speak.
I think something like Nix could help shift perception in the right direction, but I expect mindsets to take a good part of the decade to change deeply, if it happens.
Part of the reason is probably that you need the proprietary drivers to use CUDA efficiently and as they are not in the (upstream) kernel you have to use some other tools, like unofficial repositories or the native installer (which doesn't always play well with the OS package manager, system upgrades, etc.). It's a real PITA.
It motivates the need for a new take on package management by analogy to memory management in programming languages.
It contrasts how software deployment works today (“Fuck it, I’m sure /usr/bin/local/python is a reference to a python 2.7.12 install with PIL available”) to how we used to write software (“Fuck it, I’m sure 0x7FEA6072 is the address of the array of files; so clearly, assuming a 32-bit address space, the seventh file is at 0x7FEA608A.”)
In both cases, as long as your assumption was correct, things go swimmingly. But it could be easier and rely less on hoping your assumptions are correct, and more on things that are verifiably true. And that's what Nix offers: a way to build software that is insulated against assumptions and "it works on my machine".
Thesis: https://edolstra.github.io/pubs/phd-thesis.pdf