It doesn't have to be only one server in one datacenter though.
It's more work, but you can have replicas ready to go at other Hetzner DCs (they offer bare metal at 3 locations in 2 different countries) or at other cheaper providers like OVH. Two or three $160 servers is still cheaper than what they're paying right now.
It doesn't have to be one server in a single datacenter, though. It adds some complexity, but you could have a backup server ready to go at a different cheap provider (Hetzner and OVH, for example) and still save a lot.
That uses a workaround based on WiFi debugging even though it's all local. It doesn't run if you're not connected to a trusted WiFi network, you have to set it all up when connecting to a new network, etc.
Not only users are not connected to WiFi all the time, but in many developing countries people often have no WiFi at home and rely on mobile data instead. It's a solution, but not a solution for everyone or a solution that works all the time.
It's easy to use a different calendar, search engine, etc, but it's far from easy to use an Android device without Google Services. Can be done, but banking apps, contactless payments, etc, become painful or impossible.
Regarding censorship, that works only if there's no network side blocking, otherwise the unencrypted requests to root servers also get intercepted. That's why some people use DoH as the upstream for their resolver.
NMT doesn't "contain" tranformers and deep RNNs, it can use them. LLMs use a transformer architecture, not everything using a transformer architecture is an LLM. NMT can actually use an LLM, but that's not the case according to the documentation you linked, they use a parallel dataset to train their models.