Consider the best publicly available language model in 2030. For a single 2023 US dollar, how many 2K word documents can I each generate five words for?

I will pick a somewhat conservative estimate of the best inference cost I can achieve after working for three weeks with whatever funds I have at the time and without using publicly inaccessible resources.

Multimodal models that can operate on text count as LMs for the purpose of this question.

I will only accept answers that range over an OOM like so: