@hongminhee洪 民憙 (Hong Minhee) 
from the point of view of someone who is "maintaining" a JSON-LD processing fedi software and has implemented their own JSON-LD processing library (which is, to my knowledge, the fastest in it's programming language), JSON-LD is pure overhead. there is nothing it allows for that can't be done with 1. making fields which take multiple values explicit
2. always using namespaces and letting HTTP compression take care of minimizing the transfer
without JSON-LD, fedi software could use zero-ish-copy deserialization for a majority of their objects (when strings aren't escaped) through tools like serde_json and Cow<str>, or System.Text.Json.JsonDocument. JSON-LD processing effectively mandates a JSON node DOM (in the algorithms standardized, you may be able to get rid of it with Clever Programming)
additionally, due to JSON-LD 1.1 features like @type:@json, you can not even fetch contexts ahead of time of running JSON DOM transformations, meaning all JSON-LD code has to be async (in the languages which has the concept), potentially losing out on significant optimizations that can't be done in coroutines due to various reasons (e.g. C# async methods can't have ref structs, Rust async functions usually require thread safety due to tokio's prevalence, even if they're ran in a single-threaded runtime)
this is after context processing introducing network dependency to the deserialization of data, wasting time and data on non-server cases (e.g. activitypub C2S). sure you can cache individual contexts, but then the context can change underneath you, desynchronizing your cached context and, in the worst case, opening you up to security vulnerabilities
json-ld is not my favorite part of this protocol