There's an interesting dynamic around asynchronous code in that the way to do it the fastest is to have stackless coroutines which operate on state objects on a batched fashion. This limits what you can do, but completely eliminates a lot of the overheads that are associated with async.

The inclusion of generic async semantics at the language layer is itself a compromise. Rust may have gone a little too far down the "reduce overheads" rabbit hole given that this is the case. Go definitely didn't go far enough for high-performance use cases.