AST unparsing of infinity numbers produces a string which can't be evaluated because inf and infj are not builtins.

>>> from __future__ import annotations
>>> def f(x: A[1e1000, 1e1000j]): pass
... 
>>> f.__annotations__
{'x': 'A[(inf, infj)]'}

See how this problem is handled in Tools/parser/unparse.py.

There is similar problem with NaN. NaN can't be a result of parsing Python sources, but it can be injected manually in AST, and it can be a result of third-party AST optimizer.

New changeset 258f5179f9494189e6a80311c86981d2a88ba9d6 by Batuhan Taşkaya in branch 'master':
bpo-32894: Support unparsing of infinity numbers in ast_unparser.c (GH-17426)
https://github.com/python/cpython/commit/258f5179f9494189e6a80311c86981d2a88ba9d6

It seems like the fix works as expected:

>>> from __future__ import annotations
>>> def f(x: A[1e1000, 1e1000j]): pass
... 

>>> f.__annotations__
{'x': 'A[1e309, 1e309j]'}

>>> A=list
>>> eval(f.__annotations__['x'])
list[inf, infj]