This proposal defers even looking for the code. The LazyLoader already in the standard library would eagerly look for code, but just record a file path and not actually store any bytecode data, never mind deserializing or running it.

The rationale described in the PEP is that some systems try to `import`, for example, across a network share, so even searching the filesystem is slow and there is a desire to defer that (and avoid it on runs where the corresponding code isn't executed).