There has been a longstanding gap in the Python packaging ecosystem that has somewhat annoyed me, but not enough to do anything about it: we haven't really had a good way to compose multiple layers of Python virtual environments together, allowing large dependencies (like AI and machine learning libraries) to be shared across multiple different application environments without having to install them directly into the base runtime environment.
Utilities for collecting up an entire Python runtime, an application, and all its dependencies into a single deployable artifact have existed since before the turn of the century.
We've had standardised virtual environments (allowing multiple applications to share a base Python runtime and its directly installed third party packages) for almost as long.
We've had tools like wagon which allow us to ship a bundle of prebuilt Python wheel archives and install them on a destination system without needing to download anything else from the internet at installation time.