There are no universal2 wheels for NumPy 1.22+, 1.21 was the last version that had them. However there are wheels for both x86_64 and arm separately. Just install using the python you intend to use if you want the latest version. python3 -m pip install numpy should do it.

@charris - I guess I need to elaborate. Our product embeds a Python installation with numpy included. M1 flavor is being cross-compiled from an x86_64 CI node.
Python is being built as a fat binary, however fat binary is not a requirement. E.g. it would be possible to generate two distinct packages, one with ARM version of numpy, and one with x86_64 version. However, I would prefer to not deploy and M1 CI node to accomplish that. And I'm not seeing a way to deploy an arm64 version of numpy on an x86_64 machine (pip3.9 install numpy==1.22.3 --target foo2/ --platform=arm64 --no-deps fails with the exact same error as above on Intel Mac)

See #20787 for discussion of that topic.

I went through #20787 and while it's related, I don't think there's a conclusive answer to question: how to generate an M1 python+numpy package, when cross-compiling from x86_64.

I guess if we step back from the "universal2" flavor of question, it'd be how to deploy an arm64 wheel into a universal Python install structure, when running on x86_64?

This seems more a question for Pip or for Stack Overflow. We have the correct wheels on PyPI, but off the top of my head I don't know if or how it's possible to force Pip (or pypa/installer or Poetry or ...) to do this. NumPy isn't special in this regard.

This is an almost duplicate of #18143. I think the discussion there covers the questions here, and will close this. Please reopen if I missed the point.

To summarize the main point, here is how to create a universal2 wheel from separate thin wheels, if you so desire:

pip install delocate
delocate-fuse $amd64_wheel $arm64_wheel -w .