One immediate speed-up that requires no code changes: when uv creates a venv, it doesn't have to install Pip in that venv. You can trivially pass `--without-pip` to the standard library venv to do this manually. On my system:

  $ time uv venv uv-test
  Using CPython 3.12.3 interpreter at: /usr/bin/python
  Creating virtual environment at: uv-test
  Activate with: source uv-test/bin/activate

  real 0m0.106s
  user 0m0.046s
  sys 0m0.021s
  $ time python -m venv --without-pip venv-test

  real 0m0.053s
  user 0m0.044s
  sys 0m0.009s

  $ time python -m venv venv-test

  real 0m3.308s
  user 0m3.031s
  sys 0m0.234s

To install in this environment, I use a globally installed pip (actually the one vendored by pipx), simply passing the `--python` argument to tell it which venv to install into. I have a few simple wrappers around this; see https://zahlman.github.io/posts/2025/01/07/python-packaging-... for details.

In my own project, Paper, I see the potential for many immediate wins. In particular, Pip's caching strategy is atrocious. It's only intended to avoid the cost of actually hitting the Internet, and basically simulates an Internet connection to its own file-database cache in order to reuse code paths. Every time it installs from this cache, it has to parse some saved HTTP-session artifacts to get the actual wheel file, unpack the wheel into the new environment, generate script wrappers etc. (It also eagerly pre-compiles everything to .pyc files in the install directory, which really isn't necessary a lot of the time.) Whereas it could just take an existing unpacked cache and hard-link everything into the new environment.