not sure it always could. we had to continue using conda for binary packages like cudann, or anything where a wheel wasn't available (nvidia now has Jax-compatible wheels for cuda packages, but it didn't before)

You're asking for something different here. What conda does is package all binary dependencies (e.g., libpq, cuda) that normally should be dynamically linked and provided by your OS.

The pip ecosystem (including uv) doesn't encourage this. Although you could and many have, the build tools don't support it as first class.

For why? Ask what happens if you want an updated version or specific version of CUDA? The library author must add support for runtime linking or users would have rebuild the entire conda package.