If no hard versions are given, but only e.g. <= 2.1, pip will download EVERY SINGLE VERSION until 2.1 to look for metadata. That easily can take hours if it happens multiple times.
One problem we have is that support for any repository features beyond PEP-503 (the 'simple' html index) is limited or entirely missing in every repo implementation except warehouse - the software that powers pypi. So if you use artifactory, AWS codeartifact, sonatype nexus, etc, because you are running an internal repository, PEP-658 & PEP-691 support will be missing, and uv runs slower; you may not even have accept-ranges support. (and if you want dependabot, you need to have your repository implement parts of the 'warehouse json api' - https://warehouse.pypa.io/api-reference/json.html - for it to understand your internal packages)
I've been playing with https://github.com/simple-repository/simple-repository-serve... as a proxy to try to fix our internal servers to suck less; it's very small codebase and easy to change. Its internal caching implementation isn't great so I wrapped nginx round it too, using cache agressively and use stale-while-revalidate to reduce round trips, it made our artifactory less painful to use, even with pip.
Further, if a package index supports PEP 658 metadata, pip will use that to resolve and not download the entire wheel.
uv does the same but adds extra optimizations, both clever ones that pip should probably adopt, and ones which involve assumptions that don't strictly comply to the standards, which pip should probably not adopt.