[zahlman]

The issue there is how much gets downloaded from PyPI, not how much storage space it takes up. Making an archive copy of all of PyPI would only require a handful of ordinary drives. But the most popular packages get on the order of ten million daily downloads each.

When someone downloads and installs a package, it's generally a single wheel or sdist, out of a potentially massive version/platform matrix for that project. That inflates the storage cost, but it isn't why the bandwidth requirements are so high. A ton of CI systems are apparently poorly designed, individual wheels are bloated, and we can't use the best available compression. Those are the biggest issues.