Once you have control over generating the binaries, you can create a system that is more akin to what Nix offers, where other parties can setup secondary caches for their projects, and if they don't the main registry has the option to decide what to do about it without breaking any package.
Currently PyPI is railroaded into a fairly tight set of choices because it can't exercise a lot of control over binary data.
Thanks. The blog post alludes to mirrors but I didn't realize that was the primary goal of this project.
I am somewhat skeptical that these projects will have an incentive to actually set up mirrors. For example, I imagine Tensorflow likes getting free hosting.
I haven't checked, but usually when it's this visible, big tech companies tend to reciprocate so I would be surprised if Google didn't sponsor the PSF in any way.
That said, once deleting prebuilt binaries doesn't break the package anymore, if Google were to not play fair, PyPI could simply delete them. Users would experience temporary discomfort (but nothing would be irremediably broken), and soon after I'm sure that Google would decide to setup a cache for TF.
PyPI would almost certainly not even need to actually do it, the implication that it could be done would probably be enough to align everybody towards what's best for the ecosystem.
Currently PyPI is railroaded into a fairly tight set of choices because it can't exercise a lot of control over binary data.
The blog post mentions this point directly.