It seems we keep building new caches for it in every PR (the ~14 MB one), instead of building it once when merging to the main branch and reusing it following the expectation for GitHub Cache:
https://github.com/NVIDIA/cuda-python/actions/caches