Naive pandas is going to limit you to RAM sizes when parquet is deliberately meant for huge datasets.
Anyone reading who is currently using pandas and running into this issue-polars is generally faster, and now supports (still new though) out-of-core workloads.