It's looking like ghc-nix is massively slower (β10x) on our internal codebase (which is β3000+ modules). I still haven't diagnosed the exact bottleneck, though. We have some tooling for instrumenting Nix with open-telemetry that might help profile things.