My experience with F# (which was way back in version 2.0, but based on OCaml which does type inference very similarly), was that the inference would become slower as the size of the project increased, eventually taking multiple seconds. I eventually adopted a habit of adding type annotations to commonly-used functions because it mitigated the problem. These projects were not huge by any means (5-10K LOC).

There is probably a good compromise to be found here. It should be possible for a tool to suggest areas to add annotations where it would do the most good for performance. But maybe those are non issues with modern implementations.