The general rule of thumb is break it up large KML files into smaller KML files often all stored in the single KMZ file.
Then they are loaded via the parent KML file via NetworkLinks with Regions to only load those features that are in view. The parent KML should have each NetworkLink with the appropriate Region and Level of detail (Lod) element to prevent all of the KML files from loading all at once.
Google Earth can scale up to many millions of points and features but not all at once.
The exact number of what is "too large" takes trial and error. For example, 100K places may work in some situations, but crash Google Earth if each feature had a large size due to large amounts of inline HTML markup or ExtendedData, etc. Smaller groups of features such as on the order of 5K or 20K tends to work better.
As mentioned above, I posted described experiment by creating single KML files of 200K and 500K features. Things tend to break with that number of features.
--jason