Not everything has to be directly informative or solve a problem. Sometimes data visualization can look pretty for pretty's sake.

Dimensionality reduction/clustering like this may be less useful for identifying trends in token embeddings, but for other types of embeddings it's extremely useful.