Cloud models will always have massive benefits of scale.

Caching is the simplest one to understand, cloud providers often reach a 90% cache hit rate, so hosting the same request locally on the exact same model on the same hardware is often way less efficient than on the cloud where a group of users generates a healthy cache.