We Do Not Call an API. We Call the GPU Down the Hall.

+ what changes operationally when inference happens inside your own building