In short
On-premise AI is an AI deployment in which the model and the data run on the company's own servers and data is never sent to an outside service.
In on-premise AI, the language model and other models are hosted in the organisation's data centre or on its own servers. If needed, the system can run on a network fully isolated from the internet.
This keeps trade secrets, recipes and personal data away from third parties and makes compliance with data-protection and sector regulations easier. Access is role-based and every query is logged.
Deployment needs server infrastructure — usually with GPUs — sized to the model and the number of users.