You can host the model in the building.

With the hardware of the last few years, you do not have to send every prompt to a cloud you do not own. A modern workstation, a server with a proper GPU, even a box the size of a Mac Studio, can run a language model on the floor. What used to need a data centre now fits under a desk.

We keep saying this to plants in Quebec, because the question has changed. It is no longer “should we use AI”. It is “where does it run”.

The hardware is already in the quote.

Lenovo ThinkStation with an RTX card. ThinkSystem in the closet. Mac Studio when engineering already lives on Apple. We size the power, the air, and the keys the same way we size a file server. If the drawing or the packing slip must not leave the parish, the model needs a box you own.

The software is boring on purpose.

Open tools like Ollama, vLLM, or LM Studio are enough to start. Classify tickets. Draft a summary of yesterday’s PDFs. Answer a procedure question without the file leaving the tenant. A private assistant, pointed at your binder, running in-house.

Why a plant cares.

Innovate without a per-token bill you cannot explain to finance. Automate the repetitive pile with a model that has seen your domain. Protect the strategic file: it stays in the building, no third-party processing.

We will still put a human at the gate. The box drafts. A named person posts. That is how it survives March. If you want the cloud, we will say so. If you want the closet, we already sell the closet.