The question we get asked is "can we have our own AI?" The useful question is narrower: what data does it need to see, and where is that data allowed to be?
If the answer involves contracts, patient records, payroll, unreleased designs or anything a regulator would ask about, a hosted model becomes a policy problem before it becomes a technical one. Running locally removes that conversation entirely - the documents never leave the building, there is no per-seat subscription, and nothing changes underneath you when a vendor updates their terms.
The trade is real, though. You buy the hardware once and own it, along with the responsibility for keeping it running. A local model will not match the largest hosted models on open-ended reasoning. And someone has to maintain it.
Where it wins clearly: answering questions against a fixed internal corpus - your policies, your manuals, your ticket history. That is retrieval plus a competent model, and a competent model is one you can host yourself. Where it usually loses: anything needing frontier-level reasoning, or a workload so occasional that dedicated hardware sits idle.
Our rule of thumb: if the data is sensitive and the questions are repetitive, host it yourself. If the data is public and the questions are open-ended, don't.