For most practices, AI arrives as a per-seat subscription. That shape is easy to start and hard to stop: the cost scales with headcount, rises at renewal and runs for as long as anyone uses the tool. Per-seat AI licensing is heading the same way per-seat software licensing went, which is worth testing before you commit a whole studio to it.
A private deployment inverts the shape. An open-weight model running on infrastructure you control is a fixed cost plus upkeep. The marginal cost of one more user is close to nothing.
The real cost lines
Priced honestly, a private deployment has five lines: GPU hardware, or GPU capacity rented in a UK datacentre; the power and hosting that hardware consumes; patching and maintenance, exactly as for any other production system; periodic model updates as better open-weight models are released; and someone accountable for all of the above.
Practices already running GPU infrastructure for rendering or virtual workstations start from a better place than they expect. The same GPU capacity that renders a visualisation can run a private model, so the hardware line is often smaller than the sticker price suggests.
When the sums favour it
Per-seat licensing across a forty-person studio is a five-figure annual commitment that never ends. A private deployment crosses that line surprisingly quickly once usage is heavy and broad. It keeps getting cheaper relative to subscriptions as open-weight models improve.
Sometimes cost is not the deciding factor at all. Where a client agreement will not tolerate project information reaching a third-party model, a private deployment is the only compliant route. What it costs is the price of being able to hold that work at all.
Start with measurement
The right answer starts with measuring how your practice actually uses AI, usually through a contained pilot, before committing in either direction. That is a conversation worth having before the renewal notice arrives.
WIT deploys and supports private, open-weight models on Inevidesk: on your own premises, in your own datacentre or in a UK Level 4 datacentre, sized, secured and maintained like any other production system.