A private inference bench for principals who already think in unit cost. Volume stays in the room. Outside intelligence is reserved for decisions that earn it.
Three machines in one room, under one pair of hands. What they are, how they are packed and how the work is split across them is craft, and it stays here. That the work runs on hardware bought outright is the part that concerns you — because it is what makes the floor under your cost a floor you own rather than a price someone else sets.
Measured July 2026. This is the house instrument, not capacity for hire. Client systems run on hardware they own.
The commercial question is not peak theatre. It is how long real bulk takes when the path is local and the bill is electricity.
One dated run, July 2026, on one workload. It is evidence, not a service level, and it does not generalise to yours — each commission is measured on its own workload. Reproducible on request: dates, not adjectives.
Most of what a workload asks for is bulk, and bulk does not need the most expensive answer in the world. It needs a correct one, cheaply, at volume. So the queue is split by what each task actually demands.
How the tiers are packed, what runs where and on what is craft, and is not published. The routing principle is the offer; the instrument is not.
Bring the real problem. The system is designed around your work, built end to end, and handed over on hardware you own.
No CMS, no framework, no cookies — hand-cut HTML, run on owned hardware. Cheap costs electricity. Expensive is judgement. Every figure carries its date.