One box.

A local AI system for a firm of 5 to 30 people is one workstation: a tower under a desk, plugged into an ordinary outlet. The loud, expensive part of AI is training the model, and that happened long before the model reached you. Your machine only has to read, think and type, and for that, one well-built box is usually enough.

The graphics card does the reading.

The part that matters most is the graphics card. Its memory decides how much of a document the model can hold in mind at once, which is the difference between reading the whole file and skimming snippets. A bigger card buys longer files, faster answers and more people using it in the same moment. The card leads and the rest of the box is built around it: every part in a Homestead build is chosen for AI work, so the memory, the storage and the cooling keep pace with the card instead of slowing it down.

Sized by the busy hour, not the headcount.

Tiers are sized by how many people lean on the machine at once, not by how many work at the firm. A firm of eighteen does fine on one strong card until two or three people hit it in the same hour; that is where the next tier starts. The bands overlap on purpose, and the sizing conversation is a short one.

What the models need from it.

Open-weight models, the kind you can own, keep getting better at office work, and they come in sizes. Smaller ones fit modest cards and handle drafting and triage; the larger ones we install read whole case files and answer with page cites. Your machine is built to swap them: when a better model ships, the upgrade is one-time work on hardware you already own, not a new subscription.

What it costs to run.

It draws a few hundred watts under full load and a fraction of that at idle, on your own bill at your rate. No cooling plant, no water, no co-location fee. The machine is yours, serviced like the rest of your equipment. Higher tiers add network backup.

The four published tiers run $7,500 to $23,300, paid once, with hardware at cost from the receipts plus a 15% procurement, assembly and testing fee. Current figures live on the investment page. The sizing question comes down to how many use it at once, and a private demo in your office is the place to settle it.