Equipment scheduleSheet H-01

Four builds, priced in the open.

Nobody in this market publishes what a private AI machine costs. So here it is — the tiers, what they run, what they draw, and the case where buying one is the wrong decision.

Unit	Users	Comfortably runs	Form factor	Draw	From (CAD)
Desk A quiet box for a small firm	2–4	14B dense at Q8, 20–30B MoE, 32K context	Sound-dampened mid-tower	~300 W	$6,300
Workgroup A team's daily driver	10–30	70B dense at Q4/Q5, 64–128K context	Full tower, sound-dampened	~1,070 W	$17,900
Department Serious throughput and long context	25–40	70B at FP8, or 120B-class MoE, 128K+	Tower or 4U rack	~1,400 W · dedicated 20 A circuit	$44,900
Enterprise Survives a node failure	60–100	235B MoE, or 70B FP8 with a large KV cache	Two rack nodes, load balanced	~2,100 W · 208/240 V	$123,800

Indicative build cost in CAD, hardware only, before installation and GST, priced against Canadian retail on 29 July 2026. GPU and memory pricing moved sharply through 2026 — quotes are re-priced at time of order and held firm for seven days.

SizingSheet H-02

What fits in how much memory.

Three numbers pick the tier: the largest model you need, how many people hit it at once, and whether they're chatting or running agents.

Memory	Largest sensible model	Precision	Context	Concurrent chat
24 GB	14B dense, 20–30B MoE	Q4 / MXFP4	32K	2–4
32 GB	32B dense	Q4_K_M	32–64K	5–10
48 GB	32B at Q8, or 70B at Q4	Q8 / Q4	64K	8–15
64 GB	70B dense	Q4 / Q5	64–128K	15–25
96 GB	70B at FP8, 120B-class MoE	FP8 / MXFP4	128K+	25–40
128 GB	235B MoE, or 70B FP8 + large KV	FP8 / MXFP4	256K	60–100

Concurrency assumes bursty chat, not agents in a loop — agentic workloads consume 10–50× the tokens. We size against what the assessment measured, not headcount. Throughput is benchmarked on your build before any number goes into a contract.

Load calculation

When buying one is the wrong call.

A 30-person firm doing ordinary document Q&A generates about 63 million tokens a month. On a commercial API that is roughly $300. Against the Department tier, the payback period runs to decades.

~$300

Monthly API cost

30-person firm, chat-style use

~42%

Utilisation break-even

vs dedicated cloud GPU rental

We would rather lose the hardware sale than sell you a machine that sits idle. If your usage stays at chat volumes and your data can lawfully leave, an API subscription is the better buy — and the assessment will say so in writing.

Buy the box when

Your data genuinely cannot leave. A statutory trigger, a professional obligation, a residency covenant you already signed, or a file where residual foreign-law exposure isn’t defensible.

You’re running agents, not chat. Sustained agentic workloads consume 10–50× the tokens. At 30× the same firm’s API bill becomes the dominant cost and the arithmetic inverts.

You need a number you can budget. Capex plus a known monthly is approvable. An unbounded per-token bill that scales with staff enthusiasm is not.

Utilisation is genuinely high. Above roughly 42% sustained utilisation, owning beats renting dedicated cloud GPU. Below it, rent — and we’ll help you rent.

Installation notesSheet H-03

The part that sinks self-built deployments.

The machine is the easy bit. Where these projects fail is the room it goes in.

ElectricalThe Canadian Electrical Code limits a continuous load to 80% of the circuit rating. A standard 15 A / 120 V office outlet gives about 1,440 W usable, which the Department tier exceeds. That means a dedicated circuit — and above it, 208/240 V and an electrician.

HeatEvery watt in becomes a watt of heat: BTU/h = W × 3.412. The Department tier puts roughly 4,800 BTU/h into the room, continuously. A closet with no return air will cook it and throttle your throughput.

AltitudeCalgary sits at 1,045 m, so air is about 11% less dense and air cooling is correspondingly less effective — a detail most integrators miss. The upside is roughly seven months a year of viable free cooling.

NoiseA dual-GPU tower under a reception desk is a mistake you make once. Sound-dampened chassis and power limiting keep the smaller tiers office-liveable; the larger ones need a room with a door.

SecurityYou are claiming data never leaves the building. An unencrypted drive in an unlocked tower makes that claim false, and a privacy auditor will say so. Full-disk encryption and a locked space are part of the build.

Power protectionSize the UPS on watts, not VA, at roughly 1.25× steady-state load. The goal is a clean shutdown, not riding out an outage — a half-written model file is its own kind of downtime.

Next step

Get a specification with real prices.

The assessment produces an itemised bill of materials at supplier cost, with current quoted prices and honest lead times — plus the cloud comparison, so you can see both.

Book an assessment How we handle regulated data