what it costs, itemized.
every number on this page is either read off a receipt, measured on the box, or labelled as an assumption. the assumptions are the interesting part — they're the ones that decide whether any of this is actually worth it.
no metered API has ever been billed on this fleet. the production model runs on the desk.
one real inference against the primary slot, same prompt the public benchmark repo uses.
derived from the measured throughput and a stated power draw. formula below.
the hardwarepartially itemized
only the line i can defend to the dollar from a receipt. the rest is in the open, not guessed.
| component | what it does | cost | sourced |
|---|---|---|---|
| ASRock Intel Arc Pro B60 | 24 GB card, ≈21 GB usable under Vulkan. serves the production model at the measured rate above. | ≈ $500 | receipt |
| OCuLink eGPU dock | one gen4 x4 cable to the host. four lanes are enough — this card is compute-bound, not bandwidth-bound. | — | sunk |
| Thermaltake Smart 600W | feeds the dock and card only. | — | sunk |
| GMKtec EVO-X3 | the host. runs llama.cpp, the memory brain, the agents and the dashboard. | — | sunk |
| Aoostar WTR MAX | NAS. every model file lives here and is pulled to the host on demand. | — | sunk |
| nuke-standby | separate box whose only job is to watch the fleet and heal it. | — | sunk |
| router + UPS | outbound gate for agents, and a clean shutdown window for the shelf. | — | sunk |
the GPU figure comes from the parts list on /rig. the rest are older purchases whose receipts are not in front of me — i would rather leave the cell empty than put a plausible number in it.
what the box actually produces
throughput is the only input that matters to the economics, and it moves. these figures are computed from the live measurement at the top of the page, so they move with it.
| window | tokens | formula |
|---|---|---|
| per hour | — | tok/s × 3,600 |
| per day | — | tok/s × 86,400 |
| per month (30d) | — | tok/s × 2,592,000 |
| time for 1M tokens | — | 1,000,000 ÷ tok/s |
the cloud comparison
list prices for the models people actually compare against, at the time of writing. these are the vendors' numbers, not mine — check the pricing page before you decide anything from them.
| model | input / 1M | output / 1M | tokens to repay a $500 card | days at measured throughput |
|---|---|---|---|---|
| GPT-4o openai | $2.50 | $10.00 | — | — |
| Claude Sonnet 4.5 anthropic | $3.00 | $15.00 | — | — |
| Claude Haiku 4.5 anthropic | $1.00 | $5.00 | — | — |
| GPT-4o mini openai | $0.15 | $0.60 | — | — |
the "days" column assumes the box runs flat out, 24 hours a day, forever. that is not how it is used — it is used in bursts, by agents, at odd hours. treat it as the floor on the payback period, not the expectation.
electricity, with the assumption stated
nobody believes a cost claim with no meter behind it. here is exactly what this one rests on.
Power draw. 150 W for the card at full inference load, plus 65 W for the host, router and NAS idling alongside — call it 215 W, or 0.215 kW. This is a design figure, not a measurement. i have not put a meter on this box yet, and until i do, every electricity number on this page is an estimate.
Rate. $0.12 per kWh, Georgia residential. Your rate will differ.
Formula. kW × rate = $/hour → 0.215 × $0.12 = $0.0258/hour → $0.62/day → $18.60/month at 100% duty cycle.
Per token. divide the hours-per-million-tokens by the hourly cost. at the measured rate that is — per million tokens.
what this does not include
- the rest of the stack. host, NAS, dock, PSU, router and UPS are real money and they are not in the ledger. the honest total is higher than $500.
- my time. building, tuning, breaking and rebuilding this is the largest cost by far, and it is not billable per token.
- duty cycle. the box is not at 100%. most hours it is idle, drawing a fraction of the estimate above.
- model weight storage. every quant of every model that did not make the cut is still on the NAS.
- the paid lanes. some traffic does route out through a gateway. that spend is tracked separately and is not zero — see /tokenronin.
what it does include
- a hard ceiling on the bill. the model is on the desk. no usage spike can produce a surprise invoice.
- no per-token metering. the marginal cost of one more request is electricity, and it is denominated in cents per million.
- no rate-limit anxiety. the constraint is the card, not a quota. when it is slow, it is slow because the hardware is busy, and you can see it.
- data never leaves. the reason this is worth doing at all. worth more than the electricity line, and not expressible in the same units.