Cloud API only
$697K
$19,372 a month · no machines
Economics
What the work costs on machines you already own
Cloud sells inference by the token. Hardware sells it by the node-hour. This page converts one into the other, sizes the fleet a workload actually demands, and shows what changes when those node-hours come from computers that are already on your balance sheet.
Every dollar figure on this page prices somebody else’s product — a cloud API’s published per-token rate, and the hardware you would buy. None of them price PoolParty, and none of them are an observed PoolParty benchmark.
The finding
This workload asks for 137 node-hours a day. Pooled, that is 9 machines — not 25.
Sized to the work rather than to the headcount, and kept available for 20 hours instead of 8, the same tokens land on 64% fewer machines.
One machine each
$105K
$135 a month · 25 machines
A fixed 50/50 split
$451K
$9,754 a month · 25 machines
Pooled
$41K
$135 a month · 9 machines
Cash out, cumulative
Hardware lands in month nought; metered tokens never stop arriving.
Decomposition
Where the difference comes from
Starting from one machine per person and changing a single thing at a time. Watch the electricity: it barely moves, because the work is identical either way. What changes is how many machines have to exist for that work to fit.
The unflattering part
What this calculator does not price
A savings model that only counts what favours it is a brochure. These are the costs and risks on the other side of the ledger, and some of them are large.
A smaller model is a smaller model
The comparison assumes the local model is adequate for the work. Where it is not, the honest cost is the wrong answer, the retry, or the frontier call you had to make anyway. Route by what the job needs; this page cannot judge that for you.
Twenty available hours is an assumption, not a fact
It needs machines that tolerate sustained duty, thermal and power headroom, a power policy that permits it, and owners who actually opt in. Set it to eight and the pooled column still wins, by less.
Somebody has to run it
Migration, model management, monitoring and the hours spent on all three are not in any column. Neither is the cloud’s real advantage: there is nothing to run.
Peaks are lumpier than an average
Demand is modelled as a daily total with a flat allowance for peaks. Real teams arrive at nine and leave at five. Raise the spare-capacity figure until it reflects your own shape before you trust the fleet size.
Show your work
Every assumption, and where it came from
The same three states the rest of this site uses. Nothing below is an observed PoolParty result, because there are none to show yet.
| Assumption | Where it came from | State |
|---|---|---|
| Workload sizes | The light, medium and heavy definitions published in AMD’s AI Tokenomics Calculator — 5.74M input and 574K output tokens per person per day at medium. | Third party |
| $3 and $15 per 1M tokens | Back-solved from that calculator’s own output, and consistent with current published mid-tier frontier list pricing. Overwrite it with the rate you actually pay. | Third party |
| Calibration | At the defaults above this model returns the same cloud-only and one-machine-each totals AMD publishes, and the same monthly power figure. The fixed-split lane lands about half a percent apart, because electricity here follows the work actually served rather than being held flat. | Third party |
| $4,006 a machine, 275 W active | Both back-solved from the same published totals, so the hardware side is priced at a third party’s numbers rather than at ours. Applied identically to every column. | Third party |
| 700 prefill, 50 decode tokens/sec | Order-of-magnitude for a small-active-parameter mixture-of-experts model at four-bit quantisation on a current client part. It is an input, not a measurement, and the fleet ratio survives it being wrong in either direction. | Modelled |
| 30% spare capacity | A flat allowance so a pooled fleet is not sized exactly to its own average. Your peaks are your own. | Modelled |
| 20 hours a day available | The load-bearing assumption of the whole page, and the one most likely to be wrong in a given building. It is a control above for exactly that reason. | Unproven |
| Credits | Defaults to zero. No earning rate is published, and credits are redeemable only when someone is online offering compute. Credits are not money. | Unproven |
| Excluded everywhere | Migration effort, operations, networking and storage, model-quality differences, and the accuracy cost of running a smaller model on work that needed a larger one. | Excluded |
The machines are already in the building
Pooling them is the part PoolParty does: join your computers, send jobs to the ones with the hardware to run them, and reach your models through one endpoint. Free on your own hardware.
See how computers connect PoolParty for companies Back to the FAQ