Skip to content
Multigrid

Total Cost of Ownership Calculator

Inference, storage, egress, platform, evaluation and engineering time in one monthly figure, with the human share stated as a percentage of the total.

Total cost of ownership, per month
$18,091

Human time is 71.4% of it — $12,925 a month. Inference is 19.5%. Everything else is 9.1%.

Requests billed, including retries
309,000
Input tokens per month
556.20M
Output tokens per month
123.60M
Inference
$3,523
Storage
$125.00
Egress
$18.00
Platform and tooling
$1,200
Evaluation traffic
$300.00
Build, amortised over 24 months
$4,125
Maintenance and change
$6,600
On-call and incidents
$2,200
Human time, total
$12,925
Total per month
$18,091
Cost per request
$0.06
Annual run rate
$217,087
What this assumes: that the loaded hourly rate includes employer costs rather than being a salary divided by working hours — the honest figure is usually between 1.3 and 1.8 times the naive one, and which multiplier applies is your finance team’s answer, not ours. Amortising the build spreads a sunk cost so that a monthly figure can be compared against alternatives; it is not cash leaving the business this month. Retries are billed at full price including the tokens the failed attempt generated. Storage and egress are priced by your cloud, not by the model provider. Nothing here prices the cost of being wrong — the support ticket a bad answer creates, the rollback, the security review — and at most companies that is a larger number than the egress line.

The term everyone leaves out

Almost every AI cost model in circulation is a token calculator with a storage line bolted on, and almost every one of them is wrong for the same reason: the largest recurring cost of an AI feature is usually the engineering time it consumes, not the inference it buys. Sixty hours a month of maintenance at a realistic loaded rate is several thousand dollars, every month, forever — and it does not fall when a provider cuts prices, does not fall when you shorten your prompts, and does not appear on any invoice, which is precisely why it goes unmodelled.

The maintenance line is also the one people under-fill. AI features decay in ways ordinary software does not. Models are deprecated on the provider’s schedule rather than yours. Behaviour shifts under a silent update and the eval suite has to be re-run and re-baselined. Prompts accumulate special cases. Every one of those is engineer-hours against a feature that is otherwise finished, and the number is rarely under a day a month for anything in production.

What the per-request figure is for

Divide the whole thing by requests and you get the number that actually decides your pricing and your margin — and it is typically a large multiple of the token cost of a request. If you have priced a feature off the token cost alone, this is the page that explains where the margin went. It is also the page that tells you when scale fixes it: the human terms are broadly fixed, so cost per request falls steeply with volume up to the point where volume itself creates work.

The comparison worth running is this figure against the alternative you did not choose. A vendor at a per-seat price that looks expensive next to raw token cost frequently looks cheap next to the total on this page, because what you are buying from them is the maintenance line.

The reasoning behind this

Related tools

Total Cost of Ownership Calculator · Multigrid