VEN-04 · Inference cost on a napkin

Your venture delivers some unit of value — a summarized document, a matched candidate, a graded assignment, a triaged ticket. What does one of those actually cost you to produce, all in, at three different volumes? Find the volume where your gross margin breaks, and say what you would have to change: the price, the model, the promise, or the plan.

The napkin is the deliverable. If it takes twenty slides, you do not understand your own cost yet.

Teacher approves the venture concept and the unit of value in writing before Checkpoint 1. The student needs four things:

• A unit of value, defined precisely. Not "a user" and not "an API call" — the thing the customer is actually paying to receive. One summarized document. One matched candidate. One graded essay returned. Get this wrong and every number after it is wrong.
• A measured workload profile. Run the real task, or the closest honest proxy, enough times to know what it consumes: input and output size, how many calls one unit of value actually takes, and how often it has to be retried or redone. Measured, not estimated. This is the step students skip and the step that decides the answer.
• Dated, cited prices. Provider pricing pages, hosting costs, and any tool subscriptions, each captured with the date accessed. Prices in this industry move; a napkin without a date is a napkin without a shelf life.
• The costs that are not inference. Human review of model output, storage, support time, failed or unusable outputs, and the free tier you are giving away. For many honest ventures these exceed the model cost, and discovering that is a finding worth the whole brief.

Three volumes, set by the student and defended: a pilot (what one real customer would generate), a middle case, and a scale case. Do not pick round numbers because they are round.

Reference vocabulary (concepts — not financial advice):

• Your competition's own financial rubric, where one applies — DECA (https://www.deca.org/) or FBLA (https://www.fbla.org/) — so the packet fits the event the student is actually entering.
• Gross margin, contribution margin, and unit economics as your business course already defines them. Use the course's definitions, not a blog's, and say which you used.

Quoting a headline per-token price and stopping — that number is the beginning of the cost, not the end. Ignoring retries and unusable outputs, which is where the margin usually dies. Assuming cost scales linearly to the scale case with no rate limits, no volume pricing, and no support load. Forgetting the human in the loop the product actually promises. Confusing gross margin with profit. Treating today's price as permanent — prices fall, but so do the prices your competitors charge customers. Building the whole napkin on a model whose pricing page was never dated.

A one-page napkin, plus its evidence:

1. Unit-of-value statement — one sentence naming what the customer receives and pays for, and what "one" of it means.
2. Measured workload profile — the observed consumption per unit: calls, input and output size, retry and failure rate, and how many runs the measurement is based on. Say what proxy was used if the real task was not available.
3. Cost stack at three volumes — a table with a row per cost line (inference, human review, storage, support, failure waste, fixed tooling) and a column per volume, ending in fully-loaded cost per unit. Every price cited with its date.
4. Margin-break analysis — at your intended price, the volume or assumption where gross margin goes negative, and the two or three assumptions the answer is most sensitive to. State what you would change first.
5. Refusal log — what you will not claim: no projected profitability, no cost figure carried past the date of its source, no scale case presented as a forecast, no margin that depends on unpaid founder labor being free.

The napkin fits one page. The evidence behind it can be as long as it needs to be.

CT: separating a headline price from a fully-loaded cost. CR: choosing a unit of value that survives contact with a customer. CO: a peer rebuilds the cost stack from the student's own measurements and the two reconcile any gap. CM: a one-page napkin a skeptical judge reads in ninety seconds. CZ: who absorbs the cost when the margin breaks — the customer through a price rise, the worker through unpaid review, or the quality of the product.

A founder, small-business owner, CFO, controller, or accountant reviews the unit-of-value definition and the volume assumptions at Checkpoint 1 — before the measuring starts — and the margin-break analysis at Checkpoint 2. Standing instruction: reject any cost stack with no human-review line and no failure-waste line. School-supervised, both times.

R1: one unit of value, defined in a sentence. R2: every price dated and cited; workload measured with the run count stated. R3: comparator is the non-ML way of producing the same unit, costed at least roughly. R4: sensitivity named; the AI-specific risk of price and model change addressed. R5: the napkin is one page and survives ninety seconds of a judge. R6: refusal log names specific claims declined.

(a) The student multiplies a per-token price by a guessed token count, gets a number with four decimal places, and presents it with a confidence the measurement never earned. (b) The cost stack is real but contains only inference — no human review, no retries, no support — so the margin looks healthy right up until the product exists.

• Week 1–2: Venture concept and unit of value approved in writing; the three volumes chosen and defended; measurement plan agreed with the mentor.
• Week 4–5: Workload measured with run counts logged; cost stack v1 with every price dated; peer rebuild scheduled.
• Week 7–8: Margin-break and sensitivity analysis complete; napkin cut to one page; refusal log finished.

Lane A immediately (business or entrepreneurship capstone, or a DECA/FBLA financial event). Lane B through Entrepreneurship HQWBL only with the division's WBL coordinator owning the training plan and paperwork — the brief is the project, not the compliance record. No verified credit claim.