How do you calculate unit economics for an early-stage startup?

Pick the unit, one customer, one seat, one transaction, one tier member, and write what it costs and what it returns over a period you can defend. The economics of the core mechanic decide whether the product can exist. In the case the mechanic was loyalty tiers, and the economics turned on a single assumption about what a tier had to give away to be wanted.

What is the unit in unit economics and how do you choose it?

Pick the unit, one customer, one seat, one transaction, one tier member, and write what it costs and what it returns over a period you can defend. The economics of the core mechanic decide whether the product can exist. In the case the mechanic was loyalty tiers, and the economics turned on a single assumption about what a tier had to give away to be wanted.

How do AI inference costs affect startup unit economics?

Pick the unit, one customer, one seat, one transaction, one tier member, and write what it costs and what it returns over a period you can defend. The economics of the core mechanic decide whether the product can exist. In the case the mechanic was loyalty tiers, and the economics turned on a single assumption about what a tier had to give away to be wanted.

Unit economics · Business Design

§The hypotheses: one unit of the core mechanic returns more than it costs over a period we can defend, and we can name the two or three assumptions on which that arithmetic turns.

§The canvas said where money enters and leaves. asks whether, for one unit, more enters than leaves. It is the least glamorous page in business design and the one that decides whether the others matter.

§ 8.1Choose the unit#

§The unit is the thing the business does once and hopes to do many times. One customer, one seat, one transaction, one order, one active member of a tier. Choosing it is a design decision: the unit should be the one whose economics, if they work, make the whole model work, and whose economics, if they fail, nothing else can rescue.

§For a subscription , the unit is a paying account, and David Skok's SaaS metrics apply: what it costs to acquire, what it returns per month, how long it stays, and the ratio of lifetime value to acquisition cost. For a marketplace, a transaction. For the hospitality group, the unit was a member of a loyalty tier, because the tiers were the mechanic everything else hung on.

§ 8.2Cost, return, period#

§Three lines per unit. What it costs to bring one in. What it costs to serve one for the period. What one returns in the period. The period is the one you can defend with : a year for a subscription with retention data, a quarter for a new product with none, a single visit for a mechanic that has never been observed twice.

§McGrath's reverse income statement is the discipline for a first version with no data: start from the profit the unit must return for the business to make sense, and work backwards to what each line must be, so that the lines become to test rather than forecasts to hope for.

§ 8.3The AI line#

§A product that calls a model has a cost per unit that most founders underestimate twice: once because inference prices fall and they assume the fall will rescue the arithmetic, and again because usage per user rises faster than prices fall. Casado and Bornstein's 2020 warning about AI gross margins has held: the marginal cost of serving is not zero, it scales with engagement, and the most engaged users are the most expensive. The unit economics sheet has a line for it, with the model, the calls per unit per period, and the price assumed, so that when the model or the price changes, the line changes and the arithmetic is redone rather than remembered.

§ 8.4The case: tiers#

Case

The hospitality group's mechanic was tiers: a guest's status across brands, earned by visits. The unit was a tier member, and the sheet had two versions. The first, from before the field, assumed that a tier had to give something away to be wanted: points, discounts, a token economy that converted visits into value the guest could spend. Its cost per member was the giveaway, and the return was the incremental visits the giveaway was assumed to produce. The arithmetic was tight and depended entirely on the incremental-visit assumption. The second version, after the field, dropped the giveaway. Six of six customers had filled in their profiles for nothing, because being recognized was the reward. The cost per member fell to the cost of recognizing them, a glance and a data record, and the return was the same incremental visits, now resting on a mechanism that had been observed rather than assumed. The tier economics did not improve because the team found more revenue. They improved because a cost the team had invented turned out not to exist.

§ 8.5What the sheet is for#

§It is not a forecast. It is a small set of arithmetic relationships with two or three assumptions exposed, so that the team knows exactly which numbers, if wrong, break the business. In the case those were: the incremental visits a recognized guest makes, the share of guests who complete a profile, and the cost of a server's glance in seconds per cover. Three numbers, each a hypothesis, each on the list with an owner.

§ 8.6What has to be true#

§Added to the running list: the unit, named. Its cost to acquire, cost to serve per period, and return per period, each as a hypothesis with a number and the evidence or assumption behind it. The two or three assumptions on which the arithmetic turns, with owners. The AI cost line, with model, calls and price stated so it can be redone.

8. Unit economics