Where product architecture becomes finance
AI Startup Unit Economics
AI unit economics are not simply SaaS metrics with an API bill underneath them. The product's architecture, model routing, token usage, caching, free-tier design, pricing unit, and customer behavior can all change contribution margin.
By Serge Mochtchenkov, CFA · Fractional CFO for AI startups
Choose the economic unit
The right unit may be a customer, seat, query, document, minute, generated asset, agent run, workflow, or completed outcome. The purpose is to find a unit that connects customer value to variable cost and can be measured consistently.
Build the variable cost stack
At minimum, consider model/API usage, embeddings, vector database or retrieval calls, third-party APIs, cloud compute directly tied to delivery, payment processing where relevant, and other usage-driven infrastructure. Separate these costs from product R&D and general engineering so gross margin does not become a disguised headcount metric.
Want this applied to your numbers rather than described in the abstract?
Pricing-unit mismatch
If the customer pays per seat but cost is driven by tokens, the company is implicitly underwriting usage. That can be attractive if usage drives retention and expansion, but the risk should be modeled. Hybrid approaches — subscription plus included usage, usage tiers, credits, minimums or outcome pricing — can better align revenue with cost depending on the product.
Contribution margin by cohort or segment
Enterprise customers, power users and free users may have very different usage and support profiles. Segmenting contribution margin can reveal customers who look attractive on ARR but consume disproportionate model or infrastructure cost.
The margin improvement roadmap
AI margin can be engineered. Model routing, distillation, caching, prompt/context optimization, batch processing, provider negotiation, reserved capacity, and product design can all change unit cost. The finance model should show the expected timing and cash impact of those technical changes.
Frequently asked questions
- What is the key unit-economics metric for an AI startup?
- There is no universal single metric. Contribution margin per relevant economic unit, gross margin after variable AI delivery cost, retention/expansion and CAC/payback are usually part of the picture.
- Should inference cost be in COGS?
- When inference is directly required to deliver the customer service, it is commonly treated as a delivery cost; exact accounting classification should follow the company's facts and accounting policy.
- How do free users affect unit economics?
- They may generate variable compute cost without revenue, so free-tier usage should be modeled explicitly rather than averaged into paying customers.
Next step
Build the financial system behind the next decision. Book a 30-minute intro call to discuss your stage, model, runway, and next financing milestone.
Related reading