Companies built on AI models carry specific risks: dependence on model providers, inference costs that grow with usage, rights to training data and outputs that are hard to evaluate. Diligence tests each.
11 min read · Level 3
The question I find most clarifying with an AI company is what would remain if every competitor had access to exactly the same model tomorrow, because in practice they usually do. What remains - data, workflow, distribution, trust - is the business; the model is a supplier.
Four areas of risk for products built on AI models:
1. Dependence on model providers. If the product calls a third-party model, the provider controls price, availability, terms and capabilities - and may build a competing feature.
2. Inference cost. Each use of the model costs money. For a consumer product:
Cost per query 0.02 Queries per active user per day 30 Days per month 30 Monthly inference cost per user 0.02 x 30 x 30 = 18 Subscription price per month 20 Inference share of revenue 18 / 20 = 90%
Before any other cost, 90% of revenue is gone.
3. Data rights. Does the company have the right to use the data it trains or fine-tunes on? Do customer contracts allow their data to be used to improve the product?
4. Evaluation. How does the company know outputs are accurate and safe? Is there systematic testing, or only examples chosen for demos?
Because the features that make AI products impressive in a demo - broad capability from a powerful model - are often available to competitors too. The defensible parts are usually proprietary data, workflow integration and distribution.
Inference costs also make margins behave very differently from traditional software.
Educational, not personalised financial advice. Nothing in the Investor Academy is a recommendation to buy, sell or hold anything — what is right for you depends on your own circumstances and the rules where you live.
The effect of heavy users.
User group Share of users Queries per day Monthly cost per user Light 70% 5 3 Heavy 30% 80 48
0.02 x 80 x 30 = 48. At a flat 20 subscription, each heavy user loses 28 a month before other costs. Blended cost is 0.7 x 3 + 0.3 x 48 = 2.1 + 14.4 = 16.5, or 82.5% of revenue. The investor asks about usage limits, tiered pricing and plans to use cheaper models for simpler tasks.
Evaluating an AI product on demo quality without checking inference costs by user segment, dependence on model providers and rights to the data involved.
Identify which models the product relies on and the terms of use.
Model cost per user for light and heavy users against price.
Check rights to training data and customer data use in contracts.
Look for systematic testing rather than selected examples.
During diligence on any company whose product relies on AI models.
Do not assume dependence on a third-party model is disqualifying; many strong companies build on them. Judge what the company owns on top.
Venture investors and industry analysts have written that many AI-based companies report gross margins lower than traditional software companies because of model inference and compute costs, and that margins improve when companies optimise model use or reduce per-query costs.
AI products need unit economics built on real usage patterns, not software margin assumptions.
Create a free account to read the rest and apply it to your own investing.