“AI company” is too broad to describe an economic model. A model developer, a workflow application and a compute provider can sell to each other while facing different capital needs and customer risks. The three-layer comparison below is an analytical framework for reading a business, not a claim about EXAM’s holdings or a ranking of investment opportunities.
Follow the customer and the cost
Table scrolls horizontally
| Layer | Possible revenue | Question to test |
|---|---|---|
| Foundation model | Usage-based API charges or licences | Can customer revenue support training, serving and development? |
| AI application | Subscription, per-task or service revenue | Will users pay after novelty fades, including support costs? |
| Compute infrastructure | Capacity rental and managed operations | Does paid utilisation cover equipment, power and renewal? |
Test one complete customer journey
For an AI document-review application, follow one paid account from onboarding to completed, accepted work. Include data preparation, model calls, retries, human review and customer support. If these costs sit in different departments, a single API-cost figure can hide the economics of the service.
Evidence that improves a discussion
- Revenue by customer cohort, renewals, concentration, variable costs and cash collection are more informative together than a headline user count.
- Ask which dependencies could change the margin: model pricing, cloud capacity, data permissions or a customer doing the work internally. Record assumptions instead of treating growth as inevitable.
Further reading
General information only; not investment advice, an offer or a recommendation. Any participation is subject to eligibility, documentation and applicable permissions. Capital and returns are not guaranteed; some or all capital may be lost.