What happens to software economics when serving a customer costs money. Where inference belongs in the accounts, which levers actually reduce it, and why your most engaged customers become your least profitable unless somebody notices.
| # | Lesson | The question it answers |
|---|---|---|
| 01 | Cost Per Task | How do you compute it for a real feature, including everything around the call? |
| 02 | Tokens Per Task and Context Efficiency | Where does the context budget go, and what is waste? |
| 03 | Inference in Cost of Revenue | What does it do to gross margin, and who books it today? |
| 04 | Routing: Cheap Model First | How much does it save, and what does it cost in quality? |
| 05 | Caching and Repeated Prefixes | What is actually cacheable, and what does it buy? |
| 06 | Batching and Asynchronous Work | When is cheaper worth slower? |
| 07 | Margin Under Engagement Growth | What happens when your best cohort is also your most expensive? |
| 08 | The Cost of Quality | What do evals, judge calls and human review cost per task? |
| 09 | Forecasting Spend You Do Not Control | How do you budget for usage you cannot predict? |
| 10 | Falling Token Prices | Prices drop every year. Who captures that, and how do you plan for it? |
| 11 | Build Versus Buy, Financially | Self hosting, fine tuning and frontier APIs, compared honestly |
| 12 | Worked: A Margin Model | One feature, from token count to gross margin, end to end |