Reselling inference at cost is a zero-margin business. The fastest-growing AI companies are built on top of inference, but the pricing model determines whether they are software companies or payment processors. Three mechanics exist: cost-plus markup, value-based pricing, and cost optimization. Only one of them is durable.
Cost-plus caps willingness to pay at the inference line. As inference commoditizes, the 30-point markup compresses toward zero. Value-based pricing breaks the link entirely. Sierra charges only when an agent resolves a ticket, zero for failures. Devin sells Agent Compute Units, not tokens, borrowing the same abstraction Databricks and Snowflake use with credits to hide raw compute from the customer. On the cost side, distilling production traffic through frontier teacher models into a sub-8b parameter student deployed on cheap hardware produces a proprietary model competitors cannot replicate. Routing and caching are copyable. Distillation buys time. The piece also works through a specific stress test: what happens when the customer brings their own API key. Cost-plus collapses. Value-based and optimization both survive, because you are selling work and platform, not tokens.
The original is worth reading for the chart alone, which maps cost-plus markup and optimization against the inference cost line and shows visually why value-based pricing does not appear on it at all. The BYOK scenario and the Devin credit abstraction argument are the sections that will change how you think about your own pricing. Read it to find out which company you are actually building.
[READ ORIGINAL →]