Inference costs sit in cost of revenue, not research and development. A framework for reading gross margin in AI-enabled software.
Traditional software gross margins reflect a cost base that barely moves with usage. AI-enabled software does not share that property: each additional interaction carries a measurable variable cost.
Where that cost is reported matters. Classified within research and development, it flatters gross margin and obscures unit economics. Classified within cost of revenue, where it belongs, it produces a figure an owner can actually underwrite.
We therefore read three numbers together: gross margin excluding inference, inference cost per active account, and the trend in usage per account. A business with rising usage and falling margin is not scaling; it is subsidising.
Model prices will continue to fall, which is often offered as the answer. It is not a thesis. A durable business should be viable at today's prices and improved, not rescued, by tomorrow's.