The Margin Floor is our term for the structural gross-margin level an AI product settles toward once inference is treated as what it is — a cost of goods sold paid on every use, not a fixed piece of infrastructure. It is held down by a quiet mechanism: when the unit cost of intelligence falls, organisations reinvest the saving into more usage (deeper reasoning, bigger context, agentic loops) rather than banking it, so aggregate cost stays stubborn even as price per token collapses.
The point of naming it is that the floor is raisable — but only by design, on three levers: price to the cost driver (usage/outcome, not seats), gate features by margin (each use costs money), and drive COGS down (routing, caching, distillation, owning inference). It sits beside the conviction-gap — that one asks whether an idea is worth building; the Margin Floor asks whether you can afford to succeed with it. The depth treatment is The Unit Economics of an AI Product.