Every AI startup I meet is arbitraging investor money against compute bills. What does one generation actually cost you, and do your unit economics survive without subsidy?

The full executive answer

This is the question we built instrumentation for, so let me answer with the mechanism rather than adjectives. Every single AI generation writes a row to an append-only cost ledger — provider, model, input tokens, output tokens, cache tokens, and cost in dollars to six decimal places. Our engineering rate card runs from roughly eighty cents per million input tokens on the efficiency model to fifteen dollars per million on the premium model, which puts a typical governed board analysis in the cents-to-tens-of-cents range, even including the dual-provider consensus pass on flagship modules. Response caching means repeated generations cost literally zero — cache hits are logged at $0.00 with the savings estimated — and prompt-cache tokens are tracked separately because they bill differently.

Now the ratio that answers your question. Each tier carries a designed AI-usage allowance: $12 of compute against the $149 Basic price, $45 against $649 Pro, $150 against $1,799 Max. That is a worst-case cost-of-goods ceiling of roughly nine to ten percent of revenue — a floor of about 90% gross margin on AI costs even if every customer exhausts their allowance, before caching does its work. And this is not a spreadsheet assumption: the platform contains a margin engine that computes recognised revenue minus actual metered AI cost per organisation per period, with a board-readable banding where 80%-plus is classified "excellent." Our honest label for this metric is AI-attributable gross margin — it deliberately excludes general infrastructure and support overhead, so I am not smuggling a fully-loaded SaaS margin claim past you.

Roadmap and risk, candidly: model prices have historically fallen, which helps us, but we do not depend on that — allowances, per-module caps, budget alerts and the enterprise floor of $9,000 exist precisely so that no single customer can invert the economics. In Rule of 40 language: the margin structure is engineered and continuously measured from day one; the growth half of the equation is what launch has to prove.