Metric guide · AI costs
Inference efficiency ratio
Short answer
The inference efficiency ratio compares revenue with the cost of running the AI models that serve customers, called inference. Divide revenue by inference cost for the same month. A ratio of 10 means each dollar of inference brings in ten dollars of revenue. Higher is better. It falls when usage grows faster than revenue.
The formula
Inference efficiency ratio = revenue ÷ product AI cost, for the same month
Unit: Ratio. Better when it goes up.
How to calculate it
- Take product AI cost for the month: the model and cloud AI costs of serving customers, not the tools your team uses.
- Take revenue for the same month.
- Divide revenue by product AI cost.
- Track it monthly next to gross margin, because both move when inference costs change.
Mistakes that distort it
Assistants and coding tools your staff use do not serve customers. Leave them out, or the ratio reads worse than it is.
Published targets divide AI revenue by inference cost. A ratio on all revenue reads higher when part of your revenue has no AI in it.
A new AI feature often costs more per request before caching and model choices settle. Read three months before acting on it.
A published reference point
The SaaS CFO suggests 10 to 1 or better is healthy for SaaS products with AI features, with a warning below 5 to 1, and 5 to 1 or better for AI-native products, with a warning below 3 to 1.
Basis. The SaaS CFO divides AI revenue by inference cost. Sondola divides all revenue by product AI, because books rarely split AI revenue out.
The SaaS CFO, How to calculate the inference efficiency ratio · published 12 May 2026, updated 28 August 2026 · checked 24 September 2026
Where Sondola reads it from
Sondola divides revenue by product AI for the same month, and shows the source and period of each reading. Each reading keeps where it came from, the period and the date it was recorded, and you can download all of it at any time.
See all 34 integrations →Questions
What counts as inference cost?
The cost of running models to answer your customers' requests: model API bills and AI services on your cloud bill that serve the product. Training runs, experiments and team tools are left out.
Why does Sondola use all revenue?
Because most books do not split revenue from AI features out from the rest. The page says so beside the ratio, so you can compare it with a published target on the right basis.
Related metrics
Written by the Sondola team. Benchmarks name their source, basis and date. Changes to how Sondola counts a number are on the corrections page.
Open the live demo and click any number to see its source, period and read date. No signup.