A customer who uses a subscription product daily may be less likely to cancel and more likely to recommend it. With a service that generates a fresh answer for each request, heavier use can nevertheless reduce that customer’s contribution in the current month. Consider an application company that buys model access from another provider.

What goes into one answer?

Imagine a contract-review application. A customer types “What are the payment terms?” The app also sends sections of the agreement and supporting documents to the model. The billable input can be much longer than the visible question. Many text APIs split text into tokens and have different rates for input and output.

At the July 2025 launch of Gemini 2.5 Flash-Lite, Google announced prices of $0.10 per million input tokens and $0.40 per million output tokens. Using that historical tariff makes our example reproducible. A present-day budget would use the current tariff for the chosen model and contract.

Google’s launch pricing announcement, July 22, 2025

Two customers on the same $10 subscription

Suppose one job uses 40,000 input tokens and 5,000 output tokens. Ignoring additional tools, cache discounts and retries, the API bill per job is 0.04 × $0.10 + 0.005 × $0.40 = $0.006. Now suppose the application promises unlimited use for $10 a month.

Customer A runs 20 jobs. Revenue is $10, API expense is $0.12 and $9.88 remains. Customer B runs 2,000 jobs. Revenue is still $10, but API expense reaches $12, leaving a $2 deficit. Under these assumptions the API bill first exceeds the subscription fee at job 1,667. The example describes a hypothetical application and reveals no actual model provider’s customer margins.

The amount left is revenue less the external API bill alone. It must also cover storage, search, customer support and other operations. Google books the API fee as revenue. The application company treats it as a purchased input. The tariff tells us nothing conclusive about Google’s own electricity and hardware cost for that call.

Would usage billing solve the problem?

Conventional SaaS also incurs server and database expenses. For an AI-enabled product, the number of model calls, choice of model, length of the documents and output size can change the cost of serving a specific customer. A business can include usage up to a threshold, route simple jobs to a smaller model or avoid processing the same document repeatedly.

Pure usage billing aligns receipts more closely with activity, but buyers may struggle to forecast next month’s bill. Look beyond average customer contribution. Isolate the most active customers and trace which processing steps raise their cost and what they actually pay for those jobs.

Infrastructure investment is a different calculation

A provider that buys data-centre equipment spends cash on acquisition while recognizing depreciation over the equipment’s useful life. Allocating infrastructure and development costs is a different question from the cost of serving one more request. Microsoft’s September 2026 quarterly filing reported a Microsoft Cloud gross margin of 66%. Product-level margin improvements were more than offset by a sales mix shift toward Azure, bringing down the blended margin year over year. Match businesses and cost definitions when moving from an individual customer calculation to a company-wide percentage.

Microsoft’s quarterly report for September 2026 (Form 10-Q)