How Much of Your Margin Belongs to Your Model Provider?

A one-minute check for founders building products on someone else’s model.

Picture the end of the month. You have closed the revenue, and then you open the invoice from your model provider.

That line in the P&L used to sit somewhere between hosting and office coffee. In a product where AI is the core, it can be one of the largest. Before reading on, try to answer without a spreadsheet: how much of your margin belongs to your model provider? If the answer comes out as “roughly” or “it depends,” the rest of this article is for you.

What the benchmarks say

Classic SaaS earns almost everything on every new user. Gross margins of 75–85% are the long-standing norm. AI products look different. ICONIQ’s survey of roughly 300 software executives puts the average gross margin of AI products at about 52% in 2026, up from 45% in 2025 and 41% in 2024. A later ICONIQ report projects 53% for 2026 and 59% for 2027. Inference, the cost of running the model for each request, takes much of the difference.

One caution keeps this in proportion. The compression shows up in products where AI is the product. Benchmarkit’s 2026 report finds the median software gross margin still above 80% and stable across four years, and concludes that AI infrastructure costs have not yet compressed margins at the median. If AI is a feature in a mature product, your numbers will look different. If AI does the work customers pay for, treat the ICONIQ figure as your starting point.

Why your success works against you

In traditional SaaS, one more user costs cents. In an AI product, every request costs real money, and the revenue per seat stays fixed. A customer who falls in love with the product sends more requests for the same subscription, so your best customer becomes your most expensive one.

A fixed subscription on a product where every request costs money is like an all-you-can-eat buffet where the regulars eat the most. The restaurant survives only if it knows what the regulars cost.

The player you didn’t choose

There is one more participant in this equation: the model provider. They can change tariffs or limits, and your cost of service changes tomorrow. Your own prices move more slowly, because you will be revising them with customers for months.

A margin model built on today’s price per request is therefore incomplete. Add one scenario to it: what happens to your gross margin if your provider raises prices by 30%? The number does not have to be a forecast. It shows how much room you have.

What improves the margin

The encouraging part of the data is that this is treatable. In ICONIQ’s survey, two-thirds of companies reported better per-query unit economics, crediting inference cost management, smarter model routing, and revenue growth that creates cost leverage.

Three habits sit behind that result.

Measure the cost of a request by feature. A product-level average hides the two features that cost the most.

Route each request to a model that fits it. Not every task needs the most capable and most expensive model.

Keep the model behind an interface in your architecture. Switching or mixing models then becomes a configuration decision and does not require a rewrite.

No magic: accounting and architecture.

Pricing is part of the fix

The second lever is the price. Companies whose revenue grows together with usage do not suffer from the popularity of their own product. ICONIQ’s data shows that 37% of AI companies plan to change their pricing model within a year.

Usage-based pricing has its own trade-off: customers dislike unpredictable bills. A common compromise is a base subscription with usage tiers or credits, so that cost still follows usage while the customer sees a ceiling.

A one-minute check

Answer three questions:

1. How much does one request cost?
2. How much does the customer pay for it?
3. What happens to your margin if your provider raises prices by 30% tomorrow?

If any of the answers is missing, it is better to find it yourself, before the conversation with an investor.

So, how much of your revenue goes to your model provider today?

Wamisoftware has been building software products since 2014. Our senior engineers work with startups and enterprise clients on AI products where architecture decisions and unit economics are settled early.