ICONIQ's surveyed companies expected AI product gross margins to average around 52% in 2026. Classic SaaS runs at 80–90%.

Every time a user clicks a button in an AI product, you pay for it.

Investors know this, and gross margin has become one of the first numbers they pressure-test in an AI pitch.

Below you'll learn where your AI costs actually come from, which levers move your margin the most, and how to model all of it in a free spreadsheet.

Here is the full breakdown, built on one realistic example: a freemium AI product with 10,000 monthly active users, 5% of them paying $49 a month.

Inside you'll find:

For fifteen years, software had a simple promise: once the product was built, serving one more customer cost almost nothing. That is where the 80–90% gross margin benchmark came from.

AI breaks that promise. Every query runs a model, and every model call has a price. The cost of serving a user now scales with how much they use the product, not just with how many users you have.

The data reflects it. In ICONIQ's January 2026 State of AI snapshot, surveyed companies expected their AI product gross margins to average around 52% in 2026, compared with 41% in 2024 and 45% in 2025 reported by the same sample. Bessemer Venture Partners' February 2026 pricing playbook places AI companies at 50–60% against 80–90% for SaaS, and its fastest-scaling cohort as low as ~25%.

→ The takeaway is not that AI businesses are bad businesses. It is that gross margin is no longer something you inherit from your business model. It is something you engineer.

And you cannot engineer a number you have never modelled.

Most founders model AI costs per paying customer. That is the first mistake.

In our example, only 500 of the 10,000 users pay. Paid users are heavier (60 interactions a month vs 15 for free users), but there are nineteen free users for every paying one. The result: free users generate 142,500 of the 172,500 monthly interactions.

The most telling detail is the last bar. Because model and tool costs scale with interactions, roughly 83% of the AI and tools bill (about $7,700 of $9,355 a month) is spent on users who pay nothing.

→ Your free tier is not a marketing cost. It is part of your COGS, and every paid user has to carry it.

This is why the calculator separates free and paid usage. Cap free usage at 10 interactions a month instead of 15, and gross margin jumps from 57.9% to about 68.5%. Raising the paid price from $49 to $55 only gets you to 62.5%.

A user sees one answer. Your bill sees several model calls.

In the example, 40% of interactions are agentic: multi-step runs that make four model calls on average. Add a 10% overhead for retries and failed loops, and every interaction costs 2.42 model calls on average. Each call carries around 3,000 input tokens (prompt, context and retrieved documents) and 600 output tokens.

These are the five numbers that decide your margin:

With these inputs and GPT-5.6 Sol at list price, the product lands at 57.9% gross margin: $24,500 in monthly revenue against $10,305 in COGS. That is right in the Bessemer range, and 12 points below the product's own 70% target.

→ The calculator also returns the minimum price that hits your target at this usage: $68.70 per paid user, not $49.