Why Your AI SaaS Has 52% Gross Margins Instead of 80%: The LLM Cost Problem No One Budgets For
The Margin Gap Is Structural
In traditional SaaS, COGS is mostly hosting and support — costs that grow sublinearly with usage. In AI SaaS, the core deliverable is a metered, per-token cost that scales close to linearly with how much customers actually use the product. A power user isn't just more valuable in traditional SaaS economics — in AI SaaS, a power user can be a margin problem if pricing didn't account for the usage tail.
This is why median AI SaaS margin improving from 41% to 52% over two years is good news but not resolution. The gap to traditional SaaS margins is closing slowly through model price drops and smarter routing — not disappearing.
The B2B Copilot Math
A B2B copilot priced at $15/user/month — a normal SaaS price point chosen by comparing to competitor pricing, not by modeling LLM cost. In month one, light usage keeps cost low and margin looks fine. By month three, as users adopt the feature into daily workflows, actual LLM cost climbs to $11.50/user — a 23% gross margin on a product priced as if it carried SaaS-typical economics.
Interactive: Margin Calculator
Interactive: Top-Decile Usage Distribution
Click a segment to see detail.
Three Levers to Pull
- Model routing. Route simple, high-volume requests to cheaper/smaller models, reserve frontier models for tasks that genuinely need it.
- Free tier usage caps. Free or trial tiers without hard caps are where margin leaks first — there's no price signal pushing toward efficiency.
- Usage-based pricing. For features with wide usage variance, metering aligns price with the actual cost driver instead of subsidizing heavy users with a flat price.
What Investors Look For
Investors now routinely ask for gross margin broken out by AI feature, not just blended company-wide gross margin — a blended number can hide an unprofitable AI feature behind a profitable core product. They also ask how margin trends as usage scales: a margin that holds steady or improves with scale signals real unit economics; a margin that erodes as adoption deepens signals a pricing problem that worsens with growth.
// Compute per-user LLM cost from real tracked usage, not estimates
const monthlyLLMCost = await llmtrack.query({
feature: 'copilot-chat',
groupBy: 'user_id',
metric: 'total_cost',
period: 'last_30_days'
})
// Compare against plan price to get a real, current gross margin
LLMtrack's Business plan adds per-user and per-feature cost breakdowns.
FAQ
Hosting, vector database/storage, third-party APIs beyond the LLM call itself, and any human-in-the-loop review costs. LLM token cost is usually the largest line item but rarely the only one.
Not necessarily immediately, but it should trend that direction as model prices fall and routing improves. A feature stuck below 30% margin with no improving trend is a pricing or architecture problem.
Track cost per user continuously and alert on users whose monthly cost crosses a threshold relative to their plan price — the same way you'd alert on any other anomalous spend.
Know your gross margin per feature, per user, in real time
Start free. One async tracking call. No proxy and no credit card required.
Start free →