Analysis · Cost of AI

The AI gross-margin problem: why your software margins are shrinking

Classic software had a near-zero marginal cost: once built, serving one more customer cost almost nothing, which is why SaaS gross margins sat at 70 to 85 percent. AI breaks that. Every query costs compute, so an AI feature carries a real cost of goods that scales with use. Investors and CFOs are now watching software gross margins drift down toward the 50 to 60 percent range, with inference alone reported at roughly a fifth of revenue at scaling AI companies. The 80 percent margin is becoming the exception. Defending margin now depends on knowing the unit economics of every AI feature.

The margin reset
70-85%
gross margin that defined classic SaaS, on near-zero marginal cost.
a16z, industry benchmarks
~52%
reported average gross margin of AI-first companies, as compute COGS scales with use.
ICONIQ via secondary, 2026
~23%
share of revenue consumed by inference alone at scaling AI companies, before other COGS.
ICONIQ via Monetizely, 2026

The ~52% and ~23% figures come from secondary reporting of an ICONIQ dataset; confirm against the primary before quoting in a client deliverable. The structural direction, that AI adds real COGS and compresses margin, is well established.

Why the margin falls

The cause is simple: AI features have a cost of goods that software did not. Each use spends tokens, and reasoning and agentic features spend many. That cost sits in COGS and pulls gross margin down. The effect is uneven, because cheaper-per-token does not mean cheaper-per-feature: as models get cheaper, products use them more, so the cost of goods can hold or rise even as token prices fall. The result is a structural reset, with analysts projecting a new floor well below the old SaaS benchmark unless a company runs an unusually disciplined inference stack.

How to defend it

Margin defence starts with unit economics: the cost to serve one use of each AI feature, set against the price or value it earns. With that number you can do three things. Reprice features whose flat plan loses money on heavy users, including a shift to usage- or outcome-based pricing. Re-engineer the cost to serve by routing cheap requests to small models and caching context. And cut the features whose cost to serve will never be covered. None of this is possible from a blended margin; it requires costing AI at the level of a single outcome, which is what activity-based costing was built to do.

THE MARGIN RESET, VISUALISED

Illustrative. Classic SaaS gross margin sat at 70 to 85 percent on near-zero marginal cost. AI-first companies run nearer 52 percent, with inference alone consuming roughly 23 percent of revenue. Proportions are illustrative.

The 80 percent software margin was built on a marginal cost of zero. AI put a meter on every use, and the meter is now in your cost of goods.

Common questions

Why are AI gross margins lower than SaaS?
Because AI features carry a real cost of goods that classic software did not. Every query spends compute, so the more the feature is used, the more it costs to serve. Classic SaaS had a near-zero marginal cost and ran at 70 to 85 percent gross margin; AI-first companies are reported nearer 50 to 60 percent, with inference alone consuming roughly a fifth of revenue at scale.
Is the 80 percent SaaS gross margin over?
For software with heavy AI features, analysts expect a structurally lower floor, often cited in the 50 to 70 percent range, because inference cost now sits in cost of goods. Companies with light AI use or a very disciplined inference stack can stay higher. The point is that 80 percent is no longer the default once AI is embedded.
How do we protect gross margin on AI features?
By managing unit economics. Measure the cost to serve one use of each feature, compare it to the price or value it earns, then reprice the features that lose money on heavy users, re-engineer cost to serve by routing and caching, and retire the features that will never cover their cost. This requires costing AI at the level of one outcome, not a blended average.
Does cheaper AI fix the margin problem?
Not on its own. The price per token has fallen sharply, but products respond by using AI more and reaching for token-hungry reasoning and agents, so the cost of goods often holds or rises even as token prices fall. Margin is protected by managing how much AI a feature consumes per unit of value, not by waiting for tokens to get cheaper.

Defend your margin with real AI unit economics.

The Profit Check shows where margin is leaking, AI included, in five minutes.

Take the Profit Check
M
Ask us anything
usually replies in minutes
Hi. I can answer the quick questions about cost, method and timing right here. For anything specific to your business, I'll hand you to Miguel on WhatsApp.
Free. No bot loops. Straight to a person.