AI product gross margin: the metric that decides who survives the reckoning.

By COGScontrol Team · October 31, 2026

AI product gross margin is the share of revenue from an AI-powered product left after paying the direct, variable cost of serving it — model inference, cloud infrastructure, retrieval, and supporting platforms. Measured per unit the business already reports — a daily active user, a funded loan, a transaction — it tells product and revenue owners whether each AI feature earns more than it costs to run.

For two years, AI revenue growth was enough. It no longer is. Summarizing the third-quarter earnings season, Saxo’s research team wrote that investors are shifting from growth to profitability and ROI, because "higher AI revenues do not automatically translate into stronger profits." For anyone who owns customer-facing AI, that shift lands on one line of the P&L: gross margin. (For the board-level view across customer-facing and internal AI, see the AI bubble is a measurement problem.)

Why are AI gross margins lower than SaaS margins?

Because every AI interaction has a variable cost, and SaaS interactions mostly don’t. Serving one more user of traditional software costs close to nothing; serving one more AI request costs tokens and compute every time. ICONIQ’s 2026 State of AI report put AI product gross margins at 41% in 2024 and 45% in 2025, with 52% projected for 2026 — and application-layer companies lower still, at a projected 45%. ICONIQ partner Vivian Guo: “Companies are really starting to focus more on the gross margin impact of using AI.”

That gap is now a valuation question. As venture investor Mala Ramakrishnan put it in an October Forbes column, "A company earning 40 cents on the dollar and one earning closer to 70 are simply not the same business, even at identical revenue." What counts toward that cost base — and what doesn’t — is defined in what is AI COGS.

What is squeezing AI product margins in 2026?

Public software companies spent 2026 naming AI as a margin headwind on their earnings calls:

• GoDaddy told analysts its Airo AI product carries inference costs that will weigh on the business: "Yes, the gross margin is going to see a little deterioration as we move forward," said CFO Mark McCaffrey.
• Figma’s CFO Praveer Melwani described beta AI features where "we bear the cost of inference without offsetting consumption revenue" — with gross margin improving in the first full quarter of AI credit monetization.
• monday.com’s CFO Eliran Glazer, explaining a one-point dip in non-GAAP gross margin to 89%, was blunt: "There is a lot of cost related to AI."

The response is repricing: moving AI from bundled to metered so revenue scales with the cost it creates. monday.com is migrating customers to a seats-plus-credits model. Atlassian moved Rovo AI features to usage-based credits, with billing beginning December 3, 2026 for most meters. OpenAI added a $500-a-month ChatGPT tier and reintroduced its $200 plan with half the previous usage allowance, while its ads business reached a $1 billion annualized run rate to fund a vast free user base. Every one of these moves is a bet about unit economics — and a pricing change you can only evaluate if you know your cost per unit.

How do you tell if an AI feature is profitable?

Compare what one unit of the business earns with what the AI costs to serve that unit, and watch the two lines as usage grows. A company-level gross margin hides the answer, because a profitable feature and an unprofitable one average out. The working view is a per-feature margin waterfall: revenue attributable to the feature per unit, minus fully loaded AI cost per unit, tracked monthly.

The unit should be one product and finance already report: daily or monthly active users for a consumer app, seats or workspaces for B2B software, loans funded for a lender, transactions for a payments product. Three warning signs show up in this view long before they show up in the quarterly margin:

• Cost per active user rising faster than revenue per active user. The feature gets less profitable as adoption grows — the opposite of software economics.
• A heavy-user tail. A small share of users consuming most of the inference on a flat price.
• Unmonetized consumption. Free tiers, trials, and beta features carrying cost with no revenue line — the Figma pattern above.

The best outcome is the one Duolingo reported: the cost of its AI video-call feature fell from an estimated $0.30 to "under $0.01 per video call," mainly by moving to open-source models. You can only make — and prove — that kind of change if the feature has a unit cost to begin with. The full formula set, including cost per MAU and cost per interaction, is in AI unit economics.

What does cost per loan look like with AI in the stack?

For a lender, the AI question is whether it lowers the total cost of originating a loan by more than it adds — measured per funded loan, not per model call. The baseline is steep. Mortgage Bankers Association data cited by HousingWire put origination costs at $11,898 per loan in Q1 2026, against $727 of pre-tax profit per loan. At that ratio, small changes in cost per loan move profit a lot — in either direction.

The upside is real: Upstart reported that 91% of its Q2 2026 loans were fully automated, across 558,014 loans. So is the risk. Writing for the MBA, IndiSoft’s Camillo Melchiorre warned that rising token costs for premium reasoning models are breaking mortgage-AI budget assumptions: "Mortgage organizations can no longer treat AI as an unbounded utility."

A defensible AI cost per loan includes the whole stack attributable to AI in the workflow — model tokens, cloud compute, document processing, third-party data, and the human review of AI-flagged files — divided by loans funded. Reported next to total origination cost per loan and profit per loan, it answers the only question the credit committee cares about: is AI making each loan cheaper to originate?

What happens to your margin if model prices rise?

AI gross margin moves in proportion to the share of unit cost that is inference, so it should be stress-tested before prices or usage move. The direction is not obvious. Per-token prices have fallen for most models, but the labs are building margin: SemiAnalysis research reported by Crypto Briefing estimated Anthropic’s API gross margins exceed 80%, and Anthropic’s IPO filing describes growth driven by usage — "customers have expanded their usage into new use cases and increased their spending accordingly." Meanwhile reasoning models and agents can consume many more tokens per task than the chat features they replace, so cost per unit can rise even when price per token falls.

An illustrative stress test, for a product with $2.40 of monthly revenue per DAU and $0.80 of fully loaded AI cost per DAU ($0.50 inference, $0.20 cloud, $0.10 retrieval and other):

The last row is the one to put in front of the pricing committee: a plausible combination takes a healthy AI product below ICONIQ’s projected 2026 average of 52%. The levers are the ones already in play across the market — model routing (GoDaddy’s Airo switches between models, including open-weight ones), cheaper models for routine work (Duolingo), and metered pricing that passes usage through to the customer (Atlassian, monday.com).

Who should own AI unit economics?

Finance should own the number; product should own the levers; both should read it in the units the business already reports. Engineering dashboards express AI cost in tokens and models, which no pricing committee or board can act on. Finance owns the invoices and the P&L; product owns pricing, packaging, and model choices. AI gross margin per unit is the shared language between them — and it has to reconcile to what the company actually paid, or it will not survive the first finance review.

That is the job of AI value management. COGScontrol aggregates AI and cloud spend across OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google Vertex AI, and the underlying cloud, attributes it to products and features, and divides it by business metrics that arrive via CSV upload, API, direct entry, or Google Sheets — producing cost per DAU, cost per loan, and AI gross margin that tie to the invoice. Model your own product in the AI unit economics calculator.