AI Unit Economics: When LLM Features Eat Your Margin
AI unit economics means tracking per-user inference cost against per-user revenue. LLM features add a real variable cost, and that cost breaks the near-zero-marginal-cost math behind classic 80% SaaS margins. Four metrics tell the story: inference cost per active user, AI attach rate, token cost per retained user, and margin-adjusted LTV. Here's the punchline before you touch pricing. Flat pricing hides a bimodal margin distribution, and the blended average lies to you about it.
The margin assumption AI just broke
For twenty years the SaaS reflex was simple. Once you paid to build the software, serving the next user cost roughly nothing, so scale meant margin expansion. Sell more seats, watch gross margin drift toward 80-plus percent, tell the board the flywheel is working. That reflex is now wrong for any product with an LLM in the request path.
Inference is a per-query cost. Every time a user hits your AI feature, tokens get metered and you pay. ICONIQ Capital's 2026 State of AI survey — about 300 software executives — put average AI-product gross margins near 52% for 2026, up from 41% in 2024 but still well under the 75–85% traditional-SaaS range, as summarized by The SaaS CFO. And here's the part that should keep you up at night. ICONIQ found model-inference cost rising from about 20% to 23% of total spend as products mature, per Digital Applied. COGS grows as a share of cost when usage scales. The old math ran backwards.
This isn't theoretical anymore. SFAI Labs documented multiple public vertical-SaaS companies disclosing 6–9 points of year-over-year gross margin compression in Q4 2025, with explicit attribution to AI feature cost. When it shows up in a 10-K, it's real.
One caveat on that ICONIQ 52% figure, because I don't trust averages I can't see the tail of. It surveys companies still shipping AI products in 2026 — the ones that already blew up their margins and pulled features aren't in the sample the same way. Call it survivorship-flavored. And an average blends the customer running two prompts a month with the one running two thousand. The average is fine. The distribution is where your money leaks.
Four metrics for AI-era unit economics
Here's the spine. Four metrics, each with the formula spelled out, so you can drop them into a sheet today.
| Metric | What it tells you | Formula |
|---|---|---|
| Inference cost per active user (ICPAU) | Your average variable AI cost per head | Total inference $ ÷ active users |
| AI attach rate | How much of your base actually touches AI | Users touching AI features ÷ total active users |
| Token cost per retained user | Whether retained cohorts are cheap or expensive to keep | Inference $ from retained cohort ÷ retained users |
| Margin-adjusted LTV | LTV after inference eats its share | Classic LTV × gross margin after inference |
ICPAU is your headline number, but it's the one that hides the bimodal problem, so never show it alone. Attach rate matters because a 15% attach rate and a 75% attach rate are different businesses even at identical ICPAU. Token cost per retained user is the sneaky one — heavy AI users often retain better, which means your best cohort can also be your most expensive to serve. Margin-adjusted LTV is what actually belongs next to CAC. If you're still dividing raw LTV by CAC without the inference haircut, your LTV/CAC ratio is telling you a nicer story than the truth.
Assumptions in this framework — argue with me:
- "Active user" = one meaningful product action in a rolling 30-day window. Pick your own, but pick one and hold it.
- Attribution window for inference is the same 30 days. If your billing cycle differs, reconcile before comparing.
- Retained cohort = users active in month N who were also active in month N-1. Simple, deliberately.
- Inference cost is fully-loaded token spend only — not the salary of the team building the feature.
Change any of these and the numbers move. That's the point. Name them out loud so the CFO can poke them instead of discovering them later.
Worked example: the $49 seat with a $31 problem
Let me make this concrete with a tiny cohort. Product sells at $49 per seat per month. Rates below are illustrative — Anthropic's Claude Sonnet-class models ran roughly $3 input / $15 output per million tokens per Finout's June 2026 breakdown, and output is the expensive side, which is exactly why chatty AI features hurt. OpenAI's API works the same way: separate input and output rates, no flat fee, output pricier, as CloudZero documents. I'm using $3/$15 per million tokens and pricing a Sonnet-class model as of mid-2026. These get revised constantly, so recompute with today's card.
Five profiles across one team:
| User | Tokens/mo (in / out) | Inference cost | Revenue | Margin $ | Margin % |
|---|---|---|---|---|---|
| Light | 100k / 40k | $0.90 | $49 | $48.10 | 98% |
| Median | 400k / 200k | $4.20 | $49 | $44.80 | 91% |
| Heavy-ish | 900k / 500k | $10.20 | $49 | $38.80 | 79% |
| Power | 1.5M / 1.0M | $19.50 | $49 | $29.50 | 60% |
| Whale | 2.2M / 1.9M | $35.10 | $49 | $13.90 | 28% |
ICPAU across these five: total inference is $69.90, so $13.98 per user. Blended margin: total margin $175.10 on $245 revenue, about 71%. Looks survivable. You could put 71% on a slide and nobody blinks.
But look at the whale. That user generates $35 a month in tokens against a $49 seat — a 28% gross margin on a customer you're probably thrilled to have, because heavy users churn less. Your P50 user sits at 91%. Your P95 is bleeding.
A CFO I worked with called this "the smiling average." We showed her a blended 71% on an AI feature and she sat with it for about four seconds, then asked one question: "What's the worst decile?" We hadn't computed it. She sent us back to the sheet. When we returned with the P95 at 28%, she didn't kill the feature. She reweighted the forecast, because a fast-growing product pulls its cost distribution toward the heavy tail over time. The blended number was accurate and useless. The split was the actual finding.
The cost-allocation step
None of the four metrics work until you can attribute tokens to accounts. This is the unglamorous part everyone skips. If your logs say "we spent $40,000 on inference this month" but can't say which accounts generated it, you have a total, not unit economics.
In practice you meter token consumption per account — often against a credit pool so each account draws down a measurable balance — then join that to your revenue table by account ID. Some product-analytics tools, Kixo among them, tie AI usage back to specific accounts so the join isn't a monthly forensic exercise. However you do it, the attribution has to be clean before any ratio downstream means anything. Garbage allocation, garbage margin math. If you're already building a unit-economics dashboard, this is the column you add.
Three pricing responses, stress-tested against churn
You have three moves when a feature runs hot. Each trades margin recovery against retention risk, and the trade matters more than the headline saving.
| Response | Margin recovery | Revenue risk | Retention risk | Implementation lift |
|---|---|---|---|---|
| Usage caps | High | Low | High (hits best users) | Low |
| Usage-based add-on | High | Medium | Medium | Medium |
| Routing + caching | Medium-High | None | None | Medium-High |
Usage caps are the blunt fix. Cap the whale, protect the margin, done. The problem is that the whale is often your stickiest, most expansion-ready account, and a cap reads as a downgrade to the customer paying you the most attention. You recover margin by antagonizing the cohort with the best retention curve. Sometimes right. Rarely painless.
Usage-based add-ons align cost to value — the user consuming more compute pays more, which is honest. The industry signal here is loud. GitHub announced on April 27, 2026 that all Copilot plans move to usage-based billing on June 1, 2026, metering by token consumption at model API rates, per the GitHub Blog. When the company with the flagship flat-priced AI product concedes that flat pricing buckles under heavy users, that's not a footnote. That's the market telling you where this goes.
Model-tier routing plus prompt caching attacks COGS without touching price at all. Route simple requests to a cheaper model, cache repeated context. Anthropic's own numbers say prompt caching can cut cost up to 90% on long prompts, since cache-read tokens bill at roughly 10% of the standard input rate, as TechTalks reported. Stack routing, caching, and batching and SaaS Mag cites analysts putting the combined saving at 50–70% with no measurable quality loss.
My defensible take: routing and caching is the only response that doesn't put retention at risk. Caps churn your heavy users. Add-ons introduce price friction and a new billing surface. Cutting COGS touches neither the customer nor the invoice. So the order of operations is clear — attack COGS first, then reprice only what's still underwater. Recompute the worked-example whale at a 60% cost cut and that 28% margin becomes about 71%. Same customer, same price, no churn risk. You bought yourself the entire problem back.
What to put in front of your CFO
Two of the four metrics belong in the board deck. Not all four. ICPAU and margin-adjusted LTV by cohort. Attach rate and token-cost-per-retained-user are operator metrics — you use them to run the feature, not to defend it upstairs.
The one slide that changes the conversation reframes "AI is expensive" into "here's margin-adjusted LTV by cohort." Instead of a scary total inference bill, you show that your median cohort runs 90%-plus and only the top decile needs a pricing or routing response you've already scoped. That moves the discussion from panic to plan.
| Cohort | ICPAU | Gross margin after inference | Margin-adjusted LTV vs classic |
|---|---|---|---|
| Median | $4.20 | 91% | −9% |
| Top decile | $35 | 28% | −72% |
Dependency, restated: every number above holds only if token attribution to accounts is clean. If you can't tie inference to specific accounts, you don't have unit economics — you have a total and a hope. Fix the allocation first.
This is the same discipline as any efficiency argument. The MER vs ROAS fight taught marketers that a blended number can be true and misleading at once. AI margin is the same trap wearing a new outfit.
FAQ
What are good AI gross margins in 2026? ICONIQ's 2026 survey (via The SaaS CFO) put AI-product margins around 52% on average, up from 41% in 2024, versus 75–85% for traditional SaaS. Treat that 52% as survivorship-flavored — it reflects companies still shipping, and the average hides a heavy-user tail that runs far lower.
Does inference cost fall with scale? No, and this is the counterintuitive part. ICONIQ found inference rising from about 20% to 23% of total spend as products mature, per Digital Applied. Unlike classic software COGS, inference is a per-query variable cost, so it grows as a share of your cost base when usage climbs rather than shrinking with scale.
Input or output tokens — which drives cost? Output. Finout's June 2026 breakdown shows Claude Sonnet-class models around $3 input / $15 output per million tokens, and CloudZero confirms OpenAI's API prices input and output separately with output more expensive. Verbose AI features that generate long responses cost multiples of features that mostly read context.
How much can caching and routing actually save? Anthropic reports prompt caching cutting cost up to 90% on long prompts (cache reads bill near 10% of input rate), per TechTalks. SaaS Mag cites analysts putting routing plus caching plus batching at a combined 50–70% reduction without measurable quality loss. Attack this before you touch price.