AI gross margin is revenue minus the direct cost of serving customers (model inference, hosting, third-party APIs, payment fees and support), divided by revenue. AI products typically run at 25–60% gross margin, against 75–80% for traditional SaaS. ICONIQ put the 2025 average for AI products at 45%, projected to reach 53% in 2026. Model routing, caching and usage limits are the biggest levers.
Key takeaways
- Put inference in your cost of goods sold. If you leave it out, your margin is fiction.
- AI products average roughly 45–55% gross margin in 2025–26, and some of the fastest growers have been negative.
- A scheduled model price rise can take 18 points off your margin overnight, so cost your product at tomorrow's prices.
- Routing simple tasks to small models is usually the single biggest margin lever.
- Gross margin drives valuation: buyers pay higher revenue multiples for higher margins.
In the book, I set readers a simple exercise. Calculate your gross margin, then ask: is there anything you included that could be excluded?
My example was Close.com, a CRM for startups. It originally included unlimited free phone calls in its plans. Lovely for customers. Less lovely for Close, because every call cost it money. After a while, it pulled calls out of the plan and charged users per call on top of the monthly fee.
That was one feature in one CRM.
With AI, every feature is a phone call. Every summary, draft, search and agent step sends a meter spinning at a model vendor. And if you've bundled it all into a flat monthly price, you've quietly rebuilt Close's old problem across your whole product.
So let's measure it properly, see what everyone else is making, and look at what actually moves the number.
What is AI gross margin?
Gross margin is the share of each dollar of revenue left after the direct costs of delivering your product. The formula is the one from Chapter Five of the book:
Gross margin = (Revenue − Cost of goods sold) ÷ Revenue × 100
The hard part with AI is deciding what goes in cost of goods sold (COGS). Here's the list I'd use:
| Cost line | What it covers | What drives it |
|---|---|---|
| Model inference | Tokens or GPU time to serve customers | Tasks per user, tokens per task, model choice |
| Hosting and storage | Servers, vector databases, file storage | Data volume, queries |
| Third-party APIs | Search, speech-to-text, data enrichment | Calls per task |
| Payment fees | Card processing on each charge | Price and billing frequency |
| Customer support | People and tools to keep customers running | Customer count, product complexity |
| Human review | Any person checking AI output before delivery | Volume and risk level |
Training your own models, R&D experiments and sales costs usually sit below gross margin, in operating expenses. Be consistent, and don't leave inference out to make the number look better. Any serious buyer will put it back.
What gross margins do AI companies make?
Lower than you'd hope. Here's the 2025–26 evidence, with the source for each figure.
| Company or group | Gross margin | Source and date |
|---|---|---|
| Traditional SaaS | 75–80% | SaaStr, Feb 2026 |
| AI-native companies | 50–60% | SaaStr, Feb 2026 |
| Average AI product | 45% (2025), projected 53% (2026), 59% (2027) | ICONIQ survey via SaaStr, Jul 2026 |
| Bessemer "Shooting Stars" (steadier AI growers) | About 60% | Bessemer, Aug 2025 |
| Bessemer "Supernovas" (fastest AI growers) | About 25%, "often negative" | Bessemer, Aug 2025 |
| Lovable | About 35% (2025, estimated) | Sacra, via Pricing Innovation, Oct 2026 |
| Replit | 36% falling to -14% in early 2025, about 23% for 2025 (estimated) | Sacra, via Pricing Innovation, Oct 2026 |
| Cursor | -23% (quarter to Jan 2026, reported) | Contrary Research, via Tech Economics, Jun 2026 |
| Base44 (Wix) | "Nearly zero" in early 2026, targeting about 60% in H2 | Wix, via Ynet, 2026 |
For contrast, the book quotes WiseTech's CargoWise at a 92% margin and Australia's leading property portal at a 92% gross margin in 2021. That's what a software moat looks like before AI costs arrive.
ICONIQ found something else worth knowing. Only 15% of consumption-priced AI companies absorb all of their inference cost. The rest pass at least some of it on to customers through usage pricing, credits or overage. That's the market telling you what works.
Worked example: a $49 AI writing assistant
Let's build a margin from scratch, then break it, then fix it.
1,000 customers at $49 a month
Monthly cost per customer (model prices checked 27 October 2026):
| Cost line | Assumption | Per customer |
|---|---|---|
| Inference | 600 tasks on Gemini 4 Argon ($2 in, $10 out): 4,000 tokens in, 700 out = $0.015 each | $9.00 |
| Hosting and vector database | Illustrative | $1.50 |
| Search API | Illustrative | $0.80 |
| Payment fees | 2.9% + 30c per monthly charge | $1.72 |
| Support | Illustrative | $2.00 |
| Total COGS | $15.02 |
Gross margin: ($49 − $15.02) ÷ $49 = 69.3%. Healthy for AI.
Now the break. Google's pricing page says Gemini 4 Argon moves to $4 in and $20 out on 1 December 2026. Each task now costs $0.030. Inference doubles to $18, COGS rises to $24.02, and gross margin falls to 51%. Eighteen points gone, and you didn't change a thing.
Now the fix. Route the 60% of tasks that are simple (short rewrites, titles, summaries) to Gemini 3.1 Flash-Lite ($0.25 in, $1.50 out), at about $0.002 each. Keep the other 40% on Argon at the new price.
Inference falls to 360 × $0.002 + 240 × $0.030 = about $7.94. COGS drops to $13.96 and gross margin climbs to 71.5%, higher than before the price rise.
The assumptions are illustrative, but the pattern is real. Your model choice moves your margin more than almost anything else, and a vendor's price change can undo a year of careful pricing.
AI Gross Margin CalculatorModel token costs, heavy users and margin for an AI product.Open the free tool →Why do AI margins get squeezed?
There are six usual suspects.
- Flat "unlimited" plans. A few heavy users cost many times the plan price. In my example above, a customer running 6,000 tasks a month at the new Argon price costs $180 in inference alone, on a $49 plan.
- Agents. Multi-step agents resend their context at every step, so a single agent run can cost a hundred times a chat reply. The token pricing guide has the maths.
- Scheduled price rises. Google has set both Gemini 4 Argon and Gemini 3.8 Flash to double after launch.
- Supplier dependence. If you resell one vendor's model, that vendor sets your cost and can compete with you. An earlier estimate put Cursor's Anthropic bill at about $650 million a year against about $500 million in revenue.
- Price competition at the outcome level. HubSpot cut its Customer Agent from $1 per conversation to $0.50 per resolved conversation in April 2026, undercutting Fin's $0.99.
- Generous free tiers. Free users run real inference. Somebody pays.
Seven ways to improve AI gross margin
Here are the levers that real companies pulled in 2025 and 2026, roughly in the order I'd try them.
1. Route every task to the cheapest model that's good enough
Small models now cost around $0.10 per million input tokens against $4 to $10 for flagships (compare them all in the LLM API pricing comparison). Replit introduced model routing in August 2026. Cursor routes work to cheaper models and its own. Test quality on your real tasks, then route.
2. Cache what repeats
Long system prompts and reference documents are resent on every call. Cached input costs 90–95% less at OpenAI, and Anthropic's cache hits cost about 5% of the input price.
3. Batch what can wait
OpenAI, Anthropic and Google all take 50% off for jobs that can run within 24 hours. Overnight reports, bulk tagging and backfills belong here.
4. Own a cost lever
Cursor launched its own Composer model in October 2025 (see how Cursor broke "unlimited"). Wix said Base44's margin was "nearly zero" at the start of 2026 and targeted about 60% in the second half after moving to its own model. You don't need to train a frontier model, but you do need some lever over your biggest cost.
5. Put allowances and caps on every plan
Anthropic added weekly limits to Claude Pro and Max in August 2025, aimed at the fewer than 5% of users running Claude Code around the clock or sharing accounts. That's the fix for the heavy-user tail: a fair allowance, then paid overage.
6. Meter the expensive work separately
Perplexity keeps chat on a flat subscription but charges credits for its agent. Adobe keeps standard generations unlimited and meters premium ones. Separate the cheap core from the costly extras so core users aren't subsidizing heavy ones.
7. Charge for your platform, not just the pass-through
Clay admitted its old Pro tier had been run at a loss. In March 2026 it split pricing into Data Credits for the third-party data it resells and Actions for its own platform work, so it earns a margin even when customers bring their own API keys. If customers can bypass your resale margin, price the work you do.
Find out which customers you're losing money on
Model light, typical and heavy users, your token costs and your plan price. The free AI Gross Margin Calculator shows your margin by segment.
Does gross margin affect what your business is worth?
Yes, and a lot.
In the book I put high gross margin under the T in MOATS (Toughness) for exactly this reason: a higher margin means your business trades at a higher price-to-revenue multiple. Two companies with the same ARR are not worth the same if one keeps 75 cents of each dollar and the other keeps 35.
AI makes this sharper. A buyer will ask three questions: what's your margin with all inference in COGS, what happens to it if your model supplier raises prices, and how much of your revenue depends on one vendor? Have answers ready.
For the valuation side, see SaaS valuation multiples and try the valuation estimator. For a full metric guide, see gross margin for subscription businesses.
Watch the currency
Every major model vendor bills in US dollars. If you sell in Australian dollars, a weaker A$ raises your COGS with no change in usage. Track your margin in the currency you report in, and leave a buffer in your price. Remember too that GST you collect isn't revenue, so calculate margin on your GST-exclusive price.
Free Subscription Profit Review
Australian business owners can book a free 30-minute review with a BPAG Accountants adviser. An honest conversation about:
- whether a subscription model could work for your business
- how recurring revenue could cover your fixed costs each month
- how it could lift what your business is worth
No cost, no obligation. General information only until we understand your situation.
What to do next
- Rebuild your COGS with every serving cost in it, including inference, APIs and support.
- Calculate margin by customer segment, not just the average. Find your heavy-user tail.
- Re-cost at announced prices. Check every model you depend on for scheduled increases.
- Pull the cheap levers first: routing, caching and batching.
- Fix the plan structure if heavy users are still unprofitable: allowances, caps and paid overage. The AI pricing guide walks through it.
- If you're mostly reselling one model, read pricing an AI wrapper before your next price change.
Then go back to my Close.com question. Look at your plan and ask: is there anything included that could be excluded, or metered? In AI, there almost always is.
Frequently asked questions
Is inference a cost of goods sold or an operating expense?
Treat the inference you run to serve paying customers as cost of goods sold, alongside hosting and third-party APIs. Spending on training models or on R&D experiments usually sits in operating expenses. Whatever you choose, be consistent and disclose it, because buyers and investors will rebuild your margin with inference in COGS anyway.
Why do AI companies accept low gross margins?
Mostly to grow fast while model prices fall. ICONIQ's survey projects average AI-product margins rising from 45% in 2025 to 59% in 2027 as costs drop and pricing matures. SaaStr argues inference is 'the new sales and marketing spend': a cost of winning customers. That bet only works if your price holds while costs fall.
Does a free tier hurt AI gross margin?
It can. Free users run real inference, and that cost has to go somewhere. Many companies count it as a marketing cost rather than COGS, which flatters gross margin. Whichever way you book it, size free allowances to reach the 'aha' moment, then cap them, as Manus and Genspark do with daily free credits.
How fast are AI inference costs falling?
Fast for a fixed level of capability. Epoch AI estimated the price of reaching a given benchmark score fell about 40x a year, with a range of 9x to 900x by task. But frontier tiers and agents push the other way, so many businesses see total inference bills rise even as unit prices fall.
What gross margin should I use when valuing an AI business?
Use the margin with all serving costs in COGS, at current or announced model prices, not a future target. Buyers will also look at the trend and at how dependent you are on one model vendor. A rising margin with a credible cost plan is worth more than a high margin built on a promotional price.
Get both books free
The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.
Robert Coorey, MBA
Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.
Sources
- Bessemer Venture Partners: The State of AI 2025 (13 Aug 2025)
- SaaStr: 10 metrics from ICONIQ's 2026 State of AI report (10 Jul 2026)
- SaaStr: Inference is the new sales and marketing spend (9 Feb 2026)
- Tech Economics: The $60 billion company with negative gross margins (11 Jun 2026)
- Pricing Innovation: Replit's pricing journey (19 Oct 2026)
- Ynet: Wix and Base44 margins (2026)
- Clay: Introducing Clay's new pricing (11 Mar 2026)
- Google: Gemini Developer API pricing (checked 27 Oct 2026)
- Epoch AI: LLM inference price trends (12 Mar 2025)
Facts and prices checked 27 October 2026. Spotted something out of date? Tell us.