What is a good gross margin for an AI SaaS in 2026?
A healthy AI SaaS gross margin in 2026 is 55 to 65 percent. Traditional SaaS runs 70 to 80 percent or more. The average AI product sits below both: ICONIQ's State of AI series measured 41 percent in 2024, 45 percent in 2025, and projects roughly 52 percent for 2026. Below 35 percent you are in the danger zone where one power user can put a customer, or the whole company, underwater.
How do AI margins compare to traditional SaaS?
The gap is structural, not a phase. Software has near-zero marginal cost; AI products pay for inference on every request. a16z flagged this in 2020, finding AI companies at 50 to 60 percent gross margins versus the 60 to 80 percent plus SaaS norm, often with 25 percent or more of revenue going to cloud. Harvard Business Review put AI services near 30 percent gross margin in late 2025 against the 60 to 80 percent software norm.
Even the model labs are not at software margins: outside estimates compiled by Tanay Jaipuria put OpenAI around 50 percent and Anthropic around 60 percent. The same survey cites Bessemer data showing their fastest-growing AI "Supernovas" averaging around 25 percent gross margin early on, some negative, while steadier "Shooting Stars" run near 60 percent. If your margin looks worse than the SaaS books said it should, you are not failing at SaaS. You are running a different cost structure.
What do the 2026 benchmarks say?
Sources: ICONIQ State of AI 2025 and the 2026 snapshot. Two honest caveats. First, these are averages over venture-backed companies, mostly later stage than a founder at $100k to $2M ARR. Second, averages hide the distribution: margin problems in AI products concentrate in a handful of heavy customers, so a respectable blended margin can coexist with your top accounts being served at a loss. Benchmarks tell you the neighborhood. Only your own per-customer numbers tell you the truth.
Interpretation bands we use when reading a founder's margin map:
| AI gross margin | Reading |
|---|---|
| 65%+ | Healthy. Pricing is doing its job. |
| 55–65% | Safe. Watch the P99 customers quarterly. |
| 45–55% | Below the 2026 average trajectory. Usually a few specific accounts are the leak. |
| 35–45% | You are leaking. Flat pricing against long-tail usage, almost always. |
| <35% | Bleeding. Investors price this like a services business, and one whale can tip you negative. |
Why do AI margins collapse?
The public cases all follow one script: pricing that does not meter the thing that costs money, plus a small set of users who consume most of the compute.
The Wall Street Journal reported Copilot lost an average of $20 per user per month in its $10 flat-fee era, with some users costing $80. ARK Invest disputed the figures, but GitHub's own trajectory since, ending with all plans moving to usage-based billing on June 1, 2026, is the strongest confirmation that flat pricing was not covering the tail.
Aakash Gupta reported Replit's gross margin going from 36 percent to negative 14 percent in roughly two months after its coding agent shipped, because the agent consumed more model cost than pricing recovered. TechCrunch later reported margins recovering to around 36 percent by late 2025, after Replit repriced around effort-based credits. The fix was pricing, not cheaper models.
Cursor's June 2025 move from request quotas to compute-based pricing triggered enough backlash that the CEO apologized publicly and refunded surprise charges. The lesson is different here: the repricing was economically necessary, and the damage came from how it was communicated. Margin fixes fail on trust as often as on math.
TechCrunch's summary of the category in August 2025 quoted founders and investors describing AI coding startups' margins as "absolutely abysmal," sometimes "very negative". These are the campfire stories. The mechanism in every one of them is the same power-law: a minority of users consuming the majority of compute under a price that ignores usage.
How do you calculate your AI gross margin?
Gross margin = (revenue − cost of goods sold) ÷ revenue. For an AI SaaS, COGS includes everything that scales with serving customers:
- Model API fees (OpenAI, Anthropic, etc.) or GPU compute for self-hosted models
- Vector database and retrieval infrastructure billed per usage
- Hosting, storage, and bandwidth for the serving path
- Third-party per-request services in the product loop (transcription, embeddings, search APIs)
It excludes R&D, marketing, and your salary. The most common mistake we see is booking model API spend under operating expenses, which flatters the margin and hides the leak. The second most common is computing only the blended number. Blended margin is the average of your profitable customers and your unprofitable ones, and the unprofitable ones are precisely what it conceals. The per-customer computation is the one that changes decisions, and it is the subject of our step-by-step cost-per-customer guide.
Margins are a pricing problem, not just a cost problem
Cost optimization is real work with a floor: you can cache, batch, route to cheaper models, and trim context, and your cost per request will still be nonzero and still be distributed with a long tail. Repricing has no floor. Moving one threshold, capping one tier, or adding one overage rate can move gross margin by tens of points in a month, because it changes who pays for the tail. Replit's recovery came from repricing. Copilot's June 2026 change is repricing. If your margin is under 50 percent, the highest-leverage fix is almost never a cheaper prompt. It is making your heaviest users' bills track their costs. Our guide on repricing without losing customers covers how to ship that change safely.
Frequently asked questions
Do API costs count as COGS?
Yes. Anything that scales with serving customers, model APIs, GPU time, vector queries, belongs in COGS. Booking it under opex flatters gross margin and hides per-customer losses.
What percentage of revenue should inference cost?
ICONIQ measures roughly 23 percent of revenue at scaling-stage AI companies. Past about a third, reprice before you optimize.
Why are AI companies unprofitable?
Mostly because marginal cost is real and pricing was copied from zero-marginal-cost SaaS. The result is a subsidy flowing from light users to heavy users, visible only when you compute margin per customer.
Are AI wrappers profitable?
They can be, when the markup over model cost is 3x to 5x and heavy usage is metered. Wrappers fail when a flat fee meets one enthusiastic customer with an agent loop.
Find out where your margin actually goes
AI Pricing Lab maps margin per customer from your real Stripe and usage data, simulates three pricing models on your last 6 to 12 months, and hands you the blueprint plus migration pack. Delivered in 10 business days. One-time fee, no subscription.
See the AI Pricing Reset$1,500 founding rate · first 10 AI SaaS · $2,995 list