Last updated: 31 October 2026
AI startups usually run lower gross margins than SaaS. Typical SaaS is 65–80%+, while AI products are often 45–60% because inference, hosting and support sit in cost of goods. ICONIQ expects AI margins to rise from 45% to 59% by 2027.
What is gross margin?
Gross margin is (revenue – cost of goods sold) divided by revenue. For AI products, cost of goods includes inference bills, hosting and customer support. In our worked example, £1.00m revenue and £0.40m inference and hosting give a gross margin of 60%. This shows why a product can have strong revenue but a lower gross margin when each sale carries a direct compute cost. For SaaS, the same formula applies, but cost of goods is usually much smaller because adding one more user rarely adds a large marginal cost.
What is a normal SaaS gross margin?
Bessemer’s cloud portfolio, tracked from 2010 to the first half of 2021, had typical gross margins of 65–70%. The range moved by scale: 70% at $1–10m ARR, 65% at $25–100m ARR, and 70% at $100m+ ARR. a16z also places comparable SaaS businesses at 60–80%+. These figures are useful as a benchmark, but they come from different sources and periods, so they should not be treated as one single market average.
What gross margin do AI startups have?
a16z, writing in 2020, said AI companies’ gross margins were often in the 50–60% range, below the 60–80%+ SaaS benchmark. ICONIQ’s State of AI report, published in July 2026 and based on about 300 executives, gives a more recent view: average AI product gross margin was 45% in 2025, with projections of 53% in 2026 and 59% in 2027. The direction is upward, but the 2026 and 2027 figures are projections, not reported actuals.
Why are AI margins lower?
a16z gives three main reasons. First, AI companies often spend 25% or more of revenue on cloud resources. Second, many spend up to 10–15% of revenue on data labelling and ongoing model work. Third, a16z says these costs are unlikely to reach the near-zero levels seen in SaaS. There is also a structural point: every query spends inference, and inference is priced per token. That means usage is not just a support cost; it is a direct cost attached to each customer action. The worked example above, £1.00m revenue and £0.40m inference and hosting, shows how quickly that cost can take a third or more of revenue.
| Segment | Gross margin | Period / source |
|---|---|---|
| SaaS, Bessemer cloud portfolio | 65–70% | 2010 to first half of 2021 |
| Comparable SaaS, a16z | 60–80%+ | 2020 |
| AI companies, a16z | 50–60% | 2020 |
| AI products average, ICONIQ | 45% in 2025; 53% in 2026; 59% in 2027 | July 2026 survey of about 300 executives |
How can AI startups improve margin?
Our own list of levers includes routing easy queries to cheaper models, caching prompts, capping output length, and pricing on usage or outcomes. Vendor prices differ widely, so the right lever depends on the product’s traffic pattern and model mix. ICONIQ projects improvement year on year, but that does not prove any one startup will reach SaaS-like margins. Founders should check their own inference, hosting and support costs before using any benchmark as a target.
Frequently asked questions
Are AI startup gross margins always below SaaS margins?
Not necessarily. These are averages and ranges. a16z gives 50–60% for AI companies against 60–80%+ for SaaS, so AI is usually lower, but individual companies vary with product and cost structure.
What counts as cost of goods for an AI product?
For AI products, cost of goods includes inference bills, hosting and customer support. Inference is priced per token, so each query can add a direct cost. That is why AI cost of goods can be much larger than in traditional SaaS.
Can AI margins improve over time?
Yes, ICONIQ projects average AI product gross margin rising from 45% in 2025 to 53% in 2026 and 59% in 2027. Those are projections from a July 2026 survey of about 300 executives, not guaranteed outcomes.
Should founders use these benchmarks to set prices?
They can use them as a starting point, but they should check their own costs first. AI cost of goods includes inference, hosting and support, and vendor prices differ widely.
Sources
Figures and prices were checked on 31 October 2026; check the linked sources before relying on them. This article is general information, not tax, legal or investment advice.
