Why AI Gross Margins Differ From SaaS

By Brian Tr · Updated Oct 6, 2026

Gross margin is the number that most cleanly separates a software business from a services business. SaaS companies earned their premium valuations in part because each extra customer cost almost nothing to serve. AI products don't work that way. Every request runs on compute someone has to pay for, and some products still need people checking the output.

This article compares the two margin profiles side by side, explains why the gap exists, looks at whether it's permanent, and covers what it means for pricing and for how investors read your numbers. If you need the basics first, start with what gross margin means.

The Classic SaaS Margin Profile

A software product is written once and sold many times. Serving one more customer means a little more hosting, some support and payment processing. The bulk of the spending sits below the gross margin line, in engineering and sales. That's why mature SaaS businesses report high and stable margins.

Bessemer's Scaling to $100 Million benchmarks put the average gross margin for a cloud business at 65% to 70%, and note that the middle 50% of companies sit within roughly 60% to 80%. It lists average margins of 70% at $1 million to $25 million in ARR, 65% between $25 million and $100 million, and 70% above that. It also flags exceptions: Twilio ran at around 50% in its growth years because of the telecom costs underneath its product.

Public companies at scale can run higher. ServiceNow's full-year 2025 results show subscription gross margin of 80% on a GAAP basis and 83.5% non-GAAP, with 2026 guidance of 82% non-GAAP. That's a company with enormous scale and a mature product, so treat it as the top of the range rather than the norm for a startup.

The AI Margin Profile

The most cited primary source on this is a16z's The New Business of AI (and How It's Different From Traditional Software), published by Martin Casado and Matt Bornstein on February 16, 2020. Their finding: AI companies often have "gross margins often in the 50-60% range", well below the 60% to 80% plus benchmark for comparable SaaS businesses.

They pointed to three drivers:

• Cloud infrastructure. AI companies often spend 25% or more of revenue on cloud resources.
• Human support. Many companies spend up to 10% to 15% of revenue on data labeling and model maintenance, and some products need people reviewing outputs in real time.
• Model upkeep. Training a single model can cost hundreds of thousands of dollars or more, and retraining is a recurring cost because data drifts.

That essay predates the current generation of large language models, so the specific causes have shifted. The core idea has held up: AI blends software economics with services economics.

More recent data is in Bessemer's State of AI 2025. It separates fast-growing "Supernovas" from steadier "Shooting Stars." Supernovas average a gross margin of about 25%, often negative, because they trade profit for distribution. Shooting Stars average about 60%, which Bessemer describes as slightly lower than SaaS peers because of faster growth and modest model-related costs.

ICONIQ's 2026 State of AI report, based on a Q2 2026 survey of more than 300 executives at software companies building AI products, shows average AI product gross margin at 45% in 2025, projected to reach 53% in 2026 and 59% in 2027.

Side-by-Side Comparison

The figures aren't measuring the same thing in every row. The surveys use different samples, years and definitions of cost, so read them as ranges and not as a precise gap.

Why the Gap Exists

Compute is a cost of revenue, not an overhead

In SaaS, a heavier user costs slightly more to host. In AI, a heavier user can cost dramatically more, because each prompt triggers computation. That cost lands in cost of revenue, so it hits gross margin directly. The mechanics of how inference spend turns into margin are covered in AI inference costs and gross margin.

Third-party model fees

A product built on someone else's model pays that supplier on every call. You don't control that price, and you can't simply raise it. If your product is thin on top of the model, you'll also struggle to charge enough to cover it, which is one reason wrappers face thin margins.

Humans in the loop

Many AI products route uncertain cases to people, or need staff to label data and check quality. That's services-style labor sitting inside a software product. a16z flagged it in 2020 as a source of cost that SaaS simply doesn't carry at the same scale.

Usage that doesn't match the price

SaaS pricing is usually per seat and flat, which works when the cost to serve is flat. When cost scales with use but revenue doesn't, margin gets worse as the product gets better. This is the heart of the AI pricing question for SaaS companies.

Growth-first choices

Part of the gap is a decision, not a law of nature. Bessemer's Supernovas run near 25% margin partly because they are buying adoption. Subsidised usage lowers margin today in the hope of locking in customers.

How Investors Read the Difference

Investors don't treat 60% AI margin the same way as 60% SaaS margin. They ask four questions.

1. Which costs are in the number? Companies classify costs differently. Whether inference, human review and support sit above the line changes the margin you report. Make your cost classification explicit before someone else guesses.
2. Is the margin trending up? A product at 45% and rising is different from one at 45% and stuck. ICONIQ's projections of 53% in 2026 and 59% in 2027 are survey expectations, not guarantees, but they show what the market expects founders to demonstrate.
3. Is the cost tied to growth? If margin falls as usage rises, scale makes the business worse, not better.
4. What is being bought with the low margin? Margin traded for durable retention reads as an investment. Margin traded for novelty-driven usage reads as risk. Bessemer says outright that fast adoption can belie low switching costs or signal novelty.

On valuation, the standard SaaS logic of paying a premium for recurring revenue assumes high margins. A lower gross margin means each dollar of revenue produces less gross profit, so revenue multiples tend to reflect that. Compare on gross profit as well as on revenue, and be careful when borrowing SaaS multiples for an AI business. For the broader set of ratios investors use, see SaaS economics and unit metrics.

Is the Gap Permanent?

Probably not permanent, but don't count on it closing by itself. There are forces on both sides.

Pushing the gap smaller:

• Unit costs of running models have tended to fall as models and hardware improve.
• Teams learn to route simple requests to cheaper models and to cache repeated work.
• Pricing moves from flat seats to usage-aware plans, so revenue tracks cost.
• Products that automate more of the human review reduce the labor line.

ICONIQ's trajectory from 45% to a projected 59% points in this direction. The levers behind it are covered in AI model cost management.

Holding the gap open:

• Usage per customer can grow faster than cost per unit falls. Cheaper requests often lead to many more requests.
• Customers expect more capable features, and the most capable models usually cost more to run.
• Competitive pressure keeps prices from rising to cover costs.
• Products that do more work per task, such as agents running many steps, consume more compute per outcome.

The honest summary is that margins can improve, but the structural difference remains: SaaS cost of revenue is mostly fixed, and AI cost of revenue has a large variable part. Even a well-run AI product is likely to carry more cost sensitivity than a traditional SaaS product. Plan for a margin that's earned through pricing and engineering, not one that arrives with time.

Pricing and Benchmarking Implications

Price to the cost driver. If cost follows usage, at least part of the price should too. Hybrid models, a base fee with usage or outcome components, are a common response.

Benchmark against the right peers. Comparing an AI product to classic SaaS flatters nobody. Compare to AI companies at the same stage and with the same human-in-the-loop share, and report the SaaS comparison separately.

Track margin by customer and by feature. A blended margin can hide a handful of heavy accounts that lose money. Break it down.

Separate the temporary from the structural. If low margin comes from subsidised growth, say so and show when it ends. If it comes from the product's architecture, fix the architecture.

Report cost lines plainly. State what sits in cost of revenue: inference, third-party fees, human review, support. It builds trust and makes your trend easy to follow.

Key Facts

• a16z (February 16, 2020) found AI gross margins often in the 50% to 60% range, against 60% to 80% plus for comparable SaaS, with cloud costs of 25% or more of revenue and up to 10% to 15% on labeling and model maintenance (a16z).
• Bessemer's cloud benchmarks put average SaaS gross margin at 65% to 70%, with the middle 50% of companies at roughly 60% to 80% (Bessemer, Scaling to $100 Million).
• Bessemer's State of AI 2025 reports about 25% gross margin for AI "Supernovas" (often negative) and about 60% for "Shooting Stars" (Bessemer).
• ICONIQ reports AI product gross margin averaging 45% in 2025, with 53% projected for 2026 and 59% for 2027, from a Q2 2026 survey of more than 300 software executives (ICONIQ).
• ServiceNow reported full-year 2025 subscription gross margin of 80% GAAP and 83.5% non-GAAP (ServiceNow).

What This Means for Founders and Owner-CEOs

If you're building an AI product, expect lower margins than a SaaS peer early on and plan the path upward. Know your cost per customer, price with that cost in view, and be clear with investors about what's temporary. If you want the wider picture of how AI companies make money, read AI startup business models, and for where these costs sit on the income statement, what counts as cost of goods sold for an AI company.

If you're buying AI tools, the same logic explains why vendors are moving toward usage-based pricing. Their cost grows with your use, so their price will often grow with it too.

Related Reading

• Gross Margin
• SaaS Economics and Unit Metrics
• AI Inference Costs and Gross Margin
• AI Model Cost Management
• What Counts as COGS for an AI Company
• AI Wrapper vs AI-Native
• The AI Pricing Question for SaaS