Skip to content

    What gross margin will an investor accept once AI inference is in your costs?

    Alperen Zeki Gokmen6 min readWebsiteLinkedIn

    What is a good SaaS gross margin once AI is in the product?

    Benchmarkit's 2026 report, drawn from 342 B2B SaaS and AI-native companies, puts the 2025 median software gross margin at 80%, steady for four years. It calls 75% to 85% the range for strong B2B SaaS. AI products run lower: ICONIQ's July 2026 survey projects 53% for them this year.

    Below are the numbers from three investor and benchmark sources, what they say about inference costs, how pricing changes the margin, and how the number moves with company size.

    What counts in SaaS gross margin?

    In this post, gross margin is revenue minus the cost of delivering the product, divided by revenue. That is our working definition. The cost line holds hosting, the model and inference bills, and the support it takes to keep customers running.

    Benchmarkit reports the number two ways. Gross margin on total revenue includes services such as implementation. Software gross margin counts software revenue only. In its 2026 report the software figure sits 4 points higher, 80% against 76%, and Benchmarkit says the gap reflects the dilutive effect of professional services and other non-⁠software revenue.

    So a company's number and a benchmark compare only when both use the same base.

    Have AI inference costs lowered SaaS gross margins?

    Not at the median, according to Benchmarkit. In its four years of data, median software gross margin held between 79% and 81%, and it writes that "Industry-wide LLM and AI infrastructure costs have not yet materially compressed software margins at the median."

    The total revenue figure moved a little. Median gross margin fell from 77% in 2024 to 76% in 2025. Benchmarkit raised the idea that AI integration costs caused the drop and says the hypothesis "could not be proven with available data". Its next edition will report gross margin separately for legacy SaaS products and AI-⁠native products.

    The pressure shows lower down. The bottom quartile of software gross margin is 50%, and Benchmarkit names "LLM inference costs, cloud infrastructure, and per-customer customization" as the primary contributors.

    What gross margin do AI products have?

    Lower than SaaS, in every source we read.

    Source Group measured Gross margin
    Benchmarkit, June 2026 Median software gross margin, B2B SaaS and AI-native companies, 2025 80%
    Benchmarkit, June 2026 Median gross margin on total revenue, 2025 76%
    ICONIQ, July 2026 AI products at software companies building AI, 2025 45%
    ICONIQ, July 2026 Same group, projected for 2026 53%
    Bessemer, August 2025 AI "Shooting Stars", average 60%
    Bessemer, August 2025 AI "Supernovas", average ~25% (often negative)

    ICONIQ's survey of over 300 executives at software companies building AI products projects AI product margins at 59% by 2027. It reports that two-thirds of companies see better per-query unit economics, helped by managing inference costs, better model routing and revenue growth. It also writes that "As products scale, talent's share of cost falls while inference rises."

    Bessemer, a venture investor, studied 20 high-growth AI startups for its 2025 report. Its Supernovas, startups "growing as fast as any in software history", have margins it calls "often stretched close to zero or even negative". Its Shooting Stars keep gross margins "slightly lower than SaaS peers due to faster growth and modest model-related costs". Bessemer calls the Shooting Stars "the most important benchmark for AI founders to aim for."

    What do investors ask about AI margins now?

    ICONIQ opens its July 2026 report with the shift: "A year ago, boards wanted to hear the AI strategy. Now they want AI unit economics."

    Benchmarkit points to the same thing from the cost side. It recommends that companies integrating AI "instrument the ability to capture and measure the impact of variable AI costs including inference, RAG, network, on software gross margins separately from core product COGS". Among the measures it lists for a quarterly review is gross profit per million tokens.

    Our reading of the two: an investor now wants two numbers: the SaaS margin and the AI margin, with the inference bill visible between them. A blended figure that hides the model cost answers neither question.

    Does the pricing model change gross margin?

    It does in Benchmarkit's data. Companies on usage-only pricing show the lowest median gross margin on total revenue, at 62%, which the report puts down to higher infrastructure and compute costs during the usage ramp. The three subscription variants sit between 76% and 84%.

    On software margin, Benchmarkit says companies with a hybrid of fixed fee plus usage show superior margins, "effectively passing AI infrastructure costs to customers".

    Charging for AI is still the minority choice. Of the companies with AI functionality, 67% do not charge for it separately and 29% do. Benchmarkit writes that the ones that cannot charge for it "are absorbing margin pressure without a clear ROI recovery mechanism". Among those that charge, subscription plus usage is the most common model at 39%, and 13% use usage-⁠only pricing.

    Does gross margin rise with revenue?

    Yes, in Benchmarkit's data. Median software gross margin rises from 72% at under 5 million USD of ARR to 86% at 50 to 100 million USD. For a company under 5 million USD of ARR, the like-for-like figure in this data is 72%, not the 80% median.

    The same direction shows in ICONIQ's AI product margins, 45% in 2025 rising to a projected 59% by 2027, as companies work on inference cost and routing.

    What does the number look like on one page?

    Example, with invented round figures: a company bills 100,000 EUR in a month. Hosting costs 8,000 EUR, the model API bill is 14,000 EUR and customer support costs 6,000 EUR. Gross margin is 72%. Without the model bill it would be 86%.

    Shown that way, the investor sees both the SaaS margin and what the AI layer costs, and can ask the next question: whether the 14,000 EUR falls as a share of revenue as the company grows. The other numbers an investor reads next to it are in The ten numbers an investor reads first. What investors ask about the AI itself is in AI wrapper or AI company?

    Work through your numbers with us

    TBHM Growth is our four-week online programme for founders whose company has revenue, paying customers and six to eight months of runway. Its third week works through acquisition cost, lifetime value, payback, gross margin, burn and runway, then the ten numbers an investor reads first. The application form is on the Growth page.

    Questions founders ask

    What is a good gross margin for SaaS?

    Benchmarkit's 2026 report calls 75% to 85% the range for strong B2B SaaS. Its median for 2025 was 76% on total revenue and 80% on software revenue alone, across 228 and 232 companies. The top quartile on software margin was above 86%.

    Do AI companies have lower gross margins?

    On AI products, yes, in the two investor sources we read. ICONIQ's July 2026 survey reports 45% gross margin on AI products in 2025 and projects 53% for 2026. Bessemer's 2025 report put its fast-growing Shooting Stars at 60% and its Supernovas at ~25%, often negative.

    Where do AI inference costs go in gross margin?

    Benchmarkit's 2026 report names LLM inference costs among the main reasons for its bottom-quartile software margin of 50%. It recommends that companies measure variable AI costs, including inference, RAG and network, on software gross margin separately from core product COGS.

    Sources

    This page is educational content and is not investment or financial advice.

    Next in this topic

    More from the library