// BLOG / AI COMMERCIALIZATION
AI Gross Margins: Rethinking the 90% SaaS Benchmark
AI delivery cost can rise with every request. A software margin assumption is not enough to price that workload.
Traditional software's low marginal serving cost supported high gross margins. AI inference changes the cost structure, and venture firms disagree about how much current margin should influence an investment decision. Bessemer and a16z published different readings in August 2025.
AI products pay for inference and, where used, retrieval infrastructure and human review. ICONIQ's January 2026 executive survey projected AI product margins of 45 to 53 percent for 2026 across groups defined by their primary differentiator. Those are self-reported executive projections, not audited results or a causal test of architecture. Selected fast-growing AI companies in Bessemer's research ran lower margins, some negative.
The useful questions are what the reported numbers measure, which costs your product carries and what can change its margin.
Where the 90 percent benchmark came from
High software margins depend on low serving costs and the accounting boundary being used. In its February 2020 essay, a16z described a software benchmark of 60 to 80 percent and up, compared with anecdotal AI-company margins of 50 to 60 percent. Treat 90 percent as a high-margin SaaS assumption to test, not a universal historical standard or a promised AI result.
Martin Casado and Matt Bornstein's February 2020 essay treated the AI margin estimate as anecdotal and doubted that product or go-to-market optimization would fully close the gap. In October 2023, the Wall Street Journal reported that Microsoft was losing more than twenty dollars a month per GitHub Copilot user on early-2023 numbers against a ten dollar price. That figure came from one anonymous source. In a November 8, 2023 Semafor interview, GitHub’s chief executive reported positive Copilot margins; the article also states that the company denied the earlier loss figures. These are historical claims, not a current Copilot margin benchmark. Newer survey projections do not verify them.
What AI gross margins actually look like in 2026
ICONIQ's January 2026 snapshot, drawn from a Q4 2025 survey of roughly three hundred software executives, groups respondents by primary differentiator. Executives emphasizing the application layer projected about 45 percent AI product gross margin for 2026; those emphasizing proprietary model work projected 49 percent; balanced builders projected 53 percent. These groups do not establish that a particular architecture causes the margin. The chart has 269 responses and reports expectations, not completed-year results. The separate July 2026 edition, based on a Q2 survey projects a 53 percent aggregate margin for 2026 and 59 percent for 2027. Its aggregate cannot be merged with January's subgroup comparison.
Use these as dated survey context, then explain your own margin. The investor guide covers the January aggregate forecast and Bessemer's distinct growth archetypes; none is a universal Series A margin cutoff.
The cost side explains the gap without any mystery. ICONIQ's earlier Builder's Playbook, published in June 2025 from an April 2025 survey, tracked monthly inference spend climbing from about $100K before launch to between $1M and $1.6M at general availability, and $1.1M to $2.3M once the product was scaling. These are dated survey ranges, not estimates of what every product spends today. One product executive at a nine-figure-revenue AI company put it in a sentence I could not write better. "The subscription model is not working for us. Power users tend to use a lot resulting in negative margins considering LLM API costs." That is a flat-price company discovering, in production, exactly the decoupling I laid out in usage-based vs seat-based pricing. In the same survey, 37 percent of AI companies said they planned pricing changes.
Bessemer's August 2025 research describes selected high-growth companies following different strategies. Some accept thin margins for rapid growth, while others grow with margins near 60 percent. Their funding and retention matter to whether that choice is viable.
The margin fight: Bessemer and a16z, eight days apart
The disagreement concerns how to interpret current margins when growth and future serving costs are uncertain.
Bessemer moved first, on August 13, with research favoring the cohort that grows fast while keeping its margins healthy, treating margin health as evidence the business underneath is real. Eight days later, two a16z general partners, Sarah Wang and Martin Casado, published a rebuttal whose title told you everything about its posture: "Questioning Margins Is a Boring Cliche." Their argument: "lower gross margins at a moment in time are not a long term indicator of a lack of a sustainable business model." They pointed at Amazon, Netflix, and Uber, all margin horror stories at the equivalent stage, all fine now, and argued that retention and expansion tell you more about an AI company than its current COGS line. They also noted that inference costs had dropped 10x to 100x or more in eighteen months, a range they attached to no specific benchmark, though it sits conservatively inside what independent measurement has found.
On the October 27, 2025 Run the Numbers episode (10:42), Sarah Wang described unusually high AI gross margins as an "orange flag." Her point was about what a company might be leaving unbuilt or underfunded, not a rule that an 85 to 90 percent margin is always bad.
The firms weigh current margin differently. A founder still needs to explain why delivery costs produce today's margin and what would change it. That requires the cost, pricing and retention work covered in the AI commercialization guide.
Two cost curves to test against your workload
Falling prices for a fixed level of capability can reduce serving costs. Metering cost per customer covers historical benchmark-price analyses and the account measurement needed to see whether your product benefits.
MIT FutureTech's "The Price of Progress" working paper distinguishes falling prices for fixed benchmark capability from the price of running the newest frontier models. In its measured period, fixed-capability prices fell 5x to 10x per year, while frontier-model prices rose 3x to 18x per year. The paper was first posted in November 2025 and revised in March 2026; these are measured trends, not a forecast for every future model or product margin.
If a cheaper model meets the customer's quality requirements, routing, caching or distillation can lower serving costs. Test the result and include retries, retrieval and any required review. Products needing frontier capability remain exposed to its price and token demand, but that does not determine margin by itself. The charge to the customer, efficiency and workload also matter, which is why AI pricing has to respect compute reality.
How to build an AI gross margin on purpose
Use four decisions to build and explain the margin plan.
First, decide your posture. Thin margins in service of a land grab and healthy margins in service of durability are both legitimate strategies. Drifting between them because nobody chose is not a strategy. Whichever you pick, you should be able to say it out loud to an investor without flinching.
Second, meter the truth. You cannot engineer a number you cannot see, and a blended cloud bill hides more than it shows. Every account needs its own cost line, which is the instrumentation work I detail in metering cost per customer.
Third, test routing where quality allows it. Wang and Casado list model tiering and query routing as margin levers. Send simpler work to a cheaper model when it meets the customer's quality requirements, then measure the saving after retries and review. Reserve frontier-model spending for work that needs its capability.
Fourth, price against delivery cost. The quoted executive described heavy users producing negative margins on a subscription plan. A platform fee plus usage components can address that exposure if the allowance and rate cover the work. Seats can still work for stable, low-cost usage. The AI commercialization guide connects the pricing and account-cost work.
Get the account costs before setting a new rate. Without that measurement, a pricing change can still leave heavy users below cost. Choose the growth posture, test delivery changes and price with the resulting account evidence.
A period of low margin needs enough funding and a testable reason to accept the loss. State what the spending is meant to achieve, what evidence would support it and when you will change the plan. Do not assume a venture firm's growth strategy fits your company.
When you take margin to a board, bring its trajectory and the costs behind it. A 45 percent margin with a tested routing plan, account costs and two quarters of movement gives the board more to assess than the same percentage on its own. Explain what changed and which assumption the next period will test.
The AI commercialization guide connects this margin work to pricing, customer demand and ownership.
Frequently asked questions
Why do AI companies have lower gross margins than SaaS companies?
AI inference adds a variable serving cost. Retrieval infrastructure and human review add cost where the product uses them, and heavy workloads can multiply the bill. The size of the difference depends on workload, model choice, delivery process and price; not every AI request needs a human review.
Where did the 90 percent SaaS gross margin benchmark come from?
Low marginal serving costs can support high software margins, but 90 percent was not a universal SaaS result. Andreessen Horowitz's February 2020 AI-economics essay described a software benchmark of 60 to 80 percent and up. Test any high-margin assumption against your own delivery costs and accounting scope.
What do AI gross margins actually look like in 2026?
ICONIQ's January 2026 survey projected AI product margins of 45, 49 and 53 percent for groups defined by their primary differentiator. These are self-reported forecasts, not observed full-year results or a causal architecture comparison. Its separate July edition projected a 53 percent aggregate margin for 2026 and 59 percent for 2027. Bessemer's selected high-growth 2025 companies had different margin strategies, including much lower margins.
Do AI gross margins improve as inference costs fall?
Cheaper inference can improve margin when quality holds and other serving costs and pricing do not offset the saving. MIT FutureTech's working paper measured falling fixed-capability prices and rising frontier-model prices over its study period. Routing, caching and distillation can reduce exposure, but benchmark price trends alone do not predict a product's margin.
What counts as COGS in an AI product?
Include costs attributable to delivering the product under your accounting policy: model inference, relevant retrieval infrastructure, hosting and human review where used. Allocate shared serving costs consistently. Andreessen Horowitz's February 2020 estimate, from the pre-LLM era, put cloud compute alone at 25 percent or more of revenue for AI companies, with human-in-the-loop work adding another 10 to 15 percent.
Is a 50 percent gross margin disqualifying for an AI company?
The cited research does not establish one investor margin cutoff for every stage. Bessemer emphasizes margin-healthy growth, while a16z partners argue that current margin alone does not determine sustainability. Sarah Wang's Run the Numbers comment concerns the tradeoff behind unusually high AI margins. Explain the cost, pricing and growth plan behind your own trajectory.
Sources
AI product gross margins by year and primary differentiator: ICONIQ Growth, "State of AI: Bi-Annual Snapshot, The Execution Era of AI" (January 2026; Q4 2025 survey of roughly 300 software executives, N=269 on the margin chart; margins are on AI products, self-reported, and the 2026 figure is a projection). Inference spend by stage and the operator pricing quote: ICONIQ Growth, "The Builder's Playbook: 2025 State of AI" (June 2025, April 2025 survey, N=221 on deployment costs). The margin debate: Bessemer Venture Partners, "The State of AI 2025" (August 13, 2025) and Sarah Wang and Martin Casado, "Questioning Margins Is a Boring Cliche," Andreessen Horowitz (August 21, 2025); Wang's "orange flag" comment is from Run the Numbers, October 27, 2025, at 10:42. Historical baseline: Martin Casado and Matt Bornstein, "The New Business of AI," Andreessen Horowitz (February 2020; the 50 to 60 percent figure is explicitly anecdotal). Cost curves: Gundlach, Lynch, Mertens, and Thompson, "The Price of Progress," MIT FutureTech (working paper, first posted November 2025 and revised March 2026; measured April 2024 to November 2025). GitHub Copilot losses: Tom Dotan and Deepa Seetharaman, "Big Tech Struggles to Turn AI Hype Into Profits," The Wall Street Journal (October 2023; single anonymous source, early-2023 data, and company denial and CEO margin interview via Semafor the following month). Survey figures throughout are self-reported by executives, not audited results.
About the author
Jeff Brokaw is a Certified Chief AI Officer and technical-commercial operator who rebuilds the commercial layer of technically complex companies. He has been building AI systems commercially since 2016. He was the first marketing hire twice, at companies tied to $850M+ in associated exits. He reframed institutional Bitcoin mining as energy-backed digital infrastructure; a $114M institutional raise came together in under thirty days. He runs a multi-model agent system with cost-aware routing, reserving more expensive models for work that needs them.