AI Gross Margins: Why the 90% SaaS Benchmark Is Gone

Software margins used to be a birthright. Then every request started arriving with a bill attached.

For twenty years, a software company that could not clear an 80 percent gross margin had some explaining to do. The number was so reliable that nobody diligenced it. It sat in the deck the way gravity sits in a physics problem: assumed and load-bearing. AI companies do not get that assumption anymore, and the two most influential firms in venture spent August of 2025 publicly arguing about what should replace it.

The 90 percent SaaS gross margin benchmark is gone for AI companies because every AI request carries a real variable cost: inference, retrieval infrastructure, and human review. In ICONIQ's January 2026 snapshot, drawn from a Q4 2025 survey of roughly three hundred software executives, AI products are projected to average 45 to 53 percent gross margins in 2026 depending on architecture, and the fastest-growing AI startups run lower still, some negative. Gross margin stopped being a property of the business model and became a number you engineer.

Here is where the old number came from, what the new ones actually are, and what I do about it as an operator.

Where the 90 percent benchmark came from

The old benchmark was never really a number. It was a consequence. Classic software had a marginal cost so close to zero that serving the next customer was free, so once you covered your fixed costs, nearly every new dollar of revenue was margin. The 70s and 80s were structural. The 90s belonged to the best public companies on their best days, and then a thousand pitch decks rounded the whole thing up.

None of this should surprise anyone who was paying attention. Martin Casado and Matt Bornstein at Andreessen Horowitz called it in February 2020, anecdotally, at 50 to 60 percent for AI companies against a 60 to 80 percent software benchmark, and doubted that any amount of 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, a figure sourced to a single anonymous person and denied by GitHub's chief executive a month later. Both are history, and the Copilot number was always thin evidence. They get one line each here because the 2026 data says the shape they described was right.

What AI gross margins actually look like in 2026

ICONIQ's January 2026 State of AI snapshot, drawn from a Q4 2025 survey of roughly three hundred software executives, breaks the projection out by architecture. Pure application-layer AI products project about 45 percent gross margins for 2026, up from 33 percent in 2024. Companies doing proprietary model work project 49 percent. The ones balancing model depth with an application layer project 53, the healthiest of the three. Those are margins on the AI products themselves, self-reported, and the 2026 figure is a projection rather than a result.

Sit that beside the 90 the industry still recites and the gap runs thirty-five to forty-five points. I walk through the aggregate trajectory, and what investors do with it, in what investors look for in an AI startup.

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. That survey is over a year old and the absolute numbers have surely moved, but the shape is the point: the bill grows with success. 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.

The fastest-growing companies run thinner numbers than any of this, on purpose, trading margin for a land grab. Bessemer's August 2025 research split the AI cohort into two archetypes with very different margin profiles, and I break those down, with the retention data that goes with them, in the two AI growth archetypes.

The margin fight: Bessemer and a16z, eight days apart

Here is how unsettled this is: in August 2025, two of the most followed firms in venture published opposite readings of the same reality just over a week apart.

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.

Wang went further on the Run the Numbers podcast and flipped the old religion on its head, calling 85 to 90 percent gross margins at a top AI company "an orange flag." Read that again. The number that used to certify a software business now makes a sophisticated investor squint.

Notice what neither side disputes. Nobody in this fight claims AI margins currently look like SaaS margins. The disagreement is about what the gap means: symptom or strategy. So the question a founder has to answer has changed. You need to explain, mechanically, why your margin is what it is and where it goes next. That explanation is the real deliverable, and producing it is the whole discipline of AI commercialization.

The two cost curves that decide your margin

The optimist case for AI margins rests on a trend nobody disputes: the price of intelligence is collapsing. The cost of a fixed level of capability falls fast and keeps falling, and I put numbers on that curve in metering cost per customer.

The catch is which curve your product actually rides, and a working paper from MIT FutureTech put clean numbers on the distinction late last year. The cost of a fixed capability, the same benchmark score you shipped last quarter, falls 5x to 10x a year. The price of running the newest frontier models rises 3x to 18x a year, pushed up by bigger models and by reasoning workloads that burn far more tokens per task. Both curves are real. They just belong to different companies.

If last year's model quality clears your customer's bar, you ride the falling curve: route the work to older or smaller models, cache what repeats, distill where you can, and watch your COGS melt. If your product only works at the frontier, you are chained to the rising curve, and every model generation raises the floor under your unit costs. Living at the frontier is a strategic position with a price tag, and the tag belongs in your pricing, which is why AI pricing has to respect compute reality.

How to build an AI gross margin on purpose

Everything above points one direction: nobody is handed a margin anymore. You build one, and the build has four parts.

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, route by default and go to the frontier by exception. Wang and Casado list model tiering and query routing among the levers that pull margin back up, and the two-curve math above is the justification: easy work goes to cheap models, and frontier tokens get spent only where the customer can taste the difference. Treat model choice as a cost decision every time, not as a default. Make the expensive model earn its slot.

Fourth, price so revenue moves with cost. Flat seats on top of metered COGS is how that product executive ended up with negative margins on his best customers. A platform fee plus usage-aligned components keeps growth from bleeding you, and the structure is the subject of half this cluster, starting with the pillar.

The order matters more than the list. Most teams jump straight to the fourth part, because pricing is the one that feels like a decision, and they reprice without knowing their cost to serve. That produces a number that is defensible in a meeting and wrong in the P&L. Meter first. The rest of the sequence only works on real numbers.

One more thing worth saying plainly, because the venture argument tends to bury it. Both firms in that August fight are talking about companies whose growth story is big enough to make margin a strategic choice. If you are not that company, the choice is not on the table. Your margin is your business model, and the investor reading your deck will treat it that way.

So when you take the number to a board, bring the trajectory rather than the snapshot. A 45 percent margin with a routing plan, a metered cost line, and two quarters of movement behind it reads nothing like a 45 percent margin with a shrug. That is the August fight compressed: nobody agrees on what the right number is, and everybody agrees you should know why yours is what it is.

None of this is exotic. It is the same cost-to-serve discipline every manufacturer and every marketplace has always lived with, which is exactly why software founders find it insulting. Software got a twenty-year holiday from that math. The holiday is over.

Nobody is going to hand you 90 again. Build the number, meter it, and be ready to show the receipts.

This is one spoke of the AI commercialization guide. Start with the pillar: AI Commercialization: the complete guide.

Frequently asked questions

Why do AI companies have lower gross margins than SaaS companies?

Because every AI request carries a real variable cost. Classic software served one more user for close to nothing, so margins in the 70s and 80s were structural. An AI product pays for inference, retrieval infrastructure, and human review on every use, and heavy users multiply that bill. The gap comes from architecture.

Where did the 90 percent SaaS gross margin benchmark come from?

From the near-zero marginal cost of the SaaS era, rounded up by repetition. Andreessen Horowitz's February 2020 essay on AI economics put the software benchmark at 60 to 80 percent and up. The 90 percent version is the mythologized top of that range, kept alive by pitch decks rather than audited financials.

What do AI gross margins actually look like in 2026?

ICONIQ's January 2026 snapshot, from a Q4 2025 survey of roughly 300 software executives, breaks 2026 projections out by architecture: pure application-layer AI products at 45 percent, proprietary model work at 49 percent, and balanced builders at 53 percent. Those are margins on AI products, self-reported, and 2026 is a projection. The fastest-growing AI startups run lower, some negative.

Do AI gross margins improve as inference costs fall?

Only if the product can ride the falling curve. The cost of a fixed AI capability drops 5x to 10x a year by MIT FutureTech's measurement, while the price of running the newest frontier models rises 3x to 18x a year. Products that route work to older, cheaper models recover margin. Products locked to the frontier do not.

What counts as COGS in an AI product?

Everything it takes to serve one more request: model inference, the retrieval and vector infrastructure behind each answer, hosting, and any human review wrapped into delivery. 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?

No verified data shows investors applying a single margin threshold by stage. The debate is live: Bessemer favors margin-healthy growth, while a16z partners argue that margins at a moment in time do not indicate whether the business is sustainable, and Sarah Wang has called 85 to 90 percent margins an orange flag. Trajectory and architecture count for more than the snapshot.

---

Sources

AI product gross margins by year and architecture: 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 the Run the Numbers podcast. 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, November 2025, not yet peer-reviewed; 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 denied by GitHub's CEO 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 and ships AI in production, not slideware. He runs a multi-model agent system in production, routing work to the cheapest model that clears the bar, which is the same margin discipline this article argues for. He co-founded the AI fintech FaaStrak and took it from zero to $1M ARR in nine months, ran the commercial side of two media companies tied to more than $850M in associated exits, and authored the go-to-market behind a $114M institutional raise that came together in under 30 days.