I was on a call last month with a founder running an applied-AI company — [redacted Series A AI workflow startup] — and he told me his lead investor had passed. The reason confused him. The investor said the unit economics were too thin "given the model spend." When the founder pushed for specifics, the investor said something like: "you are over the thirty-percent line." Then the call ended.
That phrase — over the thirty-percent line — is a heuristic that AI investors say constantly to each other and almost never explain to founders. It comes up in pricing conversations, gross-margin debates, fundraising decks, and acquisition diligence. Most of the founders I work with have heard it referenced and have no idea what it actually means.
So here is the version I share with my own clients, drawn from what I have learned watching investors apply it across pre-seed, Series A, and Series B AI deals.
What the 30% rule actually says
The thirty-percent rule is a back-of-the-napkin test for whether an AI company has durable margins. The crude version is this: if your model and infrastructure costs are eating more than thirty percent of your revenue, investors get nervous. If they are eating less than thirty percent, you are inside the band most AI investors use to underwrite gross-margin assumptions.
The reason the heuristic exists is that AI companies have a cost structure traditional SaaS investors are not used to underwriting. A normal SaaS business has near-zero marginal cost per customer. Compute is essentially free, and gross margins routinely sit between seventy-five and ninety percent. AI companies are different. Every customer query carries a real, recurring cost — inference compute, model API calls, vector storage, retrieval. Those costs scale with usage in ways traditional software does not.
So when an investor says "what are your AI costs as a percentage of revenue," they are testing whether your business model survives at scale or whether you are subsidising every customer with venture money. Thirty percent is the threshold most of the experienced AI investors I work with treat as the upper bound for a healthy company. Below that, your unit economics resemble software. Above it, they resemble a managed service or a thin-margin reseller.
Why the threshold sits at thirty
The number is not arbitrary, though it is approximate. The logic comes from how investors back into target gross margins.
If your AI infrastructure is consuming thirty percent of revenue, your headline gross margin starts at seventy percent. After you add the rest of your cost of revenue — customer support, hosting, payment processing, data costs — you typically land somewhere between fifty-five and sixty-five percent gross margin. That is the lower edge of what investors will tolerate to call something "software-like."
If your AI costs are at fifty percent of revenue, your gross margin compresses to somewhere between thirty-five and forty-five percent. At that point, you are no longer being underwritten as a SaaS business. You are being underwritten as a services business, an outsourced ops provider, or a thin-margin platform. Your revenue multiple drops accordingly. The same revenue gets you a meaningfully smaller valuation.
That is what the investor on my client's call meant by "over the thirty-percent line." He was not saying the business was bad. He was saying the multiple he could justify had collapsed.
The mistake founders make in the spend itself
Most of the AI founders I review are not over the thirty-percent line because their product is fundamentally unprofitable. They are over it because they are using foundation-model APIs for work that does not need foundation-model intelligence.
I see this constantly. A founder builds an MVP on the most capable model available, ships it, and never goes back to optimise. Three months later they are spending forty-five percent of revenue on inference. The product works, customers love it, but every dollar of growth makes the unit economics worse.
The fix is rarely a different business model. It is technical. Move routine work to smaller, cheaper models. Cache responses where possible. Use retrieval to avoid sending entire context windows on every call. Run cheaper open-weight models for the eighty percent of queries that do not need frontier intelligence, and reserve the expensive model for the twenty percent that do.
I have watched founders move from forty-eight-percent AI-cost ratios down to twenty-two percent inside a quarter, without changing pricing, without losing customers. The work is unglamorous — prompt engineering, model routing, caching layers — but the impact on valuation is enormous.
How to talk about it with investors
If your AI costs are inside thirty percent, lead with it. Most founders bury this number. I tell my clients to put it on a pricing or unit-economics slide and let the investor see it without asking. It pre-empts the question and signals that you understand the conversation the investor is going to have internally after the meeting.
If your AI costs are above thirty percent, do not hide it. Investors will find out at diligence, and discovering it late is far worse than disclosing it early. The right move is to walk through the trajectory. What were your AI costs six months ago. What are they now. What is your plan to bring them down. What does the gross-margin profile look like at scale.
The investors I respect care less about the current number than the slope of the line. A founder running at thirty-five percent today with a credible plan to be at twenty percent in two quarters is a more attractive bet than a founder running at twenty-eight percent with no idea how the number behaves at ten times the volume.
Where the rule breaks down
The thirty-percent heuristic is a starting point, not a verdict. It does not apply cleanly to every AI business model.
Vertical AI applications with high willingness-to-pay — legal, medical, defence, regulated finance — can sit comfortably at forty-percent AI cost ratios because the value they create per query is enormous. The customer is not buying compute, they are buying outcomes that would otherwise require expensive specialist labour. The AI cost is irrelevant if the customer is happy paying ten times what the equivalent human service would cost.
Consumer AI products often look the opposite. They have low willingness-to-pay, high usage, and relentless cost pressure. A consumer AI app at fifteen-percent AI cost ratio is still a hard business. The thirty-percent rule does not save you when your end users will not pay enough to absorb any cost structure.
Infrastructure-layer AI companies — model providers, training compute, fine-tuning platforms — operate in an entirely different cost world and the heuristic mostly does not apply. They are being underwritten on different metrics.
So the rule is most useful as a screening test for applied-AI companies — the SaaS-shaped, vertical or horizontal AI products that sit on top of foundation models and sell to enterprises or SMBs. That is most of the market. That is most of the founders I work with.
What investors are actually testing
The thirty-percent rule is shorthand for a deeper question, which is: does this business have any pricing power. If your AI costs are eating thirty percent of revenue and growing, an investor wants to know whether you can raise prices without losing customers, whether you can reduce costs without degrading the product, or whether you are stuck.
The companies that escape the trap have one of three things going for them. Pricing power because they own a workflow customers cannot easily replace. Cost economics that get better with scale because they have built infrastructure that gets cheaper per query as volume grows. Or a moat in data, distribution, or product depth that lets them defend gross margins against the next foundation-model price drop.
If you have one of those three, the thirty-percent rule is a guidepost. If you have none of them, the rule is a ceiling, and your valuation will reflect that.
The questions I get asked most
The first one is always: how do I calculate my AI cost ratio. The answer is dollars spent on inference, training, embeddings, retrieval, and fine-tuning, divided by gross revenue from the AI feature. If you offer non-AI revenue too, calculate it on the AI portion only. Investors will.
The second is whether the rule applies to early-stage companies with no revenue. Sort of. At pre-seed and seed, investors are projecting forward. They want to see your modelled AI cost ratio at the volume you are guiding to in eighteen months. Get to thirty percent or below at projected scale.
The third is whether the threshold is moving as foundation models get cheaper. It is. Five years ago the comfortable line was probably forty-five percent. Today it is around thirty. In two years it may be twenty. The direction of travel is what matters — investors expect the trajectory to bend down, not stay flat.
If you want to walk through where your numbers sit and what to do about them, that is the kind of analysis we run for AI founders at Beyond Elevation. The Fractional IP CxO engagement covers this — modelling unit economics, mapping what the IP and data assets are worth, and figuring out what to fix before the next round.
By Hayat Amin — founder, Beyond Elevation. I advise pre-seed through Series B founders on patent strategy, IP monetisation, and turning research into defensible value.