AI companies are routinely valued at revenue multiples that would be absurd in most software, and the usual explanation — that investors expect enormous growth — is only half of it. The other half is that the revenue being multiplied is not all the same kind of revenue.
What a multiple is actually pricing
A revenue multiple is a compressed forecast. It encodes how fast the business is expected to grow, how much of each new pound stays as gross profit, and how likely the customer is to still be there in three years. Two companies with identical revenue can justify very different multiples if those three inputs differ.
The AI-specific wrinkle is the second input. Classical SaaS carries gross margins in the high eighties because serving one more customer costs almost nothing. A product that runs inference on every request does not work that way: usage has a real marginal cost, and it scales with how much customers use the thing.
So the same headline revenue can represent quite different economics depending on how much of it survives the inference bill. A multiple applied to revenue without asking that question is measuring the wrong quantity.
The questions that separate similar-looking companies
- Does gross margin improve with scale, or hold flat? Falling model costs and better batching push it up; a product whose usage grows faster than its efficiency gains sees it drift down.
- Is the revenue contracted or consumption-based? Consumption revenue can evaporate quietly, without anyone cancelling anything.
- How much comes from a handful of customers? Concentration is invisible in a growth rate and decisive in a downturn.
- What happens if the underlying model gets much cheaper — or much better? Both are opportunities for some companies and existential for others, depending on where the product’s value actually sits.
Where the durable value tends to sit
The pattern worth watching is whether a company’s advantage would survive its model supplier releasing something twice as good for half the price.
Businesses whose value comes from proprietary data, from being embedded in a workflow people cannot easily leave, or from distribution they already own tend to benefit — their costs fall while their moat is untouched. Businesses whose value is mostly a thin layer over a general-purpose model tend to find that the improvement arrives as competition rather than as a gift.
That distinction is not visible in a revenue multiple. It is the thing the multiple is trying, imperfectly, to guess at.
Nothing here is investment advice; it is a description of how these numbers are commonly constructed and what they leave out.
