What is left of revenue after the direct cost of delivering the product. For AI companies this is where inference and model-serving costs land, which is why their margins can look unlike traditional software.
Why it matters
Gross margin answers a narrower question than most people expect: after paying only the direct
cost of delivering what was sold, how much is left? It excludes sales, marketing, product
development and overhead entirely — those come out further down the income statement. That
narrowness is the point. Gross margin isolates whether the core product is fundamentally
profitable to deliver, before any decision about how much to spend growing it.
For an AI company, this line has become unusually important to read carefully. Traditional
software has a marginal cost close to zero — serving one more customer costs almost nothing once
the product is built. An AI product that calls a model at inference time does not: every request
has a real, variable cost attached to it. That single difference is why some AI companies post
gross margins that look more like a services business than a software one, and why a healthy
top-line number can still hide a product that loses money on heavy usage.
How it works
Gross margin is gross profit divided by revenue, and gross profit is revenue minus cost of goods
sold. The work is in deciding what belongs in cost of goods sold.
For an AI company, that typically includes model inference and API costs, cloud compute for
serving the product, and any third-party licensing tied directly to usage. It typically excludes
sales commissions, marketing spend, and the salaries of people building future features — those
are operating expenses, not the cost of delivering what has already been sold.
Where it gets contested is around engineering costs that support both current delivery and
future development. A company can reasonably allocate infrastructure engineers' time between the
two, but the allocation is a judgement call, and a seller has an incentive to push as much of that
cost as possible below the gross profit line.
Reading the trend matters more than a single figure. A gross margin that improves as revenue
grows suggests the direct cost is spread over more customers efficiently. One that holds flat or
worsens as revenue grows suggests the direct cost scales with usage roughly as fast as revenue
does — the more usage-sensitive shape typical of inference-heavy AI products.
On GetDeal
A listing's Deep Dive tab is where gross margin and the cost figures behind it are presented to
buyers and investors, and the AI analysis report reads those figures — from what the founder
uploads plus public data — when it produces its valuation range and per-section confidence. For
an AI company, that is where a reader can see how much of revenue the direct cost of inference
and compute is actually taking.
Get a free AI valuationthe in-product glossary — the same definitions, alongside your deals
Questions people ask
- Why do AI companies often have lower gross margins than traditional SaaS?
- Because serving each customer request in an AI product typically involves a real, variable cost — the model inference or API call itself, plus the compute to run it — whereas traditional software has almost no marginal cost once it is built. That variable cost sits inside cost of goods sold, so it reduces gross margin in a way that a conventional software product rarely experiences at the same scale.
- What costs get disputed when calculating gross margin?
- The most common dispute is over engineering and infrastructure costs that support both delivering the current product and building future features. A seller may want to classify as much of that cost as possible as research and development, which sits below gross profit, while a buyer will want a more conservative allocation that keeps the direct cost of delivery inside cost of goods sold.
- Does a rising gross margin always mean the business is improving?
- Not automatically. It can also happen because a company shifted its customer mix toward lower-usage accounts, renegotiated a lower-cost inference provider, or reclassified a cost below the line rather than because the underlying product economics genuinely improved. The reason behind the change matters as much as the direction it moved.
See also
More in Metrics
- ARR — The annualised value of subscription revenue that repeats — contracted and expected to continue.
- Burn rate — How much cash the company consumes each month.
- CAC — What it costs, on average, to win one customer — sales and marketing spend divided by customers gained.
- Churn — The rate at which customers or their revenue leave.
- EBITDA — A measure of operating profit that strips out financing, tax and accounting charges for past spending, so two businesses can be compared on how well the operations themselves earn..
Keep reading
Updated 2026-09-09