The 80% gross margin is gone, and diligence already knows

Software got funded for thirty years on one number. Eighty percent gross margin, sometimes ninety.

That number is why a dollar of software revenue was worth more than a dollar of any other revenue.

AI-native products are averaging closer to 52 percent in 2026. The old assumption is not a little off. It is a different business.

Before the argument, a word on those numbers. The margin figures are averages put together by analysts. They are not filed, audited or standardised.

Different companies decide for themselves what belongs in cost of goods sold. So the gap between roughly 52 and roughly 80 is the finding here. The decimal is not.

Why the number moved

Serving software costs almost nothing per user. Serving a model costs something every single time.

Every request buys tokens. A heavy user is now genuinely expensive to keep.

That cost sits in cost of goods sold, where it eats the margin directly. It does not sit in engineering, where you could argue it away as investment.

I have watched this land inside our own automation work. The model was never the expensive part of the decision. The volume was.

What investors changed on their side

Several public companies started breaking inference cost out separately in Q1 2026 filings. It runs at four to nine percent of revenue for many of them.

The ones who disclose it get credit for it. The ones who fold it into a line called infrastructure get questions instead.

That habit has moved down into private diligence. Assume the question is coming.

There is a second half to this, and it works in your favour. Inference for a given level of capability has collapsed in price over three years.

Investors are underwriting that curve, not just today’s snapshot. Which means the honest version of your numbers is defensible.

Bring the breakdown before they ask for it

A founder who volunteers this looks like an operator. A founder who gets asked and fumbles looks like someone who has not run the numbers.

Same data, completely different meeting.

You need three things in the data room. The split, the ratio, and the direction.

The prompt

Point this at your provider invoices and your revenue by month. It writes the page.

You are preparing the cost of goods sold disclosure for an AI product
company going into fundraising diligence.

Here is my monthly data for the last 12 months:
[PASTE: month, revenue, model or inference spend, hosting spend,
third party data spend, human review or annotation spend]

Produce, in plain language and no more than one page:

1. A table: month, revenue, total COGS, gross margin percent, and
   inference spend as a percentage of revenue.

2. The trend line in one sentence each for gross margin and for
   inference as a share of revenue. Say "worsening" if it is worsening.

3. Cost per unit of work: divide inference spend by [YOUR UNIT: calls,
   documents processed, tickets resolved]. Show it for month 1 and the
   latest month.

4. The three months that look worst, and for each one, the single
   question an investor would ask about it.

5. A list of anything I have NOT given you that belongs in this table.

Rules: never estimate a missing month, mark it UNKNOWN. Do not smooth
or average away a bad month. If gross margin is below 50 percent, say so
in the first line rather than at the end.

The question underneath the question

Nobody actually cares about 52 versus 80. They care whether the cost falls as you grow.

If serving your hundredth customer costs the same as your first, you have a services business wearing software pricing.

If the unit cost is dropping because you route cheap work to cheap models, you have the other thing. Say that, with the numbers behind it.

The trend beats the level. It always did.

The COGS prompt is on the resources page. Free, no email required.