Gross Margin in AI-Native SaaS: Where the Ceiling Comes From

Classic software margins improve with scale because the costs of goods are mostly fixed and get divided by a larger number. Inference is not fixed. It is strictly proportional to use, and a proportional term divided by a larger number is the same number.

Two shapes of cost of goods

Write cost of goods for N customers each using u units:

COGS(N) = F + N * u * c

  F   the fixed part: baseline infrastructure, monitoring,
      the support organisation's floor, the on-call rota
  c   the variable part per unit: inference, embeddings,
      vector queries, moderation passes, and any human
      review that touches output

margin(N) = 1 - (F / N + u * c) / R

The first term inside the bracket, F / N, is where every traditional software margin story comes from. Grow N and it tends to zero. A business whose costs are all in F has a margin that climbs toward one hundred percent and is limited only by support and hosting overhead.

The second term, u * c, does not move. It is the same for the first customer and the millionth. Growth does not touch it, funding does not touch it, and there is no operating leverage available against it whatsoever.

The ceiling growth cannot lift

As N grows without bound:

  margin -> 1 - (u * c) / R

Assume R = $40 per customer per month and a fully
loaded variable cost of $0.015 per request:

  u =   80 requests :  1 - 1.20 / 40  = 97.0%
  u =  400 requests :  1 - 6.00 / 40  = 85.0%
  u = 1200 requests :  1 - 18.00 / 40 = 55.0%
  u = 2667 requests :  1 - 40.00 / 40 =  0.0%

Every one of those is the *asymptote*. No amount of
scale improves any of them.

This single result explains most of what people find puzzling about the financial profile of AI-native companies. They are not badly run and they are not failing to find leverage. They have a term in cost of goods that is structurally immune to leverage, and the only questions are how large u * c is relative to R and whether u is growing.

And u usually is growing, which is the uncomfortable part. Every successful product initiative raises engagement. Under this cost structure, a rise in engagement moves you down the list above. A business can grow revenue, delight its customers, and watch its gross margin decline every quarter as a direct consequence of doing well — which is a genuinely new experience for software.

Two clarifications, because this argument is easy to over-apply. It concerns gross margin, not viability: a durable business can be run at a margin far below the software norm, as distribution businesses have always been. And it says nothing about the level of c, which has been falling and may keep falling for the reasons in the decomposition of a token price. The ceiling is structural; its height is not.

What the ceiling does to lifetime value

Margin does not stay in the gross-profit line. It propagates into every decision downstream, because acquisition is funded out of gross profit:

LTV = ARPU * m / churn

Assume ARPU = $40 per month and 2% monthly churn:

  m = 0.90  LTV = 40 * 0.90 / 0.02 = $1,800
  m = 0.70  LTV = 40 * 0.70 / 0.02 = $1,400
  m = 0.55  LTV = 40 * 0.55 / 0.02 = $1,100

At a fixed LTV-to-CAC target of 3:1, the affordable
acquisition cost falls in exactly the same proportion:

  $600  ->  $467  ->  $367 per customer

That is the real consequence and it is competitive, not accounting. Two companies chasing the same customer, one with a fixed-cost product and one with a metered input, can afford materially different bids for the same click, the same conference stand, the same salesperson. The margin ceiling becomes a ceiling on how loudly you are able to go to market, and no amount of efficiency elsewhere substitutes for it.

It also means the usual comparison of headline recurring revenue across companies is not a comparison of anything. Revenue at a fifty-five percent margin and revenue at a ninety percent margin buy different amounts of everything. Where you need to compare, compare gross profit, and say so.

There is a second-order effect on churn that is easy to miss. Because margin scales the numerator of lifetime value and churn divides it, the two interact: a lower margin makes retention worth proportionally more, since every month of survival now yields less gross profit and you need more of them to recover the same acquisition cost. Under a metered cost structure, therefore, work on retention competes better against work on acquisition than the same comparison would suggest in a high-margin business — and the heaviest users, who are the most expensive to serve, are frequently also the least likely to leave, which is a genuine tension rather than a tidy conclusion.

Turning variable cost back into fixed cost

There is exactly one structural way to restore scale economies, and it is to convert the variable term into a fixed one — reserved capacity, a committed-use contract, or serving a model yourself on hardware you have already paid for. Then the term you divide by N gets bigger and the term that resists division shrinks.

Per-unit API:      cost = V * c
Fixed capacity:    cost = K            (up to a throughput ceiling)

Break-even volume: V* = K / c

Assume dedicated capacity at K = $9,000 a month and
an API price of c = $0.015 per request:

  V* = 9,000 / 0.015 = 600,000 requests per month

Below V*, the API is cheaper and infinitely more
flexible. Above it, the fixed option is cheaper and
gets cheaper per unit with every further request.

What the break-even hides is that the fixed option costs you the same on a quiet month, obliges you to run the capacity, and pins you to one model at a moment when the useful models change often. The honest version of this decision — including the operational costs that do not show up in the break-even — is worked through in the self-host versus API comparison. Note also that V* falls whenever the fixed price falls or the API price rises, so this is a decision with an expiry date, worth re-running rather than deciding once.

Reporting it so it stays honest

• Put everything variable above the line. Inference, embeddings, vector storage and query, moderation, and any human review of output. Human review in particular is often parked in operating expenses, where it flatters gross margin and hides the fact that it does not shrink with scale either.
• Report margin by cohort and by plan, not blended. A blended figure is an average over a mix that is moving, and it moves most when you succeed at selling to the heavy segment.
• Track u as a first-class metric. Units per customer per period is the variable that drives the ceiling. If it is not on the same dashboard as revenue, the margin decline will be explained after the fact instead of forecast.
• Watch the tail share. The fraction of total units consumed by the top one percent of customers determines whether any flat plan is survivable — the arithmetic is in the unit-economics model.