The cloud bill and the margin
Software margins are under pressure from infrastructure and inference costs, but the pressure is uneven and the industry median has not moved. That gap between the aggregate and the cohort is the thing to examine, because a healthy blended margin can hide accounts that lose money.
The 80% gross margin belonged to a specific technical era, when replication was close to free and cloud infrastructure had finished commoditizing. Whether that era is ending is a question people answer too quickly in both directions.
The aggregate numbers are calm. Benchmarkit’s 2026 metrics report finds median software gross margin holding above 80% and stable across 4 years, and says plainly that industry-wide infrastructure costs have not yet compressed software margin at the median. Anyone claiming a general margin collapse is ahead of the data.
Particular cohorts are not calm. The 2026 Aleph and Benchmarkit benchmarks put usage-only pricing models at 62% gross margin against the 80% software median, because compute cost scales with revenue rather than sitting against a fixed subscription. ICONIQ’s 2026 State of AI report, drawn from around 300 executives at companies building AI products and published in July 2026, records average gross margin for those products at 45% for 2025, with respondents expecting 53% in 2026 and 59% in 2027. Those forward figures are expectations rather than results, which is worth holding onto: they describe what builders think will happen to their own costs.
The arithmetic at the unit level is easier to trust than any forecast. Ben Murray of The SaaS CFO worked a straightforward case: add an assistant feature to an 80 dollar per month seat, and inference, routing and supporting infrastructure can add roughly 15 dollars of direct variable cost, taking gross margin on that seat from 80% to nearer 65%.
Why a blended number hides the problem
Because cost to serve varies by account and revenue does not follow it.
A company with a healthy company-wide margin can contain a set of accounts running at a fraction of it, driven by heavy usage, unmanaged data retention, custom infrastructure or bespoke integrations agreed years ago at a price that no longer covers the cost. Nobody notices, because nobody calculates margin per customer. The account is large, visible and celebrated, and the rest of the base is paying for it.
This is not universal. Plenty of companies at this size have flat cost to serve across the base and nothing interesting to find. The point is that the blended figure cannot tell you which kind of company you are looking at, and the blended figure is usually all that is offered.
Why this is board work
Because the fix is architectural and commercial at once, and it takes several quarters.
The commercial half is repricing accounts that do not cover their cost, which is a negotiation with a customer who currently believes they have a good deal, and which cannot be done to all of them at once without a churn event.
The technical half is cost attribution before optimization. You cannot manage cost to serve that you cannot see, and many companies at this size cannot attribute infrastructure cost to a customer at all.
Neither half completes inside 100 days, and both drift the moment nobody is asking. The measurement that keeps it honest is gross margin by customer cohort, reviewed quarterly, against the plan agreed at close.
More insights
The tender was written for someone
In B2B software the best product does not always win the deal. Most buyers have a favorite before they speak to any vendor, requirement documents tend to describe the product the authors already know, and in many markets relationships and local presence count as much as features. For a buyer of a SaaS company, that changes how a pipeline and a win rate should be read.
Sold through someone else
A B2B2B software company sells to a partner, and the partner then has to sell to its own business customers. That puts 2 sales cycles, an onboarding project and a revenue share between the product and the revenue. The model can reach further than a direct team ever could, but a plan that counts signed partners as if they were selling will overstate it.
Last updated