Ninety cents of every revenue dollar: that is what Will Manidis says the median bootstrapped business-software company spends, and he argues AI is about to make most of that spending optional. His essay appeared on X on 5 October 2026, though the post itself says it was first published on 16 January 2024. Read it as a long-running thesis, not fresh analysis.
The numbers are the most useful part. Manidis splits that 90 percent into five buckets: 25 percent on selling, 24 percent on building the product, 15 percent on administration and odds and ends, 13 percent on hosting and setting up customers, and 10 percent on keeping them. He gives no source for these figures. His point is that each bucket used to mean salaries, and salaries are costs a company commits to before it knows what revenue will arrive.
That is the distinction the whole argument rests on. A fixed cost is paid whether or not customers show up, like a developer’s wage. A variable cost is paid only when work happens, like a per-task fee to a model provider. The old software model paid mostly fixed costs and sold licences on top, so every extra customer was nearly pure profit once the team was covered. If the work shifts to services billed per use, profit per customer stops being a built-in feature of the business and starts moving with how heavily customers use the product.
Manidis does not spell out the consequence for valuation, so here is the reading. Investors have paid high multiples of recurring revenue partly because they assumed margins stay put as revenue grows. He notes that the median public software company is priced at a five-times multiple of yearly recurring revenue, against more than ten during much of the spending boom since 2018. If costs track usage, a revenue dollar is worth less certainty, and those multiples have less to stand on.
He offers small, concrete signs that the shift had begun before generative AI. A basic Django web app can be bought from a Replit bounty for $650. Companies can sell through resellers such as AWS instead of hiring sales staff. Zapier and Intercom, he says, now often charge by the task rather than by user per year.
His claim about models is that they act as junior staff on demand. Software organisations have traditionally looked like a pyramid, with many junior developers under a few senior ones. He imagines them looking more like marketing teams, where senior people direct automated systems. He concedes the catch himself: unlimited junior workers also produce unlimited junior mistakes, and he does not say how hallucinations get supervised at scale.
His own example cuts against the headline promise. He says GitHub Copilot, owned by Microsoft, had about $100 million in recurring revenue yet lost around $20 per user each month on compute and training data. That is a variable cost larger than the price. Per-use costs make a business flexible, but they also make it exposed when usage is heavy and pricing is flat.
The essay then moves to financing. Manidis expects founders to pick steady cash flow over rapid growth, and to prefer credit, warrants or loans to venture capital. He closes by disclosing that he is a small investor in Lafayette Standard Co, a company founded by John Kennedy, the founder the essay discusses alongside its author, built to act on this thesis. That makes him an interested party.
What the essay does not do is show the shift has happened at scale. It forecasts, using Carlota Perez’s model of technology booms, and its figures date from 2024. No data here tracks how many companies have actually cut staff for model spending, or what happened to their margins.
For anyone pricing a software product this quarter, the practical test is to list which costs rise with each extra customer request, then check whether the price charged rises with them.
Will Manidis (@WillManidis), post on X, 5 October 2026.
