The story everyone tells about AI and software is margin compression: inference is a real cost, the era of near-zero marginal cost is over, and SaaS gross margins fall. That story is half right and mostly beside the point. The more useful question is relative: in a market that sorts every company into AI or not-AI, who is actually funding the cost of intelligence, and who is just renting it?
Building frontier AI is a capital black hole — training runs, scarce talent, and a compute bill that scales with ambition. A pure-AI company funds that black hole directly. Its entire P&L is the cost of producing intelligence, against prices that fall every few months as the next model commoditizes the last. That is a hard place to hold a margin.
A B2B SaaS company is in a different position. It rents intelligence by the call from providers competing to cut the price, and layers it onto a business that already has distribution, proprietary workflow data, and switching costs. Inference shows up as a new line in cost of goods, but it is a bounded, falling, pass-through cost — not the whole business. The eighty-percent gross margin gets an asterisk, not an obituary.
This is also why incumbents adapted fastest. They did not need to invent the model; they needed to wire it into a workflow customers were already paying for. The hard, slow assets — the integrations, the data, the compliance, the install base — were already built. AI was a feature to bolt on, not a company to fund from zero.
That is the relative-margin argument. Capital allocators are comparing two things: a software business adding AI at the cost of inference, and a pure-AI business whose opportunity cost is funding the model itself. In a world where attention flows to whoever says AI loudest, the quieter SaaS company often keeps the stronger margin precisely because it refused to fund the black hole.
The real risk is not margins; it is disruption. The same dynamic that lets incumbents rent intelligence cheaply also lets AI-native entrants rebuild a workflow without the legacy, and occasionally the model layer reaches up and absorbs the application above it. Renting intelligence preserves margins right up until someone reprices the entire workflow around it. Incumbency buys time, not immunity.
So the durable move is the same one showing up everywhere: rent the commodity, own the layer that is hard to copy. For software that means leaning into the distribution, data, and workflow depth a model cannot rent back, and treating inference as a cost to manage rather than a moat to build. It is the marketplace take-rate story in a different suit — value accrues to whoever owns the part the machine cannot route around; see Agentic Marketplace Disruption: Take-Rates vs the Meter.