The margins of AI applications
Traditional SaaS businesses ran on extraordinary gross margins — around 80%. AI application companies typically sit lower, often between 50% and 60%.
As the CEO of an AI application company, I think about this a lot. Token costs are collapsing — nearly an order of magnitude a year. But usage is climbing even faster. So inference costs stay high.
Which raises the real question: do we fight to drag margins back to SaaS levels as fast as we can? Or do we build a phenomenal business at a lower margin?
Here is where I have landed.
1. Don’t wait for SaaS margins to come back.
They might. Some of the heavy processing will eventually move onto local devices, and that could lift margins over time. Treat that as upside — not as the plan. Build to win at today’s cost structure, where 60–70% is a disciplined, realistic ceiling. That is more than enough. Some of the most valuable companies on earth live well below the SaaS line — Apple, LVMH, Roche — businesses with real delivery costs and ferocious pricing power. A lower margin has never stopped a great company.
2. Chase gross-margin dollars, not percentages.
The dollars you generate matter more than the ratio. A fast-growing business at 65% throws off far more capital to reinvest than a stagnant one clinging to 80%. Boards that fixate on defending a margin percentage may end up defending a shrinking company.
3. Price for outcomes, not access.
Solve the customer’s actual problem, and let your pricing prove it. Charging on results aligns your team with what truly matters and makes your value undeniable to the buyer. It is also how you build a moat. Becoming integral to a customer’s success is far more defensible than selling a feature list or access to a model.
4. Hold the line on tokens.
It is dangerously easy to burn compute on things that demo well — routing every prompt to the most expensive model, spinning up real-time avatars, stuffing in context nobody needs. Every product review should ask two questions. Does this improve the customer’s outcome, or does it just make a better demo? And is the team doing the unglamorous, vital work of driving token usage down?
The bottom line
The companies that win the next three years will not be the ones guarding 80% margins. They will be the ones who accept the real cost structure, price on outcomes, and stay ruthless about every token they spend.
Originally posted on LinkedIn.