From SaaS seats to inference businesses

This is the key economic difference between AI software and old SaaS, paid seats usually turn into real variable cost almost immediately. A gym can profit from members who never show up, but an AI product pays for tokens, search, and compute every time a user asks for work. That pushes companies like Adapta to win on routing, caching, and steering usage toward cheaper in house models, not on hoping customers stay idle.

• The interview makes clear that Adapta treats margin discipline as a product and infrastructure problem. The goal is to match the quality a customer could get from Claude, Gemini, or ChatGPT while keeping each task profitable through software architecture and model choice.
  1 sacra
• Cursor and Perplexity show the same pattern at larger scale. Both sell high frequency AI workflows where usage keeps climbing after purchase, and Cursor only reached slight gross margin profitability by April 2026, while Perplexity’s margins depend on keeping model and search costs below subscription revenue.
  2 sacra 3 sacra 4 sacra
• OpenRouter is the clean infrastructure comparable. Its whole value is routing requests across many models with failover and model selection, which is exactly the kind of control layer an AI application needs when there are very few gym ghosts to subsidize heavy users.
  5 sacra

Going forward, the strongest AI application companies will look less like seat based SaaS and more like compact inference businesses wrapped in great workflow software. The winners will be the teams that can keep users active, keep quality high, and continuously shift work onto lower cost proprietary or better routed models as usage compounds.