Mitigating the AI tax on margins

The core strategic issue is that AI coding products do not behave like classic SaaS, because every useful action can trigger real model spend. A seat can look profitable on light chat and simple edits, then turn expensive when the product is reviewing large repos, running tests, or executing long autonomous jobs. That makes routing, BYOK, and usage based pricing central margin controls, not add on features.

• Factory is built to shift expensive work off its own P&L where possible. Its router can send simpler tasks to cheaper models, and its BYOK setup lets customers plug in their own OpenAI, Anthropic, Bedrock, open source, or local models, so Factory is not paying for every token itself.
  2 factory 3 factory 4 factory
• The category benchmark is far below normal software margins. Replit was reported around 23% gross margin in 2023, and Lovable around 35%, because coding agents repeatedly call frontier models and often bundle other variable costs like infrastructure and payments into delivery.
  5 theinformation 6 sacra
• Factory is also aiming higher in the stack than a simple coding copilot. It sells autonomous engineering agents for coding, testing, review, and maintenance, which gives it more surfaces to meter by workload instead of hiding heavy usage inside a flat seat price.
  1 sacra

Over time, the winners in AI developer tools are likely to look less like pure seat based SaaS and more like software plus compute orchestration. Companies that can steer work across models, push spend onto customer controlled infrastructure, and charge more directly for heavy autonomous execution should widen margins as usage scales.