Unit Economics of AI Startups: Why Gross Margins Matter More Than Growth

Traditional SaaS companies enjoy 70–85% gross margins — once the software is built, serving an additional customer costs almost nothing. AI-powered products break this pattern because every customer interaction triggers real compute cost. A chatbot product charging $50/month might generate $30–40/month in LLM API costs per active user, leaving gross margins of 20–40% — closer to a managed services business than traditional SaaS. This fundamentally changes the economics. Growth-at-all-costs strategies that work in SaaS fail in AI because unprofitable customers bleed gross margin instantly rather than just delaying payback. The companies winning long-term are those obsessing over cost per inference and engineering around it. Key levers include: prompt engineering to reduce token usage (shorter prompts, structured outputs), model routing (send easy queries to cheap models, hard queries to expensive ones), caching (reuse responses for repeated queries), fine-tuning smaller open-source models for specific workloads instead of paying frontier API prices, and batching requests to improve throughput. OpenAI's Batch API, Anthropic's prompt caching, and provider volume discounts are all strategies teams use to manage gross margin. The diligence question every AI founder and investor should answer: what does gross margin look like at scale once pricing, usage patterns, and model costs normalize? Companies without a credible answer are not durable businesses regardless of their growth rate.