Traditional SaaS has near-zero marginal cost per user: serving an additional customer costs almost nothing. AI SaaS has a significant marginal cost: each additional query consumes LLM tokens, embedding compute, and retrieval resources. This changes the unit economics fundamentally. Understanding AI SaaS cost structure is essential for pricing, margin analysis, and scaling decisions.
The AI SaaS Cost Components
AI SaaS costs have five components: (1) Development (one-time): building the product, $200K-$1M+ depending on complexity. (2) Infrastructure (recurring): compute, storage, networking, $1K-$10K/month. (3) AI inference (usage-based): LLM tokens, embedding compute, $0.001-$0.10 per query depending on model and context. (4) Data pipeline (recurring): ingestion, embedding, indexing, $500-$5,000/month. (5) Operations (recurring): monitoring, support, incident response, $2K-$10K/month. The key difference from traditional SaaS: AI inference is a variable cost that scales with usage.
Inference Cost Per Tenant
Track inference cost per tenant: total tokens consumed × price per token. This reveals: which tenants are profitable (high revenue, low inference cost), which tenants are unprofitable (low revenue, high inference cost), and which features drive the most cost (generation vs classification vs extraction). Use this data for: pricing optimisation (increase prices for high-inference tenants), feature optimisation (route high-cost features to cheaper models), and tenant management (set usage limits for unprofitable tenants).
Warning
If you do not track inference cost per tenant, you will discover unprofitable tenants too late. Track from day one.
Margin Optimisation
AI SaaS margins are lower than traditional SaaS (typically 60-80% vs 80-90%) due to variable inference costs. Optimise with: model routing (use cheaper models for lower-value tasks), semantic caching (avoid redundant inference), prompt optimisation (reduce token count), usage-based pricing (pass costs to tenants), and tiered features (different models for different tiers). Target: inference costs should be under 15% of revenue. If inference costs exceed 20% of revenue, your pricing or your architecture needs adjustment.
Conclusion
AI SaaS has different unit economics than traditional SaaS: variable inference costs that scale with usage. Track inference cost per tenant, optimise margins through model routing and caching, and price based on actual cost structure. Target: inference costs under 15% of revenue.
Key Takeaways
- AI SaaS has variable inference costs — this is the key difference from traditional SaaS
- Track inference cost per tenant from day one — identify unprofitable tenants early
- Target: inference costs under 15% of revenue — above 20% requires pricing or architecture changes
- Optimise with: model routing, semantic caching, prompt optimisation, usage-based pricing
- AI SaaS margins are 60-80% (vs 80-90% traditional SaaS) — plan for this in financial models