Unit Economics for AI Startups: Why Standard LTV-CAC Fails
Key takeaways
Based on Chapter 7 of The Two Numbers
- AI startups have variable inference costs that distort traditional gross margin.
- LTV must net out per-customer compute spend — not just SaaS hosting.
- CAC patterns split: PLG can be very low; enterprise AI sales can run $50k+.
- Reforecast unit economics monthly. Costs and pricing move fast.
Quick Answer: Standard LTV-CAC frameworks assume 70–80% gross margins. AI startups often operate at 40–60% because of variable compute, API, and inference costs. Using standard SaaS formulas overstates LTV by 20–40%. The fix: use Contribution Margin LTV that accounts for per-customer variable costs.
In 2024, I reviewed a pitch deck from an AI startup — reviewing decks like this is a big part of how I work with early-stage founders. They showed a 5:1 LTV:CAC ratio. The kind that makes investors lean forward.
Then I asked one question: "What percentage of your revenue goes to API and compute costs?"
The answer was 52%.
Their gross margin was 48%. Standard SaaS benchmarks assume 70–80% gross margins. At 48% gross margin, their effective LTV dropped from $12,000 to roughly $5,760. The 5:1 ratio became 2.4:1. Below the minimum threshold for sustainable growth.
This isn't an edge case. AI companies operate with a fundamentally different cost structure, and applying standard LTV-CAC frameworks without adjusting for it produces dangerously misleading numbers.
The cost structure problem
Traditional SaaS delivers the same software to each customer. Marginal cost of serving an additional user is near zero. That's why SaaS gross margins sit at 70–80%.
AI products are different. Every customer interaction costs money. An LLM-powered feature runs inference on every request. A computer vision model processes every image. The more a customer uses your product, the more it costs to serve them.
This changes three things simultaneously:
Variable costs are higher
Traditional SaaS variable costs run 10–20% of revenue. AI products face 30–60%, depending on model size, inference volume, and whether you use third-party APIs or self-hosted infrastructure.
Variable costs scale with usage
A customer sending 10,000 API calls per month costs 10x more to serve than one sending 1,000. They may not pay 10x more, especially on flat-rate plans.
Variable costs are unpredictable
Token costs, GPU pricing, and API rates change. OpenAI has cut pricing 10x in two years. Your COGS line item can shift 30% quarter over quarter.
Contribution Margin LTV: the metric AI companies need
Standard SaaS LTV formula:
For AI companies, replace gross margin with contribution margin that accounts for variable compute:
Variable costs per customer include:
- API costs (OpenAI, Anthropic, Google, Cohere, or self-hosted inference)
- Cloud compute (GPU instances for self-hosted models)
- Per-request infrastructure (load balancers, queues, logging)
- Data processing and storage tied to usage
- Third-party per-transaction fees
Worked example
A document analysis AI charges $200/month. Variable costs per customer:
LLM API costs: $60/month
Cloud processing: $15/month
Storage and bandwidth: $8/month
Total variable: $83/month
Contribution margin per customer: $200 - $83 = $117/month
Average customer lifetime: 18 months
Contribution Margin LTV: $117 × 18 = $2,106
Standard LTV (using 75% gross margin): $200 × 0.75 / (1/18) = $2,700. Overstates by 28%.
At $800 CAC: standard formula shows 3.4:1 (comfortable). Contribution margin formula shows 2.6:1 (below threshold). Different number, different strategic decision.
The heavy-user problem
In SaaS, heavy users are your best customers. They churn least and expand most.
In AI, heavy users can be your most expensive customers. A power user running 50,000 LLM calls per month at $0.003/call costs $150/month in API fees. On a $199/month plan, your contribution margin is $49. A light user running 2,000 calls costs $6 and contributes $193.
Three pricing responses
Usage-based pricing
Solves this directly. Charge per call, per document, per minute. Aligns costs with revenue. Downside: customers dislike unpredictable bills.
Tiered pricing with usage caps
"Pro plan: 10,000 analyses/month for $299." Protects margins while capping exposure. Risk: power users churn at the cap.
Hybrid models
Combine a base fee with usage charges above a threshold. Twilio pioneered this. Predictability of subscription + margin protection of usage pricing.
CAC differences for AI startups
Demo and POC costs are higher
AI demos often need custom data ingestion or model tuning before prospects can evaluate. This adds $2,000–$10,000 per enterprise prospect to CAC.
Free tier costs are real
A SaaS free tier costs near-zero per user. An AI free tier with LLM access costs $0.50–$5.00 per monthly active user. At 10,000 free users: $5,000–$50,000/month in compute with no revenue. If 5% convert (500 paid), that's $100/customer in free-tier CAC before any marketing spend.
How to calculate unit economics for your AI startup
Step 1: Variable cost per customer per month
Pull actual API bills, compute costs, per-request infrastructure. Use last month's invoices, not estimates.
Step 2: Contribution margin per customer
Monthly revenue minus variable cost. If negative for any segment, you have a pricing problem no amount of growth fixes.
Step 3: Segment by usage tier
Lightest 20% and heaviest 20% will have radically different economics. Report separately.
Step 4: Contribution Margin LTV
Use actual contribution margin and retention data by usage tier.
Step 5: Fully-loaded CAC including free-tier subsidy
Total sales/marketing costs plus free-tier compute, divided by new paying customers.
Step 6: Ratio and payback
Contribution Margin LTV / fully-loaded CAC. Target 3:1 minimum.
Use the calculators at frameworks: input your actual variable costs rather than defaulting to SaaS assumptions.
What investors look for in AI unit economics
Based on hundreds of pitch reviews and conversations with 30+ VCs:
Gross margin trajectory
A 40% margin trending toward 60% as you optimize inference is acceptable. A flat 40% for 12 months is not.
Model cost reduction plans
Quantization, smaller models for simpler tasks, caching, provider negotiation. "We'll switch to open-source" is not a plan.
Usage distribution data
What % of customers hit each usage tier, and contribution margin at each. A Pareto distribution where 5% of users cause 50% of compute costs is manageable.
Payback under 18 months on contribution margin basis
Tighter than standard SaaS because AI costs are less predictable.
Growth Map for AI
When plotting an AI startup on the Growth Map, use Contribution Margin LTV (not standard LTV) on the vertical axis. This often moves AI companies from the Star quadrant into the Trap quadrant.
The Trap — high customer value but high acquisition cost relative to margin — is common in AI because founders see strong revenue-based LTV but ignore variable cost erosion. The response: focus on margin improvement before scaling acquisition spend.
If your unit economics look worse after reading this, that's useful information. The numbers haven't changed — your understanding has. Building strategy on accurate numbers produces better outcomes than building on flattering ones.
Cite This Data
Source: LTVCACBook.com | Author: Lech Kaniuk
Published: February 20, 2026 | Last updated: February 20, 2026
Suggested citation: Kaniuk, L. (2026). Unit Economics for AI Startups. LTVCACBook.com.
Frequently Asked Questions
Go Deeper
"The Two Numbers That Build or Break Every Business" includes sections on AI company unit economics, Contribution Margin LTV, and adjusted Growth Map positioning.
Learn moreRelated
This article draws on Chapter 7 of The Two Numbers, which covers special cases — AI, marketplaces, and businesses with shifting unit economics in full detail.
Next step
Simulate your unit economics
Stress-test LTV, CAC, churn, and margin assumptions in the unit economics simulator.
Still Building — a newsletter for founders
One email per week. Unsubscribe anytime.
Written by Lech Kaniuk, author of "The Two Numbers That Build or Break Every Business."