Unit economics for AI products and the real cost of inference

AI products have a cost that classic software mostly lacks: every request uses compute, so serving a user is never free. Unit economics for AI means measuring what each customer pays against what they cost to serve, then pricing so that the gap stays healthy even for your heaviest users.

This guide shows how to calculate it, what benchmarks suggest, and which pricing moves protect your margin.

What unit economics means

Unit economics looks at the profit from one customer, or one unit of use, instead of the company as a whole. The main pieces are:

• Revenue per customer per month.
• Cost to serve, which for AI includes inference, storage and support.
• Gross margin, the share of revenue left after the cost to serve.
• Customer acquisition cost (CAC), what you spend to win a customer.
• Lifetime value, the total gross profit a customer brings before leaving.

If lifetime gross profit is not comfortably higher than CAC, growth burns cash.

Why AI changes the math

Traditional software has high build costs and near-zero marginal cost. Serving the next user costs almost nothing, so margins of 70 to 85 percent are normal.

AI breaks that pattern. Each generation, chat turn or minute of video consumes model compute. Heavy users cost more, and sometimes the most engaged customers are the least profitable under flat pricing.

Benchmarks reflect this, though they vary by source. Bessemer's State of AI 2025 described one group of explosive-growth AI companies running gross margins around 25 percent, and a steadier group near 60 percent. CloudZero's discussion of AI gross margin notes that many AI products sit well below traditional software, with figures like 50 to 60 percent often cited. Reports about large model providers suggest even they have faced higher inference costs than they planned. These numbers come from different samples, so use them as a rough guide: expect lower margins than classic SaaS, and plan for it.

How to calculate inference cost per user

Start with a simple model for one typical action.

1. List the model calls inside the action. A single "generate" button might trigger several calls.
2. Estimate tokens or compute units per call, using real logs when you have them.
3. Multiply by the provider's price. Check current rates, since they change.
4. Add overhead such as storage, bandwidth, moderation and retries.
5. Multiply by actions per user per month for light, typical and heavy users.

Example with made-up numbers: if one video clip costs $0.40 to generate and the average paying user makes 30 clips a month, the monthly cost is $12. On a $20 plan, gross margin is 40 percent before other costs. A heavy user making 150 clips costs $60 and loses you $40.

Always compute the three user types. The average hides the problem.

The heavy-user problem

Under flat pricing, a small group of power users can consume most of the compute. Commentary on AI economics often notes that engaged customers can be the least profitable ones, since light users subsidize them.

You can manage this in several ways:

• Usage caps or credits. Include a set allowance and sell more.
• Fair-use policies. Soft limits with throttling for extreme use.
• Tiered models. Use a cheaper model for simple jobs and a premium model only when needed.
• Higher tiers for heavy needs. Let power users pay more.

This is part of why pricing in the market is shifting. The Oliver Wyman review of AI pricing discusses the move from seats toward usage, while warning it does not suit every vendor. See also our look at pricing page patterns.

Levers that cut inference cost

Cost per request is not fixed. Common levers include:

• Model routing. Send easy tasks to small, cheap models.
• Caching. Reuse results for repeated or similar prompts.
• Shorter prompts and outputs. Trim context to what the task needs.
• Batching. Group requests where latency allows.
• Open-weight models. Self-hosting can lower cost at scale, but adds engineering and hardware risk.
• Limits on retries. Failed or unwanted outputs that users regenerate multiply cost.

Sources describing inference economics suggest that even a 20 to 30 percent cost reduction can unlock meaningful capacity. That claim comes from commentary and not audited data, but the direction is sensible.

Pricing to protect margin

Once you know cost per action, pick a price rule.

1. Set a target gross margin, such as 60 percent, and work backward. A $12 cost to serve needs a price near $30.
2. Price the plan for the heavy tail, using caps so that the worst case is acceptable.
3. Charge for expensive features separately, such as video, high resolution or long context.
4. Use annual plans carefully. Upfront cash is useful, but a year of heavy usage on a discounted plan can lose money. See monthly versus yearly plans.
5. Revisit prices quarterly. Model prices change, usually downward, so margins may improve if you do not pass all savings on.

Free tiers are a marketing cost

In AI, every free user has a real cost. Treat the free tier as part of acquisition spend. If 100 free users each cost $0.50 a month and 3 become paying customers, that is $50 spent to win 3 customers, or about $16.67 each, before any ad spending. Compare that to what a customer brings in profit.

The conversion side is covered in freemium conversion rates.

Reading other companies' economics from outside

You cannot see a competitor's margin, but public signals help.

• Pricing versus costs. If a product offers unlimited video generation for $10, it either has very low costs or is losing money on heavy users.
• Plan limits. Tight credit allowances suggest tight margins.
• Model choice. A product that advertises a premium model on a cheap plan may be subsidizing usage.
• Funding and growth. Rapid growth funded by investors can hide weak margins.

Use traffic and tech stack data from a site report to see scale and which services a product appears to use, and treat the revenue figure there as an estimate that says nothing about margin.

A short checklist

Before you launch or invest, confirm that:

• Cost per action is measured, not guessed.
• Light, typical and heavy users all leave positive margin.
• The free tier has a hard cap.
• You know the price you would charge if your model costs doubled.
• Lifetime gross profit exceeds acquisition cost.

Contribution margin and payback

Gross margin is a start, but founders also need to know how fast a customer repays what it cost to win them. Divide your acquisition cost by the monthly gross profit per customer to get the payback period in months.

For example, if a customer costs $60 to acquire and pays $25 a month with a 50 percent gross margin, monthly gross profit is $12.50 and payback takes about five months. If margin drops to 30 percent, monthly gross profit is $7.50 and payback stretches to eight months.

Thin AI margins therefore lengthen payback and raise the risk of losing customers before you earn your money back. This is why retention and cost control matter together, and why paid growth works best once margins are stable.

FAQ

What is a good gross margin for an AI product?

Traditional software often runs at 70 to 85 percent, while AI products frequently report lower figures, commonly around 50 to 60 percent and sometimes much less. A target of 60 percent or more is a reasonable aim, and what matters most is that it improves over time.

How do I estimate inference cost per user?

List the model calls in a typical action, multiply by the provider's price, add overhead, then multiply by usage for light, typical and heavy users. Use real logs as soon as you have them.

Will inference costs keep falling?

Per-unit prices have generally trended down, but usage per customer tends to rise as products add features, so total cost does not always fall. Plan margins based on current prices and treat declines as a bonus.

Should AI startups use usage-based pricing?

It protects margin when costs vary by use, but it can feel unpredictable to buyers. Many products blend a subscription with included usage and paid top-ups.