Skip to main content
Back to Blog
insights8 min read

Usage Variance in AI Products: Understanding Per-Customer Cost Distribution

In AI products, the top 10% of customers often generate 80% of API costs. This article analyzes per-customer cost variance, margin calculation by cohort, and pricing model options.

BA

Blaise Albuquerque

Founder, Bear Billing

#unit-economics#pricing-strategy#cohort-analysis#ai-margins#power-users

Summary: In flat-rate AI products, the top 10% of customers by usage typically generate 80% of API costs while paying the same rate as light users. This cost distribution pattern is visible in GitHub Copilot's margin structure, Cursor's four repricing cycles in 2024, and ChatGPT Pro's unit economics at $200/month.


The Cost Distribution Pattern

A typical AI product launch:

  • $29/month flat subscription
  • Unlimited usage (or untracked "fair use" limits)
  • 100 customers in Month 1 = $2,900 revenue

Then the OpenAI bill arrives: $4,500.

The difference: Not all customers cost the same.


Three Customer Cohorts by Usage

Every AI product shows three distinct usage cohorts:

Light Users (60% of customers)

  • Use product occasionally (few times per week)
  • Generate $5-10/month in API costs
  • Pay $29/month = +$20 customer margin
  • This cohort cross-subsidizes other segments

Average Users (30% of customers)

  • Daily usage, moderate sessions
  • Generate $25-35/month in API costs
  • Pay $29/month = Break-even or slight negative margin
  • Neither margin-positive nor significantly margin-negative

Power Users (10% of customers)

  • All-day usage, complex queries, long conversations
  • Generate $80-150/month in API costs
  • Pay $29/month = -$50 to -$120 margin per user
  • This cohort consumes margin from other segments

Case Studies: Per-Customer Margin Analysis

GitHub Copilot: -$20 to -$80 Per User

GitHub Copilot charges $10/month with an average margin of -$20/user.

The cohort distribution:

User Cohort% of UsersAPI CostSubscriptionMargin
Light Users20%$5/mo$10/mo+$5
Average Users60%$15-20/mo$10/mo-$5 to -$10
Power Users20%$50-80/mo$10/mo-$40 to -$70

The 20% of light users with positive margin cannot offset the 20% of power users with negative margin.

Microsoft's position: Price increases would affect customer retention. Current approach subsidizes the product with Azure infrastructure margin.


Cursor: Four Repricing Cycles in 12 Months

Cursor reached $500M ARR with negative per-customer margins. Their pricing evolution:

  1. January 2024: $20/month unlimited — High-usage customers identified, margins compressed
  2. May 2024: $20/month with soft limits (500 completions) — Customer feedback on perceived value change
  3. September 2024: $20/month with hard limits + overage charges — Customer adjustment period
  4. December 2024: $200/month "Ultra" tier for high-usage customers — June 2025: $71 single-day charges reported

Each repricing cycle involved customer communication, retention considerations, and support volume.

The pattern: Flat pricing with variable AI costs requires adjustment as usage distribution becomes visible.


ChatGPT Pro: $200/Month Margin Analysis

Sam Altman stated publicly that ChatGPT Pro at $200/month shows negative margin on power users.

At 20x the base tier price, power users consume 50-100x more resources than average users.


The Margin Formula

The blended margin calculation for AI products:

Total Margin = (Light User Margin × 60%) + (Average User Margin × 30%) + (Power User Margin × 10%)

Example with $29/month pricing:
= (+$20 × 0.6) + (-$5 × 0.3) + (-$100 × 0.1)
= +$12 - $1.50 - $10
= +$0.50 per customer overall

Gross Margin: 1.7% (vs. SaaS benchmark of 70-80%)

Observations:

  • One power user joining offsets approximately 5 light users' contribution
  • Power users are typically the most engaged (highest product satisfaction)
  • Marketing often attracts high-engagement users (viral growth correlates with usage)
  • As scale increases, power user percentage may increase, compressing margins

Indicators by Margin Status

Margin Under 30%

  • Most engaged users show highest churn (cost-to-serve exceeds willingness to pay)
  • API costs growing faster than revenue (30%+ monthly)
  • Gross margin under 40% (SaaS benchmark is 70-80%)
  • Top 10% of users generate 60%+ of API costs

Margin 30-50%

  • Monthly infrastructure costs have high variance
  • "Unlimited" usage creates cost unpredictability
  • Price increases correlate with churn increases

Margin Above 50%

  • Gross margin above 60%
  • Usage distribution is relatively flat
  • Cost prediction accuracy within 10%
  • Pricing model includes usage-based component

Pricing Model Options

Option 1: Usage Limits

Characteristics:

  • Immediate cost alignment
  • Simple implementation
  • Direct cost-to-price relationship

Considerations:

  • High-usage customers may provide feedback
  • May affect highest-engagement segment
  • Requires clear communication

Examples:

  • Cursor: 500 requests/month limit
  • Anthropic Claude: Message caps per tier
  • Replit: Compute time limits

Option 2: Tiered Pricing

Characteristics:

  • Usage-based pricing at tier level
  • Light users remain on lower tier
  • Predictable revenue per tier

Considerations:

  • Requires clear tier communication
  • Needs usage tracking infrastructure
  • Tier selection behavior varies by customer

Examples:

  • GitHub Copilot: $10 Individual, $19 Business
  • Cursor: $20 Pro, $200 Ultra
  • ChatGPT: $20 Plus, $200 Pro

Option 3: Hybrid Model

Characteristics:

  • Base subscription + usage overages
  • Predictable base revenue with margin protection
  • Familiar "utility bill" model

Considerations:

  • More complex implementation
  • Requires real-time usage tracking
  • Communication important to set expectations

Examples:

  • Vercel: Base plan + bandwidth/executions
  • Twilio: Monthly minimum + per-usage fees
  • AWS: Reserved instances + on-demand

Option 4: Outcome-Based Pricing

Characteristics:

  • Charge for value delivered (tickets resolved, code shipped)
  • Aligns incentives (provider optimizes costs, customer pays for results)
  • Insulates from token cost volatility

Considerations:

  • Requires clear outcome definition
  • Needs sophisticated tracking
  • Customer education required

Examples:

  • Twilio AI: Per-ticket-resolved pricing
  • Some code generation tools: Per-PR pricing
  • Emerging model with limited adoption

Implementation Timeline

Week 1: Measure Current State

  1. Export usage data from OpenAI/Anthropic/your provider
  2. Segment customers into light/average/power cohorts by token usage or cost
  3. Identify highest-cost customers (top 5%)

Week 2: Model Scenarios

  1. Estimate retention impact per scenario (10-30% adjustment for power user segment)
  2. Calculate break-even point for each pricing model

Week 3: Select Approach

If gross margin < 30%:

  • Implement usage limits within 30 days
  • Add premium tier for high-usage customers
  • Consider grandfather pricing for existing customers

If gross margin 30-50%:

  • Plan tiered pricing rollout over 90 days
  • Communicate early and consistently
  • Offer migration incentives

If gross margin > 50%:

  • Add usage-based overage charges
  • Monitor cohort distribution
  • Build infrastructure for real-time tracking

Week 4: Execute and Communicate

  1. Pre-announce changes (2-4 weeks notice)
  2. Explain the rationale (transparency supports understanding)
  3. Offer credits/grandfathering to ease transition
  4. Monitor support volume and sentiment
  5. Adjust based on feedback

The Market Pattern

Flat-rate pricing for variable-cost AI products creates margin pressure.

AI companies typically follow one of these paths:

  1. Repriced (Cursor, Notion AI, Jasper)
  2. Currently repricing
  3. Subsidized by other revenue (Microsoft with Copilot)
  4. Operating with compressed margins

Companies with sustainable margins typically:

  • Track costs per customer in real-time
  • Model pricing scenarios with current data
  • Communicate changes transparently
  • Build usage-based infrastructure early

Bear Billing: Cost Visibility for AI Products

We built Bear Billing to provide per-customer cost visibility:

  1. Track costs per customer automatically (OpenAI, Anthropic, AWS, etc.)
  2. Identify usage patterns by cohort
  3. Model pricing scenarios with real usage data
  4. Implement usage-based billing infrastructure

Join the waitlist for early access.



Key Takeaways

  1. 60% of customers show positive margin, 10% show significantly negative margin — Visibility into cohort-level margins is essential
  2. Power users cost 10-100x more than light users — Flat pricing cannot accommodate this variance
  3. Even $200/month may show negative margin — ChatGPT Pro demonstrates the limit of tier-based pricing
  4. Pricing adjustments work better with lead time — Proactive changes with customer communication outperform reactive changes
  5. Transparent communication supports retention — Customers understand cost structures when explained clearly

Know your per-customer margins. Visibility into cohort-level economics enables informed pricing decisions.

Join our waitlist for per-customer cost tracking and margin analytics.

Share this article