How concentrated is your AI cost tail?
Before you pick a pricing model, see the shape of the risk: how few users drive your inference bill, how many lose money under a flat price, and how often a bad month wipes your margin. ~60 seconds, no email.
Who actually drives your inference bill
Cost-concentration curve
—
Monthly margin under flat — the tail
Simulated monthly AI-layer margin across many months (Monte-Carlo). The band spans the typical 5th–95th-percentile month; anything left of the 0% line is a month you lose money on every user.
How this simulator works
The idea — Inference Variance Allocation, one level deeper. The other calculators ask "what's my margin" and "which pricing model protects it." This one measures the thing those answers depend on: how heavy-tailed and volatile your inference cost actually is. A flat price is an un-reserved insurance policy written against your own power users — so the first question is how big the tail is.
Cost share of the top (1−F) users = 1 − Φ( Φ⁻¹(F) − σ )
Concentration (Gini) = 2·Φ( σ / √2 ) − 1
Monthly margin = 1 − (sampled cost per user) × cost-shock − other COGS → Monte-Carlo
What it shows
- Cost concentration — what share of your bill the top 1% / 5% / 10% create.
- Unprofitable users — the % whose inference cost already exceeds the flat price, and the share of the bill they burn.
- Tail risk — the spread of monthly margin, the chance of a loss-making month, and how far the worst months fall.
What it does not compute
- Your exact tiers, caps and markups → the template
- Which model to pick → the Pricing Model Calculator
- A 36-month runway / cap-table / valuation model
- Per-customer-segment splits — this blends them
Heavy-tail & reserve framing: Gomes, Your SaaS Is an Insurance Product (arXiv:2605.16699); Bai et al., How Do AI Agents Spend Your Money? (arXiv:2604.22750, agentic ≈1000× tokens, run-to-run variance up to ~30×). Concentration analogue: broadband heavy tail (top 1% ≈ 24% of traffic). Behavioural self-selection into flat: Lambrecht&Skiera (2006). Two-part-tariff fix: Oi (1971), Sundararajan (2004). Benchmarks: a16z, Bessemer, ICONIQ. Lognormal is a standard, well-behaved model of right-skewed usage; it gives a directional diagnosis from your inputs — pull your exact distribution from your COGS sheet for the precise figures.
Reading your tail like a CFO
The most overlooked line in AI unit economics is the heavy tail. In a usage-priced product the top 5–15% of users consume 4–6× more inference than the median; agentic workloads can consume on the order of 1,000× the tokens of a chat turn. A handful of users quietly drives the majority of the bill — and under a flat price they pay no premium for it.
When inference cost exceeds the flat price, a user is a loss the moment they engage — high engagement becomes worse than churn. The canonical case is GitHub Copilot: Microsoft was reported to be losing about $20/user/month on power developers, which is why it moved all plans to usage-based billing in 2026.
Mean margin hides the months that sink you. A heavy tail plus a token-price shock means your worst months can land far below the average — and the more concentrated your usage, the wider that swing. A senior investor will not accept a model that assumes flat AI cost; they ask what the bad month looks like.
This is the comfortable average — the number that looks fine right up until the tail shows up. Early-stage AI Layer GM typically runs 25–45%, not the 80–90% of classic SaaS. If your mean already sits low and your tail is heavy, the flat price is living on borrowed time.
Source: Bessemer AI Pricing Playbook → · Check your margin → Gross Margin Calculator
You've seen the tail. Now defend against it.
To act on it you need the model that turns your exposure into a contract — exact tiers, caps and a reserve buffer — plus the 36-month economy behind your round: three customer segments, AI Layer and Traditional GM decomposed, cash flow, balance sheet, cap table to exit, and 87 cited benchmarks. Built by the same fractional CFO behind this simulator.
Learn how to build the full AI startup financial model →
Turn your exposure into a contract before investors probe it:
Get the AI SaaS Financial Model — €147 → Read the book: How to Design an AI SaaS That Survives →About the author
Fractional CFO, Piedmont, Italy. I have sat on both sides of the table — raised $12M as a CFO and evaluated investment deals from the buy side. The Inference Variance Allocation lens behind this simulator is the same financial work I do with AI SaaS founders preparing for their first round.