Is your AI SaaS gross margin VC-ready?
Enter your AI SaaS numbers below to see your true AI Layer gross margin and where it leaks — about 60 seconds, no email.
Nothing is saved or sent — your numbers stay in your browser.
Calculate your true AI Layer gross margin - the metric your investor will actually ask about. Change any input and the result updates instantly; the formulas are the same ones used inside the AI SaaS Financial Model template. 60 seconds · 5 inputs · no email required.
Your AI COGS waterfall
Start from monthly variable revenue. Each step subtracts a real AI cost. What is left - the green column - is your AI Layer margin.
The Variance Buffer is shown as its own step - it is the cost surprise you are stress-testing against, not a cost you pay today. Infrastructure (GPU + vector DB) appears only when you enter a value in Full mode.
How this calculator works
AI Layer Gross Margin = (Variable Revenue − Total AI COGS) ÷ Variable Revenue
It isolates the slice of revenue that pays for AI-specific cost - LLM inference, GPU, vector DB, and Human-in-the-Loop - and ignores classical software margin, which hides the AI cost inside a comfortable blended number.
What this calculator does NOT compute
- A 3-year monthly forecast - this is one snapshot, the bundle gives 36 months
- Customer-segment split (SMB / Mid-Market / Enterprise) - this blends them
- Traditional COGS (hosting, support, payment processing)
- Cash Flow, Balance Sheet, Cap Table, valuation, cohort analysis
- LTV:CAC, Rule of 40, NDR, CAC Payback, Burn Multiple, Magic Number
- AI operational metrics (% Autonomous, AI Resolution Rate, Developer Acceptance Rate)
The formula, expanded
The 8 input variables
| Input | Meaning |
|---|---|
| Average active users | Monthly active users generating LLM calls; blended across tiers. |
| LLM cost per active user | Average per-user LLM API spend per month. |
| Heavy User Multiplier | How much more inference the heavy tail consumes versus the median. |
| % Users that are Heavy | Share of users in that heavy tail. |
| Variable revenue per user | Per-user revenue from Per Activity / Output / Outcome components. |
| AI Cost Variance Buffer | Stress multiplier on base AI COGS for a cost surprise. |
| HITL cost % | Human quality-assurance time as % of revenue - counted as COGS. |
| GPU + Vector DB fixed | Common AI infrastructure not allocated per user. |
Default values and why
| Variable | Default | Basis |
|---|---|---|
| Variance Buffer | 1.20× | Practitioner default for Series A models - D. Perelygin |
| Heavy User Multiplier | 2.0× | Practitioner estimate, aligns with the Microsoft Copilot tier analysis |
| % Heavy users | 10% | Practitioner estimate, typical Pareto tail |
| HITL cost | 7% | Practitioner estimate, Series A median - D. Perelygin |
Sources
- a16z - State of AI / AI app economics
- Bessemer Venture Partners - Cloud Index and the AI Pricing Playbook for founders
- David Skok - SaaS unit-economics benchmarks (forentrepreneurs.com)
- David Sacks, Craft Ventures - Burn Multiple
- Microsoft GitHub Copilot heavy-user economics - press reporting, 2023
- Practitioner observation - Dmitry Perelygin, fractional CFO, where labelled as estimate
Some source URLs are pending final verification before this note is expanded.
Reading your result like an investor
Four takeaways, each tied to a number you just entered. They update with the calculator above.
Classical SaaS assumes 80-90% gross margin, because serving one more customer costs almost nothing. AI SaaS breaks that: every LLM call has a real per-token cost that scales with usage. When a VC reads your gross margin line, the question is no longer "how much do you spend on hosting" - it is "what does that blended number look like once the AI Layer is separated out?" That AI Layer slice - inference, vector DB, embeddings, HITL - usually lands at 25-45% in early stage, not 80%.
In the bundled Helix AI demo, Blended GM looks reassuring at 79% and Traditional GM at 96% - but the AI Layer GM sits at 27% in Year 1 and only recovers to 41% by Year 3, through caching, model routing and price optimisation.
Source: Bessemer AI Pricing Playbook →The most overlooked line in AI SaaS unit economics is the heavy-user tail. In a usage-priced product the top 5-15% of users consume 4-6× more inference than the median - their cost can be 8-12× the average user while their revenue is capped by the subscription tier.
The canonical public case is Microsoft GitHub Copilot: reporting in 2023 indicated Microsoft was losing about $20/user/month on power developers. The flat fee could not absorb their inference appetite - which is why Cursor, Replit and Claude Code all restructured pricing toward usage-based caps. If you model AI Layer GM without splitting Light and Heavy cohorts, you systematically overstate margin.
Source: Wall Street Journal, Microsoft Copilot economics →Frontier LLM prices trend down over time, but day to day they can spike - a model deprecates, a provider raises rates, an outage forces fallback to a costlier provider. The Variance Buffer is a multiplier on base AI COGS: 1.20× means "I expect my modelled cost, but I am stress-testing margin against a 20% surprise."
A senior investor will not accept a model that assumes flat AI cost - they will ask what happens at +20% and +50%. Practitioner estimate - Dmitry Perelygin, fractional CFO: 1.20× realistic for Series A models, 1.25× conservative, above 1.50× as a separate stress test.
A frequent error in early models is treating Human-in-the-Loop time as OpEx - founder and engineering hours absorbed into payroll. That leaves a gross margin line that looks healthy because the messy human cost is hidden one level below. It is wrong by accounting logic and worse by investor optics: time spent on AI quality assurance and exception handling is direct cost of delivering the output, so it belongs in COGS.
Practitioner estimate - Dmitry Perelygin, fractional CFO: HITL lands at 10-20% of revenue at Seed, 5-10% at Series A, under 5% at Series B+. The trajectory matters - investors want the curve falling as % Autonomous rises.
What this snapshot still cannot see
These also decide your round - and none of them can be answered by a single-month calculation. They need a full 3-year model.
- Net Dollar Retention - >110% median, 130%+ best
- Rule of 40 - >40% healthy, >60% top-tier
- LTV:CAC - >3× healthy, 5×+ excellent
- CAC Payback - <12 months healthy, <6 top-tier
- Burn Multiple - <2× acceptable, <1× top-tier
- % Autonomous & AI Resolution Rate - the AI operating curve
One snapshot is not a financial model.
To walk into a VC meeting you need the whole economy modelled - 36 months, three customer segments, AI Layer and Traditional GM decomposed, cash flow, balance sheet, cap table to exit, and 87 cited benchmarks. That is the AI SaaS Financial Model bundle - built by the same fractional CFO behind this calculator.
See the full AI SaaS model →
Common questions
What is AI Layer Gross Margin and why does it differ from Blended GM?
AI Layer Gross Margin isolates the slice of revenue that pays for AI-specific costs - LLM calls, vector DB, embeddings, GPU rent, HITL - and divides it by the variable revenue that this slice generates. Blended GM averages this with classical software margin and hides the AI-specific cost inside a comfortable number.
What is a healthy AI Layer GM for a Series A AI SaaS?
Based on the Bessemer Cloud benchmarks and a16z State of AI: Seed/Series A typically 25-45%, mature AI products 45-50%, top-tier 50%+. Below 25% is not fatal early on, but you need a credible path upward.
What is the Heavy User Multiplier?
It captures that the top 5-15% of users in a usage-priced AI product consume 4-6× more inference than the median. The Microsoft Copilot loss case is the canonical example. Ignoring it overstates margin.
What is the Variance Buffer and why is 1.20× the default?
A multiplier on base AI COGS that stress-tests margin against a cost surprise. 1.20× is the practitioner default for Series A models; 1.25× conservative; above 1.50× as a separate stress test.
Why is HITL counted as COGS, not OpEx?
Human time spent on AI quality assurance and exception handling is direct cost of delivering the AI output. It belongs in COGS, not OpEx. Treating it as OpEx inflates visible gross margin.
How accurate is this calculator versus the full model?
The calculator gives one blended snapshot. The full AI SaaS Financial Model bundle computes the same number across 36 months and three customer segments, plus cash flow, balance sheet, cap table and 87 cited benchmarks.
Can I use this calculator for non-AI SaaS?
It is built for AI SaaS where inference cost scales with usage. For classical SaaS with near-zero marginal cost, the AI-specific inputs add little.
Why does the calculator not collect my email?
There is no email gate. Nothing is stored or sent. Save builds a bookmarkable link that holds your inputs in the URL - your numbers stay in your browser.