AI Startup Unit Economics: Gross Margin and ARR

Written By Mark Sue

This guide is for founders of AI-native companies, where every customer request has a real compute cost. SaaS math no longer fits. The guide shows how to measure cost per task, report gross margin with compute included, and present ARR that holds up when investors re-cut it.

What's in the guide

  • AI-native vs. SaaS economics: margins, biggest COGS, pricing models, commitments and moat
  • AI gross margin data: SaaS median around 77% vs. AI products around 45% in 2025, and rising
  • Unit economics per task: tokens × blended price + overhead = cost per task
  • Pricing in the wild: Intercom Fin, Salesforce Agentforce, Cursor and Decagon
  • Seven levers to cut cost per task, from model routing and caching to batching and self-hosting
  • Compute commitments: sizing them to the downside and adding unused commitments to burn
  • What counts as ARR, what to show separately and what to exclude
  • The investor diligence test with pass marks, an AI-native KPI page and accounting watch-outs
  • Five AI-native risks to rate each quarter, plus a finance stack and glossary for board decks

Questions this guide answers

What gross margin should an AI startup have?

The guide cites a classic SaaS median of about 77% and an average of about 45% for AI products in 2025, rising as compute gets cheaper. Its diligence pass mark is 60%+ and rising, with a path to 70%. Always show gross margin with compute included.

How do you calculate cost per task for an AI product?

Multiply tokens per task by your blended price per million tokens, then add overhead per task. In the guide's example, 12K tokens at $8 per million plus $0.03 overhead costs $0.126 against a $0.50 price.

What counts as ARR for an AI company?

Contracted recurring platform fees and committed usage minimums count. Show variable usage separately as labeled run-rate, and exclude pilots, paid proofs of concept and one-time setup. Annualizing one hot month is not ARR.

How much compute should an AI startup commit to?

Commit no more than about 70% of base-case usage and size it to the downside case. Unused commitments are cash out the door, so add them to burn and runway.

What do investors check in AI startup diligence?

The guide's test asks for cost per task by quarter with gross margin of 60%+ and rising, net dollar retention above 110%, under 70% of spend on one model provider, training rights in contracts with all top-20 accounts, and a margin model built bottom-up at 10× scale.

How Strativ helps

Strativ sets up finance stacks that tie usage billing and compute cost to the same customer, so margin is visible per account. Mark helped create and taught the Finance & Accounting Technology (FATE) Certificate, and his Wall Street background at RBC Capital Markets helps founders present ARR and margins the way investors will check them.

Request the full guide to get the complete one-page reference, or book a call to talk through your numbers with Mark.

Mark Sue