AI Unit Economics: The Numbers Your Investors Will Ask About (That Most Founders Can't Answer)
The question came at 4pm on a Tuesday, midway through our Series A due diligence. Our investor was looking at a line in the P&L: "AI Infrastructure: $31,000/month." He looked up and asked: "What does that look like per customer?"
Dead silence. We had total spend. We had monthly trends. We could say it was growing. We could not say who it was growing because of. We were a data company that had been running blind on our biggest variable cost for six months.
Why total AI spend is the wrong metric
Total AI spend tells you if you have a problem. It does not tell you what to do about it. A $31k AI bill looks completely different depending on how it's distributed. Is it 600 customers at $51 each? 50 customers at $620 each? Eight customers at $2,000 each and everyone else at $10? Each of those is a different business problem with a different solution — and the blended average hides all three.
The metric your investors are actually asking about is contribution margin per customer: revenue from that customer minus the direct cost to serve them, including AI. If you can't answer that, you can't answer whether your pricing is sustainable. You definitely can't answer whether your enterprise tier is more profitable than your Pro tier — or whether it's secretly the most expensive segment you serve.
The metrics that actually matter
AI cost per customer is the baseline — what does it cost, in API spend, to serve one customer per month? Not the average, the distribution. One customer at $3,000 and ninety-nine at $10 gives you an average of $39. That average is useless for every decision you need to make.
AI as a percentage of revenue by pricing tier is the next layer. Your free tier is 100% cost and 0% revenue — that's the floor. Your paid tiers should have AI costs well below the revenue they generate. If your Pro customers are generating $99/month each and costing $60 in AI to serve, you don't have a growth story. You have a pricing problem.
AI gross margin by feature is the most granular — and usually the most surprising. Our document summarisation feature was our most prominent marketing hook. It was also our least profitable. High token usage, underpriced tier, heavy use by free users. We were optimising acquisition around a feature that was actively hurting margins.
What we found when we finally had the numbers
Our enterprise segment — the accounts paying us the most — had a 12% AI gross margin. Our Pro segment had 71%. Enterprise customers used more of every AI feature and generated far more tokens per session. We had priced for the revenue uplift of moving to Enterprise. We had not priced for the cost increase that came with it. The two numbers had never sat next to each other before.
Once we had cost by customer, the board conversation changed. We weren't defending a large spend line. We were presenting a clear picture: here are the profitable segments, here are the ones we're fixing, here's the plan. That's a completely different conversation than "AI is expensive but necessary."
If you're heading into a board meeting without these numbers, PerUnit gives you cost attribution by customer, feature, and pricing tier — with your Stripe revenue alongside so you can see margin, not just spend. Our free AI margin calculator is a useful place to start with your current numbers.