Cost Economics · 2 min read
Why 2× customers doesn't mean 2× AI cost
The most dangerous assumption in an AI-native financial model is that cost scales linearly with users. It rarely does — and the gap is where runway quietly disappears.
By Akhil Anand · September 7, 2026
Open almost any AI-native company's model and you'll find the same load-bearing assumption: cost per customer is a constant. Add 10,000 users, multiply by average cost, done. It's clean, it's intuitive, and it's usually wrong.
Cost per customer is a distribution, not a constant
Customers are not identical units of consumption. One runs a single prompt a week; another wires your product into an agent that loops for hours. When you average them into a single "cost per user," you've collapsed a wide, skewed distribution into a point — and the point is dominated by whoever you happened to have this month.
Now scale. The next 10,000 customers won't have the same mix as your first 1,000. Heavier-usage cohorts, deeper workflows, longer contexts, more tool calls — any shift in composition moves the average. 2× the customers can easily be 3× the cost, and nothing "broke" to cause it.
Where the non-linearity hides
A few compounding effects turn linear assumptions into optimistic ones:
- Context growth. As features mature, prompts get richer. The same request costs more tokens than it did last quarter.
- Model drift upward. Teams quietly route more traffic to premium models as quality bars rise.
- Workflow depth. Agents and multi-step flows multiply calls per user action.
- Cohort composition. Your power users arrive later and cost more.
None of these show up in a users × avg_cost model. All of them show up on the invoice.
Nothing malfunctioned. Every request succeeded. The economics simply didn't scale linearly — and you found out from a bill instead of a forecast.
The fix isn't a better average
You can't fix a distribution problem with a better point estimate. The fix is to model consumption as a distribution and let customer count drive it — so "what happens at 50,000 users" becomes a question you can answer before the invoice, not after.
That's the entire reason AtlasBurn treats AI cost as a probabilistic risk problem instead of an accounting one. Historical accounting tells you what you spent. It can't tell you what your own growth is about to cost you.