Why your AI product has SaaS pricing and utility-grade costs

Why the widely cited “token prices have collapsed” statistic can be true at the same time AI companies report worse gross margins than the software they’re replacing

The specific mechanism, agentic multi-step workloads, currently outpacing the price-deflation curve, with the 2026 numbers behind it

Why one AI lab’s margin can climb from 38% to over 70% in a year while its own customers’ margins compress, using the same usage growth as the cause

A framework for telling whether a given AI cost behaves like software (falling toward zero at scale) or like a utility (scaling with usage)

How to reread a vendor’s pricing page, a startup’s unit economics slide, or an earnings call transcript once you’re looking for this distinction

What Uber’s four-month AI budget and Microsoft’s internal Claude Code cancellation actually tell you about where this is heading

Token prices have fallen by roughly two to three orders of magnitude since 2022, yet gross margins on AI-native products are falling too, from a SaaS-standard 70 to 90 percent range down toward the low 50s in 2026.

The reason: agentic workflows consume far more tokens per completed task than the falling price curve accounts for, so total spend rises even as the unit price drops.

The margin math splits by layer. Model owners are capturing rising margin from usage growth. The companies building on top of those models are absorbing the rising cost as COGS.

This is already visible in the market: Uber, Microsoft, Cursor, GitHub, and Salesforce have all had to publicly reprice or restrict AI usage in 2026 because flat pricing collided with real, scaling compute costs.

The strategic implication: AI-native software should be underwritten like a utility with a real marginal cost, not a software business with a marginal cost of zero.