📈 Why AI bills rise as costs fall

There's a paradox sitting at the heart of enterprise AI spending right now, and it's the kind of thing that makes finance teams break out in a cold sweat. Token prices have been dropping for months, sometimes dramatically, as model providers compete on efficiency and inference costs shrink. And yet the invoices keep going up. Not down. Up.

The explanation isn't mysterious once you sit with it. When something gets cheaper, people don't use less of it, they use a lot more. Call it tokenmaxxing: engineers and product teams start piping bigger context windows into every call, chaining multiple model requests together for a single task, running agents that loop and retry and self-correct. Each of those individual actions costs less than it used to. But the volume explodes, and the volume is what actually shows up on the bill at the end of the month.

This is exactly the pattern Exponential View flagged in earlier coverage of what CFOs are up against, and it maps almost perfectly onto Jevons' paradox from 19th-century economics, where more efficient coal use led to more coal consumption, not less. AI cost forecasting is hard right now precisely because nobody has a stable baseline. A team might run the same workflow twice in a week and get wildly different costs, not because pricing changed, but because someone added a retrieval step, swapped to a bigger model for accuracy, or let an agent take twelve extra turns to finish a task nobody double-checked.

What happens as the industry gets better at this? Probably a shift from reactive expense-tracking to genuine cost engineering, the same evolution cloud computing went through a decade ago, when FinOps teams turned unpredictable AWS bills into something closer to a managed budget line. Expect more granular metering, per-agent cost caps, and tooling that treats token spend the way SREs treat latency: a thing you monitor, alert on, and design against from the start, not something you discover in an invoice.