The same artifact keeps coming up in customer conversations: an AI spend dashboard, every team's token consumption broken out to the dollar. The numbers are precise. They're also measuring the wrong half of the problem.

Every infrastructure and martech vendor is racing to build a layer that ties token spend to business outcomes. Ramp calls it spend intelligence. Alexander Atzberger calls it Return on Token. The ambition is reasonable, because the pressure is real: Ramp reports AI spend up 13x since the start of 2025, and a single runaway agent loop can burn fifty thousand dollars before anyone notices.

The trap is hiding inside that ambition. Sarah Guo put the mechanism in one line: a thing you can measure is a thing you can train against. The tokens you can cleanly attribute to an outcome are the legible ones, the retrieval and boilerplate and generic reasoning that has a public right answer. That is exactly the work open models already do for a fraction of the price, and it gets cheaper every quarter. So the attribution layer everyone is building does its sharpest work on the half of your spend that's already racing to zero.

The tokens that hold their value are the ones no dashboard can see into. A token reasoning over your company's private data, inside a workflow where a person puts their name on the result, is worth far more than a generic-question token that's worth almost nothing. Its value lives inside a walled system with no public verifier, so there's no clean line from that token to a figure on a slide. The most valuable work is illegible by construction. This is the money version of something I wrote about last week: a dashboard can't see when the work got better, and it turns out it can't see where the value lives either.

Token spend is really a portfolio, and the two tiers call for opposite moves. The bottom tier substitutes for labor, so you minimize it. One team cut its token spend 60% just by stopping engineers from routing every trivial question to the most expensive model. The top tier compounds judgment, so you protect it and invest in it. Matan Grinberg framed it as the first time technology costs roughly what a person costs. Run fleets of agents in parallel and you are already spending around thirty thousand dollars of tokens per engineer, which makes allocation a deliberate call rather than an accounting cleanup.

Which leaves the question none of the vendors can answer yet. We have a decent primitive for metering the cheap half of the token economy and nothing for valuing the expensive half. Until that exists, most companies will govern the AI spend that matters least with great precision, and guess at the spend that matters most.