The Token Bill Is the New Payroll

For most of human history, intelligence was the scarce input. You hired it, or you spent years building it in yourself, and either way it arrived slowly and priced itself accordingly. That constraint shaped every org chart ever drawn. It is gone. Intelligence is now something you summon, one token at a time, at a price that keeps falling.

Everyone reads that as good news. It is — but not in the way most operators assume.

What Cheap Actually Does to Demand

When something scarce becomes cheap, consumption does not hold steady and hand the savings back. It explodes. Cheap electricity did not leave households with the same lighting bill and more money in the tin; it lit entire cities. Cheap computing did not stop at the mainframe budget; it put a computer in every pocket. The unit price collapsed and total spend climbed, because the falling price made a hundred new uses viable that had never penciled out before.

Intelligence is on that curve now. Every time a model gets cheaper, some workflow that was not worth automating last quarter becomes worth automating this quarter. Your per-token cost drops and your bill grows, and both of those things are true at once. Anyone reading only the first half of that sentence is going to be surprised by the invoice.

The Number Nobody Is Watching

Ramp sits on the card spend of a wide slice of the startup economy, so it has an unusually honest view of the curve. Its customers' monthly token spend rose roughly twentyfold between June 2025 and June 2026, and now runs at about $220 million a month across that base. Ramp built a token spend-management product for one reason: finance teams could not see where any of it was going.

The spend is not the problem. The blindness is. Most companies now carry a five-figure or six-figure monthly line item that nobody owns, nobody attributes, and nobody can break down by team, by model, or by workflow. It is the only meaningful cost centre in the business with no meter on it.

Waste Has a Shape

Unattributed spend is not evenly wasteful. It fails in specific, boring, findable ways.

A production workflow gets pointed at a frontier model for work a cheaper model handles identically, and nobody re-checks it after launch. A high-volume prompt runs with no caching, paying full price to re-read the same context thousands of times a day. A config change ships on a Friday and quietly multiplies a cost for the whole weekend. Ramp reports finding exactly these — one customer whose model setting had multiplied a workflow's cost 6x over a weekend, another that found caching missing from a high-volume path and cut roughly $10,000 a month of avoidable spend the same day.

None of those are strategy failures. They are meter-reading failures. You cannot fix a number you have never seen, and you cannot argue about a number nobody publishes internally.

We Put a Meter on Ourselves

This is not a spectator argument here. CYSTEMS runs an agent estate that thinks all day — writing, auditing, reviewing, publishing, checking its own work — and we built our own token ticker before anyone tried to sell us one, because the first serious month of agent work made the bill impossible to reason about by feel. It is deliberately plain: every machine computes its own daily total and pushes it to a local hub, and the running number only ever moves up.

Having it changed the question. Not "is this too much" — that question has no answer — but "what did this buy." A review round that catches a defect before it ships is worth its tokens several times over. A retry loop regenerating the same refused draft three times is not. Those two look identical on an invoice and completely different on a per-workflow meter. The meter is the whole difference between a cost you manage and a cost that manages you.

Intelligence Per Dollar Is the Metric

The instinct when a bill grows is to cap it, and that is the move most finance teams will reach for first. It is the wrong one. Spending less on intelligence in an era when intelligence is the input is a decision to move slower than the people who learn to spend well.

The metric that matters is intelligence per dollar — how much real output, in decisions made and work shipped and errors caught, you get per unit of spend. That number improves through routing cheap work to cheap models and reserving the expensive ones for where they earn it, through caching, through retiring loops that produce nothing, and through knowing which workflows are actually load-bearing.

That is a Cystems problem, not a procurement problem. Attribution first, then routing, then the boring efficiency work, in that order — because you cannot route what you cannot see.

The companies that come out ahead over the next few years will not be the ones with the smallest AI bill. They will be the ones who can tell you, line by line, what theirs bought.