Uber reportedly spent its entire 2026 AI budget by April. The cause was not a failed project. It was a successful one, after the company expanded its use of Claude Code.
That is the strange part of agentic AI economics. The more useful the tool becomes, the more it costs, and the bill looks the same whether the money produced software or produced loops.
A meter cannot tell value from waste
A recent piece in MIT Technology Review, written with HPE, laid out the mechanics. When a developer hands a task to a coding agent, the agent plans, writes code, runs tests and corrects itself. One task can use 10 to 50 times more tokens than a basic chat prompt.
The pressure is visible across the market. GitHub Copilot moved from flat-rate pricing to metered credits. Gartner predicts that AI coding token costs will pass average developer salaries by 2028.
The usual reaction is rationing. Usage caps, seat limits, approval queues before anyone launches an agent. I understand the instinct, because I have watched finance teams do it with every new line item for sixteen years.
But a cap on tokens is a cap on consumption, and consumption is the wrong number.
Some token spend is waste. Retries, runaway loops, an agent rewriting the same function four times. Some of it is the actual work of shipping software. A spending cap cuts both with the same blade.
Count the thing that ships
The metric the article proposes is cost per merged pull request. Total AI-assisted development cost, divided by the number of changes that pass review and get merged. Track rework and defects next to it, so cheaper changes do not hide lower quality.
If that number falls while AI usage rises, the team is converting consumption into output. If it rises, the extra spend is going somewhere other than finished work. That is a conversation you can actually have with a CFO.
Marketing went through the same lesson with paid media. For years, teams reported cost per click and cost per lead because those numbers were easy to pull. The useful number was always cost per closed deal, and it took a decade for most companies to wire the data together to see it.
AI spend is at the cost-per-click stage right now. The invoice is easy to read. The output is spread across repositories, tickets and review queues, so almost nobody connects the two.
The spend lives with a few people
The most useful data point in the piece is about distribution. At HPE, the top 5% of users generated most of the consumption, and the heaviest users consumed roughly 450 times more than the lightest on the same license.
That breaks per-seat budgeting completely. If power users grow from a small minority to half the workforce, monthly spend can rise about 4.4 times with flat headcount. Nothing about the org chart changes, and the bill quadruples.
Garry Tan gave the extreme version on Walter Isaacson's podcast, reported by TechCrunch. The Y Combinator CEO said he spends about $100,000 to $200,000 a year on tokens. He also published his set of 23 Claude Code skills, called gstack, so other people can work the same way.
I use a version of that setup myself, and I believe the spend is justified for someone working at that level. The point is that one person can now cost more in tokens than a junior employee costs in salary. Averages across licenses hide that completely.
Route the work, not the person
The fix is architecture, not restriction. Agentic coding does not need the most powerful model at every step. Planning, novel reasoning and hard debugging justify a frontier model. Generating tests, repeated refactoring and boilerplate usually do not.
A gateway can route each request by complexity, data sensitivity, cost and latency. HPE says it routes routine workloads from more than 150 internal applications to open-weight models it runs on its own infrastructure, and cut token costs by 30 to 50%. The line I liked most: meter the exception, not the workload.
We apply the same logic when we build agent systems at Madison AI. The expensive model writes the plan and checks the result. Cheaper models do the repetitive middle. The quality of the output depends on the plan and the check, not on paying frontier prices for every keystroke.
Owned capacity is not for everyone. Low or unpredictable usage favours paying per call, with no capital tied up in idle hardware. Ownership pays off when sustained demand keeps it busy. Most companies will end up hybrid.
The routing layer is what makes hybrid work.
What I would measure this quarter
I would start with two numbers, the same two HPE says it tracks. Cost per active user, because averaging across every license hides the power users who drive the bill. And cost per finished unit of work, which for engineering is a merged pull request.
For marketing and operations teams, the unit is different but the logic holds.
Cost per published asset that passed review. Cost per qualified lead the agent actually researched. Cost per support ticket closed without escalation. Pick the unit your business already values.
Baseline both numbers before you change vendors, caps or infrastructure. Otherwise you will not know whether the new setup improved anything or just moved the bill.
I wrote earlier about how AI costs are getting metered and about what happened when Uber hit its cap. The metering was the first act. Measurement is the second.
The real risk is not that AI gets expensive. It is that companies cut the workflows producing value because the only number they can see is the invoice.
Count what ships. The token bill is just the receipt.