Editorial documentary photograph of a quiet finance executive's desk; a framed editorial illustration above the desk shows a small stack of token-shaped chips with one corner of the stack burning steadily and a thin trail of smoke rising upward, the rest of the stack intact but visibly being consumed.
Posted on by Gérard

Your AI budget is already gone.

Three Uber executives, three different seats, told the same story this spring. The cost category most boards govern quarterly is moving to hourly.

By April, Uber’s CTO had blown the AI budget he set in December. Three weeks later, the CEO said he was metering headcount and leaning further in. Two weeks after that, the COO asked aloud whether any of it was producing value. Three quotes, three seats, one cost category nobody had experience with. Here is why token spend breaks the quarterly cadence finance was built on, and the three questions a board should be asking by the next meeting.

Praveen Neppalli Naga, Uber’s CTO, told The Information in April: “I’m back to the drawing board, because the budget I thought I would need is blown away already.” Claude Code adoption at Uber had moved from 32% to 84% of a 5,000-engineer organization, at $500 to $2,000 per engineer per month.

Three weeks later, on the Q1 earnings call, CEO Dara Khosrowshahi reframed the same fact as strategy. “If every person at this company can increase their throughput by 20%, 30%, 50%, 100%, then metering headcount growth and leaning in on AI investment is going to be well worth it.” Roughly 10% of code changes at Uber, he added, are now produced by autonomous agents.

Two weeks after that, COO Andrew Macdonald said the quieter part out loud to Business Insider. After talking to senior engineering leaders, he could not find a proportional link between token consumption and useful new features for riders. “That link is not there yet, right?”

Budget gone. Strategy doubled down. Value unproven. All three are right at the same time, which is the new shape of the problem. If Uber can blow through a 12-month AI budget by April and publicly disagree about whether it bought anything, the issue is not competence. The standard IT cost playbook was built for categories that do not behave like tokens.

The cost category nobody has experience with

Three structural facts make tokens different.

The first is shape. Server costs grow in steps. SaaS seats grow in contracted increments. Token spend grows on a continuous curve whose slope is set by how the workload is built, not by how many people use it. A chat interaction runs about 2,000 tokens. An agentic workflow, the kind where the model plans, retrieves, calls tools, and checks its own output, can run 1,000 times faster on the same task. Stanford’s Digital Economy Lab measured a 30x cost variance running the same agent on the same task twice. The same employee can multiply the run-rate by Tuesday because someone shipped a better agent on Monday.

The second is elasticity in the wrong direction. Unit prices fell by roughly 99.7% across the leading models in 18 months, according to NavyaAI’s tracking. Enterprise AI bills tripled to $37 billion over the same period. Cheaper inputs produced higher bills because adoption breadth and agentic depth expanded faster than unit prices compressed. Jevons in a hurry. The CFO who modeled a price drop into a forecast and called it discipline is already wrong.

The third is feedback latency. Finance’s tightest cadence is the monthly close. Token spend moves on hours. A new model deploys on Wednesday at 3 p.m.; by Friday, the run-rate has reset; the monthly report, when it lands three weeks later, describes a regime that no longer exists. Governance whose tightest loop is the monthly close is running open-loop against the thing it is supposed to govern.

Editorial typography card on a warm cream background, with a large Source Serif 4 statement reading "Unit prices fell 99.7%. Enterprise AI bills tripled," a thin horizontal Orion-blue rule beneath it, and a smaller Source Sans 3 attribution line reading "AI tokens, 2024 to 2026. NavyaAI, FinOps Foundation."

Why the obvious fixes are the wrong fixes

The first instinct, when a budget blows up, is to cap it. The second is to centralize approvals. Both push the highest-leverage decision (what is this AI doing for us, and is it worth it?) onto the people furthest from the answer.

Capping team-level spend taxes the workloads with the highest ROI. The team building an agent that replaces three vendors gets throttled to the same ceiling as the team running chatty copilots that produce nothing. The cap penalizes both, and the second one was supposed to be killed anyway.

The other temptation, the one Macdonald named when he said tokenmaxxing, is the AI equivalent of cutting marketing to hit a margin target. Drive consumption up to look AI-native, or down to look disciplined. Declare victory. Quietly stop asking what they bought. The discipline version of theatre.

What is missing in both cases is the lines on the P&L.

The line you do not have

The line that needs to exist, in every company past the experimentation phase, is the one revenue already has. Owned. Forecast. Reviewed. Decomposed by group. Tied to a unit of value the business already counts.

Four moves make it real.

Build AI consumption into the budget architecture as its own line item, not buried inside cloud or R&D. Decompose by group the way revenue is decomposed by segment.

Set the cadence to weekly. The monthly close is the wrong instrument for a cost that moves daily. The FinOps Foundation’s emerging AI framework is one useful starting point; treat it as vocabulary, not as the answer.

Measure cost per outcome, not cost per token. Cost-per-token rewards underuse. Cost-per-ticket-resolved, cost-per-qualified-lead, and cost-per-shipped-PR rewards workloads that earn their tokens and starve the ones that do not. If a workload cannot name its unit, it probably should not exist yet.

Make the whole C-Suite accountable. Not the CFO alone; the CFO and the head of the function whose work is being multiplied. The CFO defends the math, the operator defends the value. Neither alone is enough.

What a board should be asking by next month

If you sit on a board or run an audit committee, Uber is a free dress rehearsal. Three questions will tell you whether management is governing this category or merely talking about it.

What is the total AI spend by the group this week, in dollars? Not annualized. This week.

What unit of value are you measuring against, and what is the cost per unit by group?

What cadence reconciles actuals against forecast, and who owns the variance?

If any answer hesitates, the company is in the same place Uber was in December. The calendar has simply not run far enough to find out.

The interesting move is not to spend less on AI, and not to spend more. It is to know, on a Tuesday, what you spent on Monday and what it bought you. Until that loop closes, the budget conversation is happening in the wrong room, at the wrong speed, about the wrong number.

Editorial watercolor illustration on a warm cream paper background showing a single closed clockwise loop drawn as a confident hand-painted brushstroke; four small watercolor markers sit at the cardinal points around the loop, and the closing arc from the lower-left marker back to the upper-left marker is rendered in a slightly heavier deep blue brushstroke with a small mint dot at its arrowhead, marking the step that closes the loop.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *