How Uber Burned Its Entire 2026 AI Budget by April

One number should make every executive planning an AI budget pause. According to reporting, Uber burned through its entire 2026 AI budget by April. A company with world-class engineering and effectively unlimited resources ran out of AI money a third of the way through the year.

Read that not as a story about Uber, but as a warning about a structural trap almost everyone is walking into. The problem underneath it is simple and widespread: organizations are struggling to connect AI spending to measurable returns, and when spend floats free of value, it doesn’t plateau — it runs.

In short:

• Reports say Uber burned through its entire 2026 AI budget by April — a vivid symptom of a wider problem: AI spend that isn’t tied to measurable return.
• Costs run away when agents are given open-ended tasks with no ceiling, no scope, and no unit economics.
• The fix isn’t spending less on AI. It’s knowing the cost and the value of every task before you scale it.
• Scoped, measured builds are boring on a slide and beautiful on a P&L.

Why AI Costs Run Away

Traditional software has predictable costs. You buy a license, you know the number. AI, especially agentic AI, doesn’t behave like that. Its cost scales with usage in ways that are easy to underestimate and hard to see until the bill arrives. Give an agent an open-ended task — “research this,” “handle that” — with no ceiling and no scope, and it will happily spend tokens generating, retrying, checking, and re-checking. Each individual action looks cheap. Multiply it across every user, every day, with no unit economics underneath, and cheap becomes catastrophic.

The seductive part is that it often works. The output is good. So usage grows, and the cost grows with it, silently, until someone in finance asks why the line has gone vertical. By then you’ve built dependence on a system whose economics you never modeled.

The Missing Number

Here’s the discipline almost nobody applies before scaling: what does one task actually cost, and what is one task actually worth? Both numbers, side by side. A workflow that costs a little more per task but returns a faster, better, more reliable result can be radically more efficient than one that looks cheaper on paper and then generates retries, reviews, and rework downstream. Cheapest-per-token is not the same as cheapest-per-outcome, and the gap between them is where budgets die.

If you can’t state the cost per task and the value per task for an AI workflow, you are not ready to scale it. You’re ready to pilot it, measure both numbers, and only then decide. Scaling an unmeasured workflow is how you become the cautionary tale in someone else’s article.

Scope Is a Cost Control, Not a Limitation

This is why scoped builds beat open-ended ambition on economics as well as on delivery. A system built for a defined workflow has a defined cost envelope. You know what it’s allowed to do, how often, and to what end. It can’t quietly wander into an expensive loop because it wasn’t handed an open mandate. That constraint isn’t a weakness — it’s the thing that keeps the P&L honest.

The runaway-cost failures share a shape: broad mandate, no ceiling, no per-task accounting, scaled on the strength of a good demo. The controlled successes share the opposite shape: narrow mandate, known economics, scaled only after both numbers checked out.

What Discipline Looks Like

Three practices keep AI spend sane. First, scope every workflow to a defined job with a defined cost envelope — no open-ended agents in production without a ceiling. Second, measure cost per task and value per task before you scale, and refuse to scale anything where the second number isn’t clearly bigger than the first. Third, monitor spend continuously, not quarterly, so a runaway is caught in days, not discovered in the annual review.

None of this means spending less on AI. Spend as much as the returns justify — that’s the whole point. It means never spending on AI you can’t account for. The companies that win won’t be the ones who spent the most or the least. They’ll be the ones who always knew, task by task, exactly what they were paying for and exactly what they got back.

“If you can’t state the cost per task and the value per task, you’re not ready to scale it.”