WritingNovember 3, 2026
Uber Blew Its AI Budget in Four Months. The Metric Was the Problem.

Uber blew through its entire 2026 AI coding budget in four months. The fix it reached for — a hard dollar cap per employee — treats the symptom and ignores the disease.
If you're running Claude Code, Cursor, or Copilot across even a five-person team, this is worth ten minutes. Not because you're Uber-sized. Because the thing that broke Uber's budget is the same thing sitting un-tracked in your own billing dashboard right now: token spend that has no relationship to what actually shipped.
The budget burned in four months
Uber encouraged its engineers to use AI coding tools "as much as possible." It ran internal leaderboards ranking usage. Then, by April, the company's own CTO told the org it had spent its entire planned 2026 AI budget — with eight months left on the calendar [1].
The response: a $1,500-per-employee-per-month cap on agentic coding tools, applied separately to Claude Code and to Cursor. Spending is now trackable on an internal dashboard, and employees can request an exception if they need to go over [1]. Uber's COO, Andrew Macdonald, put the real problem plainly: "it's very hard to draw a line" between how much a team spends on AI and what consumer-facing feature actually came out the other end [1].
That's not a budgeting failure. That's a measurement failure. Uber was tracking activity — tokens burned — and had no number connecting that activity to output. You can't manage what you're not measuring, and "tokens consumed" was never the thing that mattered.
Source: TechCrunch — Uber caps employee AI spending after blowing through budget in four months
Why a flat cap is a lazy fix
A per-seat dollar cap punishes your best and worst users identically. The engineer burning tokens on a genuinely hard migration and the one burning tokens because they let the agent retry a broken prompt twenty times both hit the same wall.
Worse: a cap doesn't tell you anything. It just stops the bleeding at an arbitrary number. Six months from now, Uber will still not know whether $1,500 a month per engineer is cheap or expensive, because the cap was never connected to what that spend produced.
A dollar cap tells you when to stop spending. It never tells you whether the spending was worth it.
What to track instead
The fix is boring and it's not a new idea: stop counting tokens, start counting cost per merged pull request. Take total AI-assisted development spend for a period and divide it by the number of changes that actually got reviewed and merged. Track rework and defect rate alongside it, so a cheap-looking number that ships broken code doesn't hide [2].
This single change flips the conversation. "We spent $40,000 on AI coding this month" is a scary, context-free number. "We spent $40,000 and shipped 600 merged PRs at a lower defect rate than last quarter" is a result you can defend to a founder or a client.
It also tells you something a token count never will: which tasks are worth routing to an expensive frontier model, and which are routine enough to hand to something cheaper. HPE says it built exactly that kind of routing — a hybrid platform sending workloads from more than 150 internal applications to on-premises open-weight models — and that doing so cut its own token costs 30 to 50% [2]. That's HPE's own figure about its own platform, not an independently audited number, but the mechanism is sound regardless of who's reporting it: not every coding task needs your most expensive model.
| Metric | What it measures | What it misses | Where it leads |
|---|---|---|---|
| Raw token spend | Size of the API bill | Whether the tokens shipped anything | Panic caps, like Uber's $1,500/tool/month limit |
| Usage leaderboards | Who uses the tool most | Rework, wasted calls, wrong-model routing | Rewarding the heaviest user, not the most effective one |
| Cost per merged PR | Dollars per shipped, reviewed change | Nothing you weren't already tracking in git | Routing decisions — which tasks need a frontier model |
| Cost per PR + defect rate | Cost and whether the shipped thing held up | — | Catches "fast but sloppy" before it costs you in rework |
The 2028 cliff nobody's budgeting for
This isn't a one-company problem. Gartner predicts that by 2028, AI coding token costs will overtake the average developer's monthly salary — and its baseline for that salary is roughly $2,000 a month globally [3]. Gartner's own research already finds 6% of organizations paying more than $2,000 per developer per month in AI token charges today [3].
Part of why the math gets ugly fast: a single agentic coding task can trigger five to thirty separate model calls, and output tokens cost three to five times more per unit than input tokens on Claude models [3]. Nitish Tyagi, a senior principal analyst at Gartner, named the actual failure mode: "Token discipline will not emerge through developer choice alone, as developers tend to optimize for speed and convenience over cost efficiency. Without a governed engineering operating model, costs can escalate faster than the productivity gains these tools are designed to deliver" [3].
Read that twice. Nobody expects your engineers to self-police token spend, because that's not their job and it was never going to be. If there's no metric connecting spend to output, the spend grows until someone in finance notices — which is exactly what happened at Uber.
flowchart TD Start([You scale AI coding<br/>seats across the team]) --> Q1{Tracking raw token<br/>spend only?} Q1 -->|Yes| A[Spend looks fine<br/>right up until it isn't] A --> B[Budget exhausted<br/>mid-year, no warning] B --> C[Panic: blanket<br/>per-seat dollar cap] C --> D[Cap punishes your best<br/>and worst users the same] Q1 -->|No — tracking<br/>cost per merged PR| E[Cost is tied<br/>to shipped output] E --> F{Cost per PR<br/>climbing?} F -->|Yes| G[Investigate: rework,<br/>or wrong model for the task] F -->|No| H[Scale seats<br/>with confidence] G --> I[Route routine codegen<br/>to a cheaper model]
The SMB-sized version of this audit
You don't need Uber's engineering headcount to have Uber's problem. You need exactly two things it didn't have in January: a metric that ties spend to output, and a routing decision for which tasks get your expensive model.
Here's the five-minute version. Pull last month's AI coding spend by tool. Divide it by merged PRs for that same period — not commits, not lines of code, merged and reviewed changes. If you don't have that number today, that's the actual finding, not a footnote. Then ask the harder question: of that spend, how much went to boring, repetitive codegen that a cheaper or open-weight model could've handled just as well? That's the stack audit I run for clients, and it's the same instinct that turned a $1,800/month Zapier bill into a $40/month n8n instance for The Hub — the tool wasn't the problem, the missing metric was.
So what do you do with this
If your team's AI coding bill feels fine right now, that's not evidence you're fine. It's evidence you haven't hit your April yet. Pull the cost-per-merged-PR number this week, before finance pulls it for you. If the number doesn't exist, building it is a half-day of work, not a project — and it's the kind of audit I run as part of fractional AI automation engagements for teams that would rather find this out on their own schedule than their CFO's.
Sources
[1] TechCrunch — Uber caps employee AI spending after blowing through budget in four months — techcrunch.com
[2] MIT Technology Review — Stop counting tokens, start counting what ships — technologyreview.com
[3] Gartner — Gartner Predicts AI Coding Costs Will Surpass Average Developer's Salary by 2028 as Token Consumption Surges — gartner.com
The short version
- Uber exhausted its full 2026 AI coding budget by April, then capped Claude Code and Cursor spend at $1,500 per employee per month
- A flat spending cap doesn't fix anything — it just stops the bleeding at an arbitrary number with no link to output
- The better metric: cost per merged PR, tracked alongside rework and defect rate
- Gartner predicts AI coding token costs will overtake the average developer's salary by 2028; 6% of orgs are already past $2,000/developer/month
- Route routine codegen to a cheaper or open-weight model once you can actually see which tasks are worth the expensive one
- If you can't produce your own cost-per-merged-PR number today, that gap — not your current bill — is the real finding
Drafted with Claude, reviewed and edited by Bryan before publish.
