The story went round every tech newsroom: Uber consumed its entire 2026 AI budget in four months, then introduced a cap of $1,500/month per employee per agentic coding tool - Anthropic's Claude Code and Cursor are named explicitly. The news, which started as a Bloomberg dispatch, was relayed by ActuIA and Challenges, and analysed by Simon Willison on his blog.
The dominant narrative: euphoria gives way to sobriety, the bill catches up with the promise. That is true. But stopping there misses the point. Uber's budget overrun is first and foremost a symptom of success.
An overrun, yes - but because it works
Let us look at the numbers before sounding the alarm. In early 2026, close to 95% of Uber's roughly 5,000 engineers were using AI tools every month. The company had even gamified adoption with an internal leaderboard. The result: it worked - perhaps too well. An engineer was generating between $500 and $2,000 of tokens a month; according to Pragmatic Engineer, 84% of Uber's developers already use coding agents.
The budget itself had been set in late 2025, before Claude Code and Cursor became everyday tools. That is the whole problem: a 2025 usage pattern was budgeted for a 2026 reality. As Simon Willison put it, the $1,500/month cap is not a brake, it is a rational price signal for the industry - roughly 11% of an Uber engineer's total package. At that price, if the productivity gain follows, the maths still works comfortably.
Uber's mistake was not aggressive AI adoption. It was setting adoption targets without matching cost governance.
The real issue: drawing a line between spend and value
The most lucid passage comes from Uber COO Andrew Macdonald, who admitted on a podcast that it is "very hard to draw a line" between token consumption and useful features shipped to users. The link, he says, "isn't there yet". That is the heart of the problem - and it is not financial, it is organisational.
When 10% of the code is written by agents, that does not mechanically translate into 10% more features. Without connecting spend to an outcome - a feature, a version, a deliverable - the "AI" line item stays a growing number with an uncertain return. Uber's answer (a uniform cap) is rational but blunt: it is a tourniquet, not steering. The real answer is instrumentation.
Our conviction: tokens are the customer's money. A blind cap treats every use the same. Yet a session that ships an expected feature has nothing in common with a session going in circles. The right unit of measure is not "per employee", it is "per feature shipped".
Measure better, don't spend less
The lesson from Uber is not "slow down adoption". It is "build the governance that makes adoption sustainable". On that ground a few principles are broadly agreed - and they are exactly the ones we apply at Survol:
- Attach every token to its object. Track consumption per session, per version and per feature - not just per person. That is what finally lets you "draw the line" Uber is looking for.
- Dashboards from day one. Visibility changes behaviour before limits do. Many engineers have no idea what their usage really costs until they see it.
- The right model for the right task. Reserve heavy agentic sessions for refactors and multi-file debugging; matching the model to the task can cut the bill by 40 to 60% with no loss of productivity.
- A tailored context. Send the spec, the decisions and the files concerned - nothing more. A minimal context means fewer tokens at equal quality.
- Caps with exceptions, not walls. Uber allows case-by-case overruns: friction discourages waste without blocking high-value work.
# Capping "per employee" vs attaching spend to value uber_cap: "$1,500/month/tool" # blind: a tourniquet survol_tracking: feature: "Promo codes & gift cards" version: "v2.0" tokens: "42,100 - model selected for the task" delivered: "preview approved on 18/06" # the line is drawn
Worth noting too: Uber already gives every employee a dashboard of their consumption and allows managed exceptions. That is the right direction. What is missing is the link between that tracking and what is actually produced. A Bain & Company survey quoted by Bloomberg confirms it: 40% of companies that track their AI spending have not met their cost-reduction targets - yet 83% of CFOs plan to raise their AI budget by more than 15% over two years. The money keeps coming. What is missing is the compass.
The takeaway
- An annual AI budget consumed in four months is first of all a sign that the tool found its audience.
- $1,500/month per tool is now a market reference point: companies are willing to pay if the ROI follows.
- A uniform cap is a patch; the durable solution is connecting every token to a feature shipped.
- Every team deploying agents today will have this budget conversation within 6 to 12 months. Better to have it on your own terms.
Uber did not discover that AI is expensive. It discovered that it was flying without a dashboard connected to the product. That is a lesson you can learn from reading the press - or by fitting the right instruments right now.
