Summary: Bloomberg broke the news on June 2, 2026: Uber now limits every employee to $1,500 a month on each agentic coding tool, with Claude Code and Cursor both covered by the rule. It also quietly kills the leaderboard logic: what counted as a virtue signal in Q1 had turned, by mid-year, into a line item someone has to defend. An internal dashboard shows how much each person has consumed, and with the right approval the ceiling can be lifted in specific situations.
Bloomberg broke the news on June 2, 2026: Uber now limits every employee to $1,500 a month on each agentic coding tool, with Claude Code and Cursor both covered by the rule. An internal dashboard shows how much each person has consumed, and with the right approval the ceiling can be lifted in specific situations.
What happened beforehand is the bigger story: by April, Uber’s CTO had already conceded that four months were all it took to exhaust the company’s annual AI budget. Swapping “use AI as much as possible” for month-by-month rationing puts Uber among the first large US tech companies to formally cap how much AI its employees may use — and drags enterprise AI cost governance out into the open.
When “Use More AI” Turns Into Rationing
Earlier reporting from The Information described how Uber pushed staff to use AI “as much as possible,” going so far as to rank employees against each other on internal leaderboards by how much AI they consumed. April flipped the mood: the CTO demonstrated internally that tools like Claude Code are capable of devouring an AI budget — a full year’s allocation lasted barely a third of the year. The policy announced on June 2 amounts to a formal confession that agentic tool usage, left unchecked, now generates per-person costs large enough to demand limits. It also quietly kills the leaderboard logic: what counted as a virtue signal in Q1 had turned, by mid-year, into a line item someone has to defend.
Look at how it is built: a ceiling, a dashboard, and an exception path — not a ban. Flexibility survives, and for the first time every employee gets an unobstructed view of their personal AI bill.
How Far Does $1,500 a Month Go
Token consumption is the billing basis for agentic coding tools — a world away from the old flat per-seat subscription. An engineer who leans heavily on Claude Code or Cursor, with agents endlessly reading and rewriting code, invoking tools, and retrying failed steps, will reach a four-figure monthly bill without much effort. This is the very reason developers howled when GitHub moved Copilot to token-based billing in late May: the pricing model is still shaking itself out, yet finance departments are expected to plan a full year around it.
As a compromise, $1,500 lands in sensible territory: roomy enough for everyday engineering, firm enough to cut off runaway agent loops and the “let it run overnight” brand of waste. It also opens the conversation the industry has dodged so far — namely, how much AI spend per engineer per month is actually reasonable.
Measuring Value Is the Real Bottleneck
Saving money is not the hard part; proving what the money bought is. Speaking on a podcast in May, COO Andrew Macdonald acknowledged that it is “very hard to draw a line” between AI usage and new consumer features actually shipping. A Bain survey that Bloomberg cited on June 1 points the same way — AI has delivered less cost reduction than most firms predicted — and the Wall Street Journal finds Corporate America already starting to ration AI as expenses soar.
Every company shares Uber’s bind: money goes out on the spot and lands on the ledger, while the payback hides inside productivity gains that resist clean attribution. Burning the annual budget in four months merely set that structural fault off ahead of schedule. A cap is a tourniquet, not a cure — the cure is running AI spend as an engineering resource with unit costs rather than handing it out like a perk. Until teams can price a shipped feature in tokens, budgets will keep getting written in a panic instead of by plan.
Three Takeaways for Engineering Teams
Three lessons worth acting on. One: monitoring for agentic tools belongs inside the system from day one, not bolted on after the budget blows up — dashboards, unit costs, and anomaly alerts should ship as part of the default setup. Two: pricing is sliding from per-seat to consumption-based, so procurement contracts and cost models need rework, and invoices broken down to the token level will only multiply. Three: the numbers to watch are completed tasks and rework rate, not raw token counts — the very theme our lessons from Cursor’s cloud agents keeps returning to: the more agents you run, the heavier governance and feedback loops weigh.