Uber cuts AI coding-tool spend to $1,500 per employee per tool each month

• Uber told Bloomberg it now limits employees to $1,500 a month in token spend for each AI coding tool, covering products like Cursor and Claude Code, according to simonw's linked report and Bloomberg's report.
• The cap is per tool, not one shared pool, so an engineer using multiple agents can still run multiple four-figure budgets in parallel, as simonw noted while PYMNTS' writeup added that employees can request exceptions.
• zeeg's reaction and hwchase17's post both framed the move as a sign that inference costs have graduated from experimental slush fund to managed operating expense.
• Simon Willison's back-of-the-envelope math in simonw's tweet and his blog post puts two active tools at a $36,000 annual ceiling per engineer, which is large enough to sit next to compensation and headcount planning, not ordinary SaaS seat budgeting.

You can read Bloomberg's original report, Simon Willison's concise cost breakdown, and a shorter TechCrunch follow-up. The oddest useful detail is that the limit is attached to each tool separately, while PYMNTS says Uber also exposes usage in an internal dashboard and allows employees to ask for permission to go over.

The core detail is unusually concrete for enterprise AI spend. Bloomberg reported that Uber capped employees at $1,500 in monthly token spending for each AI coding tool, and that the policy applies to agentic coding products such as Cursor and Claude Code, per Bloomberg's report.

That per-tool wording matters more than the headline number. As simonw emphasized, spending on one coding agent does not reduce the allowance for another, which makes this a product-level budget control rather than a single companywide wallet.

Simon Willison's framing in his post is the clearest quick benchmark: if an engineer actively uses two tools, the cap implies a theoretical ceiling of $3,000 a month, or $36,000 a year. He compared that with Levels.fyi's median Uber software engineer compensation package of about $330,000, which lands the cap at roughly 11 percent of compensation.

That helps explain why the reaction on X converged on cost discipline, not product quality. zeeg argued that $18,000 a year for one tool is easier to absorb than the $20,000 to $100,000 range some teams appear to be drifting toward, while hwchase17 treated Uber's policy as evidence that “costs start to matter.”

A few replies pushed the logic further. Yuchenj_UW's post cast the choice as a tradeoff between unlimited token burn, budgets, and headcount, which is glib but captures why finance teams care about these numbers in a way they never cared about ordinary editor subscriptions.

The follow-on operational detail is that Uber reportedly did not just set a cap and hope for the best. PYMNTS says employees can track usage across tools in an internal dashboard, and can seek permission to exceed the default limit.

That makes the policy look closer to cloud spend governance than seat licensing. TechCrunch's summary likewise described the new rule as a monthly cap per employee and per agentic tool, with usage visible to employees rather than hidden in back-office procurement systems.

The budget cap landed a month after Uber management started putting numbers on actual AI-assisted output. In a May earnings-call report cited by Business Insider via Yahoo Finance, CEO Dara Khosrowshahi said autonomous agents produce about 10 percent of Uber's code changes, with humans still reviewing the result before merge.

rohanpaul_ai paired that 10 percent figure with the new spending limit, which is the real reason this story is useful as a reference point. Plenty of companies say they are “using AI.” Uber is one of the few that has now put hard numbers on both sides of the ledger: a measurable share of code output, and a measurable per-tool ceiling on what it will pay for the agents producing it.