Uber, Meta, Amazon impose caps on employee AI tool spending
In brief
- Uber exhausted its full 2026 AI budget by April, spending three times the planned rate.
- The company implemented a $1,500 monthly cap per employee on agentic coding tools.
- Meta employees engaged in 'tokenmaxxing'—competitive overuse consuming tens of trillions of tokens.
- Meta and Amazon removed internal AI-usage leaderboards after costs spiraled exponentially.
- Walmart, Microsoft, and AT&T imposed similar restrictions as per-employee costs climbed.
The Runaway Cost Problem
Uber's entire AI budget for 2026 was gone by April. The company had allocated funds expecting a gradual ramp-up in adoption and usage. Instead, the company was spending at roughly three times the rate it had planned. In response, Uber rolled out the $1,500 monthly cap on June 2, 2026, capping per-employee spending on tools like Claude.
The pattern isn't unique to Uber. Meta indicated in mid-June 2026 that it plans to limit AI tool usage, citing an "exponential increase" in costs. Amazon has signaled similar restrictions. Walmart, Microsoft, and AT&T have all imposed similar usage restrictions as per-employee AI costs climbed into the thousands.
How Internal Competition Amplified Costs
What drove Meta's spending spike? Employees engaged in a practice dubbed "tokenmaxxing," essentially competitive overuse of AI tools. This resulted in the consumption of tens of trillions of tokens.
The culprit was organizational design. Both Meta and Amazon had set up internal AI-usage leaderboards, originally designed to encourage productivity and adoption. Instead of driving efficient usage, the leaderboards created perverse incentives. Employees raced to log the highest token consumption, treating AI tool access like a competitive sport. Both companies subsequently removed the leaderboards.
The lesson is sharp: visibility without guardrails amplifies costs. Leaderboards work for metrics you want to maximize (bugs fixed, features shipped). They backfire when you're trying to manage spend. Tech leaders are now learning this the hard way.
