Several large technology firms experienced rapid growth in AI token consumption after deploying Anthropic's Claude, producing unexpectedly large invoices and budget overruns.

Internal leaderboards and tokenmaxxing

Within Meta, a voluntary intranet leaderboard called Claudeonomics ranked 85,000 employees by AI token usage and created incentives to increase consumption.

Over a 30-day period, Meta consumed 60T tokens, with the top user burning 281B tokens per day on average during that month.

«My top engineer spends as much on tokens as he earns in salary. But he is 5–10 times more productive. It's easy money. No limits.»

Industry signals and managerial comments

Executives from other vendors commented on the trend, suggesting high token spending had become a proxy for productivity in some teams.

One remark attributed to Jensen Huang indicated concern if an engineer with a $500 K salary did not expend at least $250 K on tokens, reflecting market pressures on usage.

Budget exhaustion and company examples

Pravin Neppali, CTO of Uber, confirmed his company exhausted its AI budget for 2026 within four months after wider adoption of Claude Code tools.

The share of engineers using the service reportedly rose from 32% to 84%, with monthly token expenses per engineer ranging from $500 to $2K.

Large bills and agent-driven consumption

An anonymous CFO reported that Anthropic issued a charge of $500 M for a single month where no consumption limits had been set for agents.

The explanation centers on autonomous agents rather than simple chat queries; such agents can run continuously and consume orders of magnitude more tokens.

• Meta: internal leaderboard and 60T tokens burned in 30 days.
• Uber: full 2026 AI budget spent in four months after wider adoption.
• OpenClaw team (Peter Steinberger): approximately $1.3 M in token costs in one month.
• Anonymous finance report: a single month invoice of $500 M.

Implications for corporate AI spending

These cases illustrate how rapid adoption and lack of consumption controls can generate unexpectedly large operational costs for enterprise AI deployments.

Companies have faced trade-offs between granting open access to agent tooling for productivity gains and enforcing strict token limits to control expenses.