Something strange is happening among Meta's 85,000 employees: a competition over who can consume the most AI tokens. Over 30 days, total usage crossed 60 trillion tokens — and the top individual burned through 281 billion tokens alone. At public pricing, that's one person spending over $1.4 million worth of compute.
This isn't just an internal meme. "Tokenmaxxing" is spreading across Silicon Valley as a genuine cultural shift, and Nvidia CEO Jensen Huang put it bluntly: "If a $500K engineer spent less than $250K on tokens, I'd be deeply concerned."
What Is It?
It started with an internal leaderboard called "Claudeonomics" — built voluntarily by a Meta employee and named after Anthropic's Claude model. It tracked AI token consumption across the workforce and published a ranked list of the top 250 users.
Claudeonomics — a homemade dashboard named after the Claude model
AI token consumption across all 85,000 employees, with a public ranking of the top 250
Gamified titles like Token Legend, Session Immortal, and Cache Wizard
60 trillion tokens in 30 days; the top individual consumed 281 billion tokens
Here's the thing — neither Mark Zuckerberg nor CTO Andrew Bosworth made the top 250. The heaviest users were engineers who actually use AI hands-on for coding, day in and day out.
But the leaderboard was taken down within two days of going public. Internal data leaked outside the company, and the creator shut it down voluntarily — leaving behind a note that read: "This was meant to be a fun look at tokens, but since dashboard data got shared externally, I'm closing it down for now."
What Changes?
Claudeonomics is gone, but the trend it reveals is much bigger.
| Traditional Performance Metrics | The Tokenmaxxing Era | |
|---|---|---|
| What Gets Measured | Lines of code, commits, completed tasks | AI token consumption, agent uptime |
| How People Are Rewarded | Salary, bonus, RSUs | Salary + token budgets (a new fourth element) |
| View of Productivity | Based on direct output | AI proficiency determines output |
| Risk | Overwork culture, burnout | Meaningless token consumption (input gaming) |
Silicon Valley leaders aren't shy about where they stand:
Andrew Bosworth (Meta CTO): "The best engineers are spending as much on tokens as their own salary, and their productivity is up 5–10x. Easy money. Keep going. No limits."
Jensen Huang (Nvidia CEO): "I'm going to ask every engineer at year-end how much they spent on tokens. If they say $5,000, I'm going to completely lose it."
Jensen Huang (GTC 2026): "Going forward, every engineer gets an annual token budget — roughly half their base salary, added on top as tokens to multiply their output by 10x."
Meta went even further. In November 2025, Chief People Officer Janelle Gale sent a company-wide announcement: "AI-driven impact will be a core expectation in performance reviews starting 2026." Individual AI usage is now an official performance metric.
And this isn't just a Meta story. OpenAI has its own employee token usage leaderboard — during one week in March, the top user consumed 210 billion tokens. Sendbird, a startup, awards the title "AI God" to employees who use over 100 million tokens a day, with rewards ranging from coffee gift cards to extra PTO.
But Is This Actually Working?
There's a classic management paper from 1975. Steven Kerr's "On the Folly of Rewarding A, While Hoping for B" — on the absurdity of incentivizing one thing while expecting another. It's no coincidence that Ethan Mollick, a Wharton professor, cited exactly this paper when commenting on the Token Legends phenomenon.
The core dilemma: We're rewarding token consumption (input) while hoping for real productivity (output). Some employees run AI agents in idle loops for hours just to climb the rankings. That's like judging a writer's skill by how much paper they've printed.
According to reporting from The Information, two current Meta employees confirmed that colleagues were "maxing out tokens on research tasks just to boost their rankings." Even Sendbird's CEO acknowledged: "Eight of our top 10 are genuinely productive — but the other two are more experimental."
Getting Started
The Token Legends phenomenon has a clear lesson: accelerate AI adoption, but don't fall into the measurement trap. Here's a framework for designing an AI adoption culture in your organization.
- Measure outputs, not inputs
Token consumption is just a leading indicator. What actually matters is the features shipped, lead times cut, and quality improvements made using AI. Meta's mistake was gamifying the input. Build dashboards — but make them show "results changed by AI use," not raw consumption. - Treat token budgets as investment, not cost
Jensen Huang's "half your salary in tokens" formula sounds extreme, but the core idea is right. Think of AI compute as infrastructure investment, not overhead. Just make sure you're running per-team ROI dashboards alongside it. - Use gamification carefully
Meta's Level Up game and the Claudeonomics leaderboard work well for early-stage adoption. But the moment chasing titles becomes the goal, distortion follows. Use usage badges during the exploration phase — then switch to outcome badges once AI use is established. - Include AI capability in performance reviews — but evaluate how, not how much
Meta's direction of baking AI-driven impact into performance reviews is right. But the standard shouldn't be "how much did you use it" — it should be "how did you use it, and what did you actually achieve." - Design guardrails first
Per-team token quotas, cost dashboards, purpose tagging for usage. Claudeonomics came down in two days because it raised transparency without any guardrails. Roll out data security policies alongside any usage initiative.
Deep Dive Resources
"On the Folly of Rewarding A, While Hoping for B" — A classic on incentive design. It covers the organizational distortions that emerge when what you measure and what you actually want are two different things. This one paper explains exactly why Meta's Token Legends setup is risky.
A roundup of tokenmaxxing culture across Silicon Valley — Meta, Nvidia, Databricks, Sendbird, and more. Packed with specific examples of how companies are actively encouraging employees to consume more compute.
The story behind AI-driven impact becoming a core expectation in Meta's 2026 performance reviews. Covers the context behind Janelle Gale's internal memo and initiatives like the Level Up game.
