Meta's 85,000 Employees Burned 60 Trillion Tokens in One Month

Everyone assumes more AI usage equals higher productivity. The opposite is happening.

Meta just revealed the scale of corporate tokenmaxxing—their 85,000 employees consumed 60 trillion tokens in a single month. The company was so proud they built an internal leaderboard to rank workers by consumption. Until the backlash hit and they quietly removed it.

This isn't just Meta being Meta. It's a Silicon Valley-wide delusion where organizations treat token volume as productivity proof. Some companies now tie AI usage directly to employee compensation. The logic seems bulletproof: more tokens = more AI work = higher output.

Except it's completely backwards.

The real story: Higher token consumption typically means lower quality outcomes. More expensive bills. Slower systems. Noisier outputs. When developers optimize for token count instead of results, they generate:

• Redundant prompts that repeat the same request
• Over-iteration on problems already solved
• Massive context windows stuffed with irrelevant data
• Complex reasoning chains for simple tasks

It's the workplace equivalent of measuring programmer productivity by lines of code. We learned that lesson decades ago, yet here we are again.

The Elephant in the Room

Tokenmaxxing reveals something darker about developer psychology right now. As AI automates coding tasks, programmers face existential pressure about their career relevance. Some are responding with performative metrics—generating impressive-looking AI activity to prove their value.

The leaderboards feed this anxiety perfectly. Visible, quantifiable, seemingly innovative. Perfect for dashboards and performance reviews. Terrible for actual productivity.

Reid Hoffman tried defending usage tracking, arguing it could work if paired with understanding how tokens get used. But that misses the fundamental problem: incentives drive behavior. Rank people by token consumption, and they'll optimize for token consumption.

Model Routing Exposes the Lie

The technical solution to tokenmaxxing already exists: model routing. Direct simple tasks to smaller, cheaper models. Reserve flagship reasoning models for complex analysis that actually benefits from their capabilities.

Classification, extraction, and FAQ responses work fine on lightweight models. Planning, coding, and synthesis justify the premium models. This approach slashes costs while improving outcomes.

But model routing destroys the tokenmaxxing game. It optimizes for efficiency rather than volume. Smart developers using routing will show lower token consumption while delivering better results. They'll rank poorly on Meta's leaderboard while their tokenmaxxing colleagues burn through expensive flagship models on trivial tasks.

This is why Salesforce created Agentic Work Units (AWUs) to measure completed tasks and business outcomes instead of token volume. Singapore Airlines uses AWUs to track customer service resolution times. Williams Sonoma measures product recommendation performance.

Actual results. Novel concept.

Tokenmaxxing represents everything wrong with modern performance measurement: confusing activity with achievement. The most productive developers will often be the ones using fewer tokens, not more. They'll craft precise prompts, choose appropriate models, and avoid unnecessary iterations.

Meanwhile, their colleagues will be gaming leaderboards with expensive theater, generating millions of tokens that accomplish nothing meaningful. Guess who gets the performance bonus?

The industry needs to kill tokenmaxxing before it kills productivity. Measure outcomes, not inputs. Reward efficiency, not volume. Stop building leaderboards that incentivize waste.

Or keep burning through 60 trillion tokens monthly while wondering why AI hasn't delivered the productivity revolution everyone promised.