Tokenmaxxing Was Silicon Valley’s Dumbest Trend, and I’m Living Proof You Don’t Need a Leaderboard to Use AI a Lot

I run ecommerce operations for my day job, and I lean on Claude for a real chunk of that work. My monthly bill runs around $1,500, which works out to roughly 100 million tokens a month. Compared to what Silicon Valley was doing to itself earlier this year, I’m not even a rounding error.

For a few months in 2026, some of the richest companies on earth paid employees to compete over who could burn through the most AI tokens. Not who shipped the best feature. Not who made customers happiest. Just raw, unfiltered token consumption, tracked, ranked, and rewarded like a leaderboard in a video game. They called it tokenmaxxing, and it was every bit as dumb as it sounds.

Meet “Claudeonomics,” the leaderboard nobody asked for

Somewhere inside Meta, an employee built an internal dashboard ranking roughly 85,000 coworkers by how many AI tokens they’d burned. They named it Claudeonomics, after Anthropic’s Claude, and handed out titles like “Token Legend” and “Session Immortal” to the top performers.

One employee burned through 281 billion tokens in a single month. That’s nearly 3,000 times my entire monthly usage, in 30 days, from one person. Across the whole leaderboard, Meta staff used 60.2 trillion tokens in 30 days, which at retail API pricing would run about $900 million. Even at a steep internal discount, that’s still real money, possibly north of $100 million, spent partly on work an insider later described as “disposable garbage.”

Mark Zuckerberg, who presumably approved a culture where this was a thing employees competed over, didn’t even crack the top 250.

It gets dumber. People built bots to lose on purpose

Here’s the part that’s actually funny. Once token count became the metric that mattered for performance reviews, employees did exactly what smart people do when you reward a number instead of an outcome: they gamed it.

At Amazon, staff reportedly spun up AI agents to run wholly meaningless tasks, just to keep their stats up. One venture capitalist described watching Meta engineers build bots that, in his words, “just run in a loop burning tokens as fast as they can.” Other workers fed entire codebases or stacks of redundant PDFs into the AI for questions that needed one sentence of context, purely because bigger inputs meant bigger token counts.

There’s even a term for the dread behind it: “token anxiety,” the fear that if your usage numbers look low, your manager assumes you’re falling behind on AI, regardless of whether your actual work is good. People weren’t tokenmaxxing to be productive. They were tokenmaxxing to not get noticed for being unproductive. That’s a strange thing to build a corporate incentive structure around, and an even stranger thing to compete in.

Then the invoices showed up

This is the part of the story that actually matters.

Uber’s COO said the company blew through its entire 2026 AI token budget in the first four months of the year, largely from heavy use of AI coding tools. His quote on it might be the most honest thing a tech executive has said about AI all year: autonomous agents were writing about 10% of Uber’s code, but he flat out couldn’t draw a line from that to anything users actually noticed. “That trade becomes harder to justify because it’s not free,” he said.

Salesforce’s CEO said his company is staring down a roughly $300 million bill from Anthropic this year, and admitted he wished there was a way to automatically route easy questions to cheaper models instead of paying premium rates for everything. Microsoft reportedly canceled Claude Code subscriptions in several divisions. And Meta, after the bad press hit, quietly took its own leaderboard down.

Nobody got fired over this, as far as anyone’s reported. They just all collectively arrived at “huh, maybe rewarding volume instead of outcome was a bad call,” which feels like something a first-year MBA student could have told them for free.

The actual lesson, from someone who pays his own AI bill

I bring up my own usage not to brag, 100 million tokens a month sounds impressive until you remember one Meta employee did nearly 3,000x that, solo, in the same window. But it’s a useful contrast. Nobody’s paying me to run that number up. If anything, the opposite, it comes straight out of operating costs I’d love to shrink. The only reason I keep spending it is that the output justifies it: work that ships, problems that get solved, time that gets saved. If a month went by where I burned through that much and couldn’t point to what it bought me, I’d cut back and figure out what went wrong.

That’s the whole test tokenmaxxing failed. Strip away the AI buzzwords and this is a story as old as business itself: a company measured the wrong thing, and the people inside it optimized for the wrong thing, exactly as fast as you’d expect smart, competitive people to do. The surprising part isn’t that it happened. It’s that it took a Wall Street Journal story and a few hundred million dollars in invoices for anyone at the top to notice.

Tokenmaxxing is dead now. Long live whatever the next version of it turns out to be.