Tokenmaxxing
The Expensive New Way to Look Busy
Remember when Jensen Huang stood on stage at GTC and called data centers “token factories”? When he said the future belongs to whoever can generate the most tokens-per-second?
He wasn’t lying. He just didn’t mention the other side of the equation: everyone in this industry wants you to spend more on more expensive tokens.
That includes your employer.
The New Status Game
At OpenAI, engineers burn hundreds of billions of tokens per week. At Anthropic, monthly bills hit six figures. At Meta and Shopify, managers tie performance reviews to token count.
Welcome to tokenmaxxing: the anxious status game where tech workers compete on leaderboards to prove they’re productive by burning through billions of tokens.
Nobody’s measuring whether any of this produces better software.
What Changed
Until recently, hitting millions of tokens took effort. ChatGPT Q&A, essay revisions, coding help — you’d rack up thousands per session. Billions was impossible without days of typing.
Then agentic coding tools arrived. Claude Code, Codex, OpenClaw — systems that work unsupervised for hours. They spawn subagents, review entire codebases, write programs from single prompts, run 24/7 while you sleep. Each subtask generates thousands of tokens. Stack a few agents in parallel and suddenly you’re burning hundreds of millions per week.
This isn’t iteration. It’s a phase shift. And it turned token consumption from a byproduct of work into a signal of work — whether or not the work is any good.
The Leaderboard Trap
Some companies run internal leaderboards showing each employee’s token count. The logic: more A.I. use = more productivity. Top of the board gets rewarded. Holdouts get chastened.
But leaderboards don’t measure output quality. They measure activity. And once you start rewarding the metric instead of the outcome, you get Goodhart’s Law: people optimize for token count, not for shipping useful code.
The incentive structure is backwards. The top of the leaderboard isn’t the best engineers — it’s the ones who figured out how to game the count. Stack subscriptions. Exploit bugs. Run agents that produce garbage but rack up tokens.
Token Anxiety Is Real
Tech workers know the script: A.I. is coming for white-collar jobs, adapt or die. So they’re signaling hard. Dinner conversations used to start with “What are you building?” Now it’s “How many agents do you have running?”
Career survival instinct in an industry that decided the future belongs to whoever can command the biggest swarm of A.I. agents.
“Inside large tech companies, it’s becoming a career risk to not use A.I. at an accelerated pace, regardless of output quality.”
If your manager ties your performance review to token count, you’re not going to argue about whether the code is maintainable or even correct. You’re going to spin up more agents and chase that leaderboard position because not doing so might get you fired.
The Economics Are Insane
Anthropic doubled its revenue projections in two months, mostly from agentic coding tools. OpenAI’s Codex tripled weekly active users; token usage increased 5x. Google processes 1.3+ quadrillion tokens/month.
Companies are paying massive bills because they’re terrified of being “left behind.” But if nobody’s tying this spend to measurable outcomes, it’s just expensive signaling all the way up the chain:
Engineers signal productivity via token count
Managers signal leadership by rewarding high token use
Executives signal innovation by subsidizing unlimited agent swarms
A.I. companies signal growth by celebrating whale users
Nvidia signals dominance by selling chips to power the whole circus
And at the end of this chain? Nobody knows if the output is any good.
Follow the Incentives
Jensen wants you to run more tokens because every token needs compute, and every compute cycle runs on Nvidia chips.
OpenAI wants you to run more tokens because that’s how they bill.
Your employer wants you to run more tokens because token count is easier to measure than actual productivity.
The entire industry is aligned on one thing: make you spend more.
Not “make you ship better software.” Not “make you solve harder problems.” Just: use more tokens.
This is the AI equivalent of selling you a gym membership and then measuring success by how many hours you spend in the building, not whether you’re actually getting stronger.
The Problem Isn’t Agentic Tools
Running background automation (heartbeat checks, cron jobs, monitoring) isn’t tokenmaxxing. That’s legitimate infrastructure. Offloading repetitive tasks to an agent that can handle them unsupervised is rational.
The problem is using them performatively because your company decided token count = productivity and you’re afraid of getting outcompeted by someone who doesn’t care if their 10 parallel agents are producing garbage.
What Happens Next
Scenario 1: The bubble pops. Companies realize they’re paying millions for token theater, start demanding ROI metrics tied to actual outcomes (revenue, shipped features, bug reduction). A.I. companies adjust pricing. Brutal correction, but healthier long-term.
Scenario 2: The anxiety wins. Token count becomes a permanent proxy for productivity because nobody wants to admit we don’t know how to measure knowledge work anymore. The industry locks itself into an expensive, wasteful equilibrium where appearing productive matters more than being productive.
I’m betting on a mix. The companies that figure out how to measure real outcomes early will win. The ones that don’t will burn cash until investors force the correction.
The Real Lesson
Activity ≠ productivity. Never has, never will.
You can rack up billions of tokens and ship nothing useful. You can also ship transformative work with a fraction of that if you know what you’re building and why.
Agentic coding tools are legitimately powerful. But using them to compete on a leaderboard? That’s not engineering. That’s status performance wrapped in automation cosplay.
The question every company should be asking isn’t “How many tokens are our engineers using?” It’s “What are we shipping, and is it better than before?”
Until that shift happens, tokenmaxxing is just another way to look busy while the real work gets ignored.
Coda
If you’re a tokenmaxxer: no judgment. You’re responding rationally to broken incentives. But when the leaderboard disappears (and it will), make sure you’ve got something to show beyond a high score.
If you’re a manager tying performance reviews to token count: stop. You’re optimizing for theater, not outcomes. Measure what ships. Measure what works. Measure what users care about.
And if you’re Jensen Huang: congratulations. You built the perfect incentive loop. Every token burned is another chip sold. You win either way.
The rest of us? We’re paying for it.