Meta Shut Down Its 60-Trillion-Token Leaderboard: Don't Become Fuel in the AI Frenzy

A 60-Trillion-Token Farce

In April 2026, Silicon Valley giant Meta played out a jaw-dropping management experiment inside its own walls. To encourage employees to embrace AI, executives launched an internal leaderboard called "Cloud Dynamics" that tracked each employee's monthly AI compute consumption (measured in tokens), rewarding top performers with flashy Platinum and Emerald badges.

The result? In just 30 days, Meta's 85,000 employees burned through a staggering 60 trillion tokens chasing leaderboard glory. The number-one ranked employee was particularly egregious — running 2,811 data tasks in a single month, which at prevailing compute costs translated to roughly $1.4 million of the company's money spent by one person alone.

Sounds like a group of million-dollar-salary elites furiously creating value? According to analysis by Bilibili creator "金融老兵" (Finance Veteran), it was in fact "a disaster of garbage generation." To inflate their rankings, engineers wrote scripts that fed AI endless loops of trivial questions already answered in documentation, or deliberately generated mountains of throwaway code that would never ship. Some even let machines blindly run core business logic for the sake of speed, ultimately causing system crashes. Within days, as compute costs spiked and the junk data backfired, the leaderboard was taken down in an emergency, and leadership was forced to broadly restrict employees' unchecked use of AI tools.

When Compute Becomes a KPI: Are You Becoming a "Digital Laborer"?

What this farce exposed is the deep-seated anxiety inside every modern professional. That anxiety has spawned a disease: performative diligence. When compute consumption becomes a performance metric, these Silicon Valley elites didn't become masters of AI — they became "digital laborers" frantically feeding data into a system.

What's alarming is that this isn't an isolated incident at a big tech firm — it's a spreading epidemic of "compute-driven rat racing." Several authoritative data points paint a clear picture:

• Forbes data: Of all enterprise AI initiatives launched in the U.S., a staggering 95% of pilot projects fail to generate meaningful business returns.
• S&P Global survey: 42% of companies have already shut down at least one AI project.
• Upwork survey: Of companies that adopted AI, 77% of employees report their workload has not decreased — it's actually increased.

Underlying all of this is a fatal logic: the Productivity Paradox.

The Productivity Paradox: More AI Tools, More Burden

Imagine buying a robotic vacuum, a floor scrubber, and a window-cleaning robot to save yourself effort — only to find you're spending more time refilling water tanks, emptying dirty water, and troubleshooting malfunctions than you ever did with a simple broom. The essence of the Productivity Paradox is this: adding more tools does not equal more real output. More often than not, it only adds enormous, invisible maintenance costs.

The same logic applies when you use AI to expand a single sentence into a 5,000-word flowery email to a colleague. You're not boosting efficiency — you're creating "toxic inventory" in your workplace. You think you're commanding AI; in reality, you're working for the algorithm, becoming a "human trigger" that exists solely to keep the AI system running. What you've paid a premium for is a flashy but hollow shell — a modern retelling of the ancient Chinese parable of buying the box and returning the pearl.

A note on the Productivity Paradox: The term was first coined by economist Robert Solow in 1987, who famously observed that "you can see the computer age everywhere except in the productivity statistics" — hence it's also called the Solow Paradox. Its core logic: there is a structural mismatch between technology investment and actual output. The learning costs, coordination costs, and maintenance costs introduced by new tools often offset — or even exceed — their efficiency gains in the short term. Historically, both electrification and the mass adoption of personal computers went through similar "output lag" periods, often taking 10–20 years to show up in macroeconomic productivity data. This phenomenon is especially pronounced in the AI wave: companies don't just pay for model API calls — they must also invest heavily in correcting AI errors, managing compliance risks, and absorbing the cognitive load employees bear when switching tools. This explains why 77% of workers in the Upwork survey report heavier workloads: they're effectively doing two jobs — their original work, and playing "AI babysitter."

Lines of Code Are Hollow; Profit Is Real

The video's author draws on a personal experience from twenty years ago to provide a sharp historical parallel for today's AI frenzy. Back then, he received a software outsourcing business owner at a Beijing branch office. The company had annual revenues of hundreds of millions of yuan, its books looked great, and the owner bragged that his team of 200-plus programmers churned out millions of lines of code every month.

But the author didn't look at lines of code. He zeroed in on one core metric: the payment collection cycle. He found the company's receivables cycle had stretched from three months to nine, with downstream debts piling up and cash flow on the verge of collapse. He rejected the multi-million loan application on the spot. Within a year, that company's funding chain snapped, and it went into liquidation.

Lines of code are hollow. Only profit that can truly be converted into real money matters. No matter how much code you write, if it can't be sold, it's a ticking time bomb sitting in a warehouse.

This judgment remains razor-sharp today: Token consumption, instruction counts, and word generation volumes could all be the new "lines of code" — vanity metrics that appear to signal prosperity while masking emptiness.

The video likens today's AI wave to a global "weather system": a massive expansionary force pushing everyone to buy compute and generate data at all costs. But commercial logic is ruthless — when the false boom hits a wall, contraction inevitably follows. Meta's emergency shutdown of the leaderboard and the 95% AI project failure rate are exactly this contraction force asserting itself.

A note on Tokens: A token is the basic unit of text processing for large language models. Roughly speaking, one English word equals about 1–2 tokens, and one Chinese character also equals about 1–2 tokens. Every time a model generates or receives text, it consumes tokens, and cloud APIs charge by token volume. Just how large is 60 trillion tokens? At commercial pricing of roughly $1–5 per million tokens for mainstream models, 60 trillion tokens represents a raw compute cost somewhere between $60 million and $300 million. Even with lower marginal costs from in-house infrastructure, the GPU compute-hours, electricity, and cooling that number represents are still enormous. This is precisely why token consumption is so vulnerable to gaming as a performance metric — a script can generate astronomically high token counts with minimal code complexity and zero business value generated. This is the fundamental design flaw of the leaderboard mechanism.

Three Practical Strategies for Not Becoming "Fuel"

In an era when machines can generate perfect answers at near-zero cost, how does a living, breathing person build a survival architecture that can't be replaced? Three core strategies are worth serious reflection for every professional.

Step One: Reclaim the Right to Ask the Questions

Don't just follow the options the machine hands you. In the workplace of the future, people who can't ask their own questions will only ever be assistants to machines. AI can draft a flawless equity agreement for you, but it can never judge a business partner's true character, look them in the eye, and assess whether they'll betray you under extreme pressure. Your job is to find the most critical pressure point within chaos and uncertainty — not to passively tick boxes on a machine-generated multiple-choice test.

Step Two: Deliberately Preserve "Clumsy Friction"

In an age obsessed with speed and frictionlessness, you need to consciously slow down. Which is worth more: a perfect client relationship email AI generates in three seconds, or the phone call you make late at night when a client faces a crisis — stumbling, exhausted, but genuinely caring? The answer is obviously the latter.

Real competitive moats aren't built through endless cost-cutting — they're built on scarce, hard-won trust. In a world flooded with artificial perfection, the emotional resonance and authentic interaction you're willing to invest time in are trust assets that no machine can ever replicate.

Step Three: Be the Person Who Makes the Call and Takes the Responsibility

When people lean heavily on AI, it's often because they're afraid to make decisions — if something goes wrong, they can blame the algorithm. But that means outsourcing control over your own risk. In the business situations where everything is on the line, the ability to choose to believe in a person or a cause even when all the data points to danger — that kind of courage and conviction is something no algorithm can simulate. In a contraction, what you compete on is your foundation — and that foundation is your accountability as a human being.

Closing: True Mastery Looks Simple

The Tao Te Ching teaches: "Great skill appears clumsy." In this grand human-machine competition, the truly wise have never been the fastest sprinters or the most prolific data generators. AI is always just the whetstone that sharpens your sword — you are the swordsman who decides where the blade falls.

Authenticity with all its imperfections, irreplaceable lived experience, and the "stubborn foolishness" of holding on to a belief against all odds — these are the most expensive assets in the world. Learning to use AI well is important, but more important than knowing how to use it is this: don't mistake your tools for your moat, and don't let yourself become fuel in the compute frenzy.