Meta Built a Leaderboard for Burning AI Tokens. Then Had to Shut It Down.

METAMAXXING

6:36

The stat making the rounds is real: Meta employees consumed 73.7 trillion AI tokens in about 30 days, which works out to roughly $221 million a month and $2.65 billion a year at list price. That's not exaggerated. It's in an internal memo, first reported by The Information, sent to roughly 6,000 employees warning that internal AI usage costs are approaching billions of dollars in 2026.

But the number by itself misses the actual story, and the actual story is better than the number.

The Leaderboard Nobody Meant to Build

Earlier this year, an employee at Meta built an internal dashboard called "Claudeconomics" — a nod to Anthropic's Claude, one of the third-party AI tools widely used inside the company — that ranked the top 250 power users by token consumption. It was meant to encourage adoption. What it actually did was turn AI usage into a competitive sport.

Employees started chasing titles like "Token Legend" and "Cache Wizard." According to SemiAnalysis, which talked to more than 50 enterprises about this exact trend, some employees had agents run for hours doing research tasks with no real output goal, purely to climb the rankings. The dashboard tracked over 60 trillion tokens burned in 30 days before The Information reported on it, and Meta shut the leaderboard down two days later.

CTO Andrew Bosworth's follow-up memo did not mince words: "All motion is not progress and token usage alone is not a measure of impact of any kind." Meta is now building a centralized "AI Gateway" dashboard to track spend in real time, with formal token budgets rolling out company-wide in 2027, and is steering employees toward its own MetaCode assistant instead of paying third-party API costs for tools like Claude.

Meta Isn't the Only One

Uber ran into the same wall from a different direction. The company burned through its entire 2026 AI coding budget in four months and responded by capping employee spending at $1,500 per month per tool. Nearly 95% of Uber's engineers use AI tools monthly, and close to 70% of committed code is now AI-generated — but COO Andrew Macdonald has openly said the link between token spending and measurable output "is not there yet."

A KPMG survey found only 26% of companies have comprehensive visibility into their own AI costs. That's the real headline hiding underneath the viral number: this isn't a Meta problem, it's an industry-wide blind spot, and Goldman Sachs is projecting a 24x increase in enterprise token consumption by 2030, reaching 120 quadrillion tokens a month across the industry. If companies can't see their spend today, at a fraction of that volume, the visibility gap only gets more expensive from here.

The Part That Actually Matters

Here's where SemiAnalysis's reporting adds the nuance the meme version leaves out entirely: after talking to enterprises directly, they found the panic narrative around Meta and Uber is overstated, and it stems from poor incentive design and bad allocation structures at those specific companies — not from AI tools inherently being wasteful. Most organizations they spoke with never had a gamified leaderboard problem in the first place, because they never built one.

What's actually happening across the enterprise landscape is a shift from unmetered access to formal token budgets, with numbers ranging anywhere from $250 a month to tens of thousands, depending on the role. Companies are downgrading default models for routine tasks, turning off premium tiers by default, and treating token spend as a line item that needs the same discipline as cloud compute spend got a decade ago.

That comparison is the one worth sitting with. When cloud computing went mainstream, companies spent years figuring out that unmetered access to infinitely scalable infrastructure needs guardrails, which is exactly why FinOps became its own discipline — a dedicated function just for managing and forecasting cloud spend. Token budgeting is heading down the identical path, just compressed into months instead of years, because the entities doing the spending this time also built the tools and know exactly how fast the bill can move.

Why This Is the More Interesting Story

Everyone's capex conversation about AI right now is about chips, power, and data centers — the hardware layer. Meta alone is planning to spend up to $135 billion on AI infrastructure through 2026 and has committed $600 billion to data center buildouts through 2028. That's the story that gets the headlines.

But there's a second, quieter cost layer sitting on top of all that infrastructure: what happens once you actually hand every employee a frontier model and tell them to go use it. That's not a capex line, it's an opex line, and it turns out humans are extremely good at turning any measurable metric into a leaderboard, whether or not that metric means anything. Meta didn't get burned by AI being expensive. It got burned by forgetting that if you rank people by a number, they will maximize that number, and the number was never the point.

The actual lesson isn't "AI costs are spiraling out of control." It's that the operational discipline required to run AI tools at scale inside an organization — governance, incentive design, usage visibility — is turning into its own competitive advantage, separate from whichever model or chip you're using. The companies that figure out token budgeting now, quietly and without a viral headline, are the ones who'll actually see a return on the trillion-dollar buildout everyone else is still arguing about.

---

Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.