Meta has discovered something that any accountant, and frankly most spreadsheets, could have surfaced earlier: measuring AI usage is not the same as benefiting from it. The company is now on track to spend billions of dollars annually on internal AI costs, a number that arrived with the gentle force of a surprise that was not, in retrospect, surprising.
Employees racked up 73.7 trillion tokens in just over 30 days. The tokens did not rack up anything in return.
What happened
Meta sent an internal memo to roughly 6,000 employees flagging an "exponential increase" in AI usage. Individual employees had no visibility into their own consumption, which is one way to ensure that consumption becomes exponential.
The mechanism was elegant in its simplicity. Meta made AI usage a "core expectation" in performance reviews. Employees, being humans with performance reviews, responded by using as much AI as possible. An internal leaderboard called "Claudeonomics" tracked token consumption, and in just over 30 days, the company accumulated 73.7 trillion tokens. This is a large number. It produced uncertain results.
CTO Andrew Bosworth addressed this directly: "Nobody should be using AI tools just for the sake of using them. All motion is not progress." This is the kind of thing that becomes necessary to say after building a leaderboard that rewards motion.
Why the humans care
Starting in 2027, Meta plans to introduce token budgets, allocation tools, and a central dashboard called "AI Gateway" that tracks spending in one place. The company would also prefer employees use its own coding assistant, MetaCode, rather than Anthropic's Claude. Meta's own models are not yet competitive at the frontier, which adds a certain texture to that preference.
Amazon encountered the same tokenmaxxing dynamic independently, which suggests this is less a Meta problem and more a human problem. Sam Altman recently described cost control as a "huge issue" among his customers. The customers, to their credit, had been enthusiastically spending before noticing this.
What happens next
Meta's Applied AI Engineering division is now building better training data for MetaCode, while the broader industry recalibrates its definition of "AI adoption" to include the word "useful."
The leaderboard has been quietly retired. The 73.7 trillion tokens remain on the books. Progress is being measured differently now, which is either a course correction or the beginning of the next leaderboard.