Recently, I noticed an interesting phenomenon: Meta’s founder, Mark Zuckerberg, has actually started writing code himself. This isn’t just a routine code submission—it’s his first substantive code contribution in 20 years. It’s reported that he’s using the Claude Code CLI tool developed by Anthropic, and in one of his commits, it even received approval from more than 200 engineers.

So what does this reflect? AI coding tools are drawing company founders back into system development. Even Y Combinator CEO Sam Altman returned to coding after 15 years—this trend truly shouldn’t be underestimated. Meta is even more ambitious internally; according to documents leaked in March this year, they plan to have 65% of engineers use AI to write more than 75% of their code by mid-year.

But things start to get interesting here. To push forward generative AI applications, Meta has spawned a ranking list called Claudeonomics, tracking the token consumption of more than 85,000 employees. In just 30 days, employees consumed as many as 60 trillion tokens, and the average token consumption of the top users on the list reached 281 billion tokens. The company has also set up titles like “Token Legend” to encourage employees.

This sounds insane, but even more insane is that some employees, to boost their performance numbers, simply leave AI agent programs running idle for hours, causing waste of computing resources. Meta CTO Andrew Bosworth has mentioned that the tokens consumed by a top engineer are equivalent to their annual salary; he said this kind of comparison to illustrate the scale. Nvidia’s CEO Huang Renxun has also said that if an engineer earning $500,000 a year fails to consume tokens worth $250,000, he would feel worried.

Plainly put, a system that ties KPI targets to token consumption eventually turns into a show. Employees chase numbers for the sake of numbers, and performance reviews lose their support from any real business outcomes.

What’s somewhat worrying is that Meta previously tripped up in the metaverse. It once poured about $80 billion into building Horizon Worlds and VR/MR equipment, even changed the company’s name, and yet still didn’t reach the expected user scale. Now that it’s shifting to the AI track, whether it can avoid the same pitfall of overinvestment—turning the internal token-consumption frenzy and the acquisition of startups into products that truly have commercial value—is the key. Otherwise, no matter how high the internal metrics are, they’re just a numbers game.