Meta Re-bets on AI: Zuckerberg uses Claude to write code, while employees spark a Token burn battle to hit KPIs

Meta is doubling down on AI, and Zuckerberg is personally using Claude Code to write programs. To hit KPIs, an internal token consumption free-for-all has emerged, bringing chaos and wasting resources. Learning the lesson from the metaverse’s $80 billion loss, Meta is actively acquiring startups, hoping to turn the technology into tangible value.

Meta doubles down on AI: Zuckerberg writes code with Claude Code

Meta, a tech giant that owns social media platforms such as Facebook, Instagram, and Threads, is recently putting the company’s resources fully into the generative AI space.

Meta CEO Mark Zuckerberg has started personally using an AI code-writing tool, Claude Code, to write code, breaking his years-long record of not being involved in direct development.

Overseas media have also reported that internally, Meta has recently sparked a token consumption competition: many engineers consume large amounts of tokens to boost their individual performance indicator (KPI) figures.

AI coding is all the rage; founders return to the development front

In March 2026, Zuckerberg submitted 3 code diffs to Meta’s single repository; this is his first tangible code contribution in 20 years.

The terminal coding assistant he used is Claude Code CLI, developed by Anthropic, and one of the submissions received approvals from more than 200 engineers.

His actions reflect how AI coding tools are attracting corporate founders back into system development. Y Combinator CEO Garry Tan has also returned to coding after 15 years and open-sourced a system that combines Claude Code.

According to internal documents leaked from Meta in March 2026, the company set aggressive goals, planning that by the middle of 2026, 65% of engineers will use AI to write 75% or more of their code.

Image source: flickr, photographed by Niall KennedyMeta founder Zuckerberg speaking at the Facebook F8 developer conference in September 2011

Meta’s internal token consumption competition; KPIs become a performance

To promote the adoption of generative AI applications, Meta has internally developed a phenomenon that links token usage to productivity. Tokens are the smallest unit processed by large language models for text; in Chinese they are often called “symbol units” or “tokens.”

《The Information》reports that an internal leaderboard called Claudeonomics has appeared at Meta, tracking AI token consumption across more than 85,000 employees. The data shows that employees consumed as many as 60 trillion tokens in just 30 days, and the top user’s average consumption reached 281 billion tokens.

The leaderboard sets up titles such as Token Legend, encouraging employees to integrate AI tools into their day-to-day work.

《Forbes》reports that Meta CTO Andrew Bosworth has mentioned that a top engineer consumes a token amount equivalent to his annual salary; NVIDIA CEO Jensen Huang has also said that he would be concerned if an engineer earning $500,000 could not consume $250,000 worth of tokens.

However, this KPI system that drives token consumption also comes with downsides. To boost their performance figures, some Meta employees let AI agent programs run idle for hours, causing a waste of computing resources.

Moreover, treating employees’ token consumption directly as a productivity metric makes consumption behavior turn into a performance, creating challenges for performance evaluations that lack support from concrete business outcomes.

Learning from metaverse failure experience, Meta’s next challenge in the AI race

Before making heavy investments in AI, Meta’s bet on the metaverse ended in failure. The company had previously invested roughly $8 billion to build the virtual world Horizon Worlds and VR/MR devices, and even changed the company name to “Meta,” but it still failed to reach the user scale expected by the market.

Lily Liu, president of the Solana Foundation, also expressed a pessimistic view in the comment section on a social platform when discussing the development of blockchain gaming and the metaverse, criticizing the virtual economy model in the past for lacking substantive content support.

Image source: Meta Meta’s metaverse platform Horizon Worlds; in the initial version, showing Zuckerberg’s virtual character

Now that Meta has shifted its focus to AI, it is actively laying out its market strategy. Besides launching its own large language model LLaMA, it is also gradually advancing an AI model plan called “Avocado.”

A recent report from 《Axios》also revealed that Meta acquired the agent community touted as the “AI version of Reddit,” Moltbook; Moltbook founders Matt Schlicht and Ben Parr will join Meta’s team.

• Related report: 《The Lobster community was bought! It’s rumored that Meta acquired Moltbook, and the founder successfully transitioned from media to AI》

Outside observers are still watching to see whether Meta can avoid repeating the metaverse’s path of over-investment while lacking real applications—turning the current internal token consumption frenzy and acquisition deals targeting startups like Moltbook into real products with commercial value, thereby securing a foothold in the fiercely competitive generative AI market.