Meta’s AI Token War: Employees Curse Execs, Limits Imposed, Zuckerberg Admits Mistakes

In early 2024 Meta ran an internal competition called tokenmaxxing , where engineers deliberately consumed large numbers of AI tokens to prove themselves heavy AI users. The practice was celebrated and even displayed on a leaderboard that ranked the top 250 token spenders.

Within months the cost of internal AI usage began to climb dramatically, with internal estimates projecting that AI token consumption alone would cost tens of billions of dollars by 2026. In response, Meta shifted from encouraging tokenmaxxing to a Tokenminimizing policy that caps employee token usage.

The company built an internal platform, described as an “AI Gateway” dashboard, to monitor real‑time AI usage, set budgets, and enforce spending limits for roughly 6,000 employees. The memo also mentions an upcoming automatic‑alert system that will flag abnormal spending spikes.

To curb external AI costs, Meta is urging engineers to replace third‑party tools such as Anthropic’s Claude with its own internal coding assistant, MetaCode (formerly Devmate), while still permitting limited use of new external models.

“We see AI usage growing exponentially; at the current trajectory internal AI spend alone will reach tens of billions of dollars in 2026,” the memo states. “Employees and teams lack visibility and control over how much they spend on AI.”

Meta’s CTO Andrew Bosworth reiterated that token consumption should only occur when AI tools demonstrably boost productivity, warning that “using AI for the sake of using AI” does not equate to impact.

Employee dissatisfaction surfaced quickly. The Applied AI Engineering (AAI) team, about 6,500 engineers and product managers, faced criticism for being forced to generate training data for AI models rather than traditional software work. An internal live demo was disrupted when a participant shouted an expletive at a senior AI executive, highlighting growing morale issues.

More than 1,600 staff signed a petition opposing a project that would monitor clicks and keystrokes for AI training data, underscoring concerns about privacy and the dual role of employees as both AI users and data producers.

Organizational strain is evident: the AAI team originally employed a very flat hierarchy, sometimes with one manager overseeing up to 50 engineers, which amplified feelings of uncertainty and loss of control.

CEO Mark Zuckerberg later acknowledged the missteps, stating that Meta “made mistakes” in its AI workforce transformation and expects further errors as the transition continues, while promising stability in future reorganizations and no company‑wide layoffs.

The episode illustrates that once AI moves from a strategic slogan to everyday workflow, the primary challenges shift from encouraging adoption to managing cost, governance, and employee roles in a sustainable manner.