Meta's Mosseri: AI token budgets likely need per-engineer caps within 1–2 years

Instagram head Adam Mossori recently told Lenny's Podcast that Meta will likely need to implement per-engineer caps on AI token spending within a year or two. His rationale is straightforward: the computational cost of running AI queries for a strong engineer could soon equal or exceed that engineer's annual salary and employment costs. At that point, Mossori argued, companies will have no choice but to impose spending limits.

The urgency of this concern stems from Meta's own recent experience. The company shut down an internal leaderboard that ranked employees by their AI token spend after realizing the costs could reach billions of dollars in 2026. This is not an isolated problem. Uber exhausted its entire 2026 AI coding budget by April of this year, forcing a rapid reassessment of spending priorities. Microsoft took a more direct step, canceling its Claude Code licenses and consolidating engineers around its own Copilot CLI tool to consolidate costs.

Mossori explained that managing token budgets should follow the same principles as managing other constrained resources within the company—GPU and CPU capacity, storage, RAM, labeling budgets, and headcount. Each team receives an allocation based on company priorities and expected return on investment. Token budgets, he said, should be capped on a per-engineer basis in proportion to the company's confidence in that engineer's ability to spend the budget in an "ROI-positive" way. Currently, Meta has no such caps in place for any employee.

Looking further ahead, Mossori expressed optimism that token costs will decline as AI model makers compete for users by lowering prices. In the near term, however, Meta has started to reduce token burn by eliminating what Mossori called "silly things"—such as the internal leaderboard itself—that consumed tokens without producing measurable business value.

AI token spending has emerged as a material operational challenge for major technology companies in 2025–2026. Meta's shutdown of its internal token spend leaderboard—a metric that gamified resource consumption—signals that unconstrained AI experimentation can quickly become expensive. Mossori's framing of token budgets alongside payroll, GPU allocation, and operating expenses reflects a broader industry shift: AI is no longer a nascent experimental domain but a cost line item requiring disciplined resource management, much like traditional infrastructure and labor.

The timing Mossori suggests—one to two years before caps become necessary—appears grounded in the observation that token costs are rising faster than engineer salaries. However, he also expects competitive pricing pressure among AI model providers to eventually bring costs down, which could delay or reduce the need for hard caps. In the interim, Meta has already taken steps by eliminating internal token-burning activities, a pragmatic approach to cost control without formal caps.

No comments yet. Be the first to share your thoughts!