Meta’s Adam Mosseri says AI “token budgets” may become capped per engineer as AI costs rise

Adam Mosseri, the head of Instagram, said Meta and other companies will likely have to manage AI usage with clearer spending controls, similar to how they manage payroll or other operating costs. In comments highlighted by Yahoo Finance, Mosseri suggested that “token budgets” could eventually be capped per engineer, a policy designed to limit how much generative AI input and output a developer can consume over a period of time.

The phrase “token budget” refers to the unit of work used by large language models. Tokens cover both the text or data you send into an AI system and the text the system returns, so higher usage generally means higher computational cost. Mosseri’s central point, as reported, is that these consumption levels are not just a technical concern, but a budgeting problem that grows more urgent as companies scale AI tools across teams.

Mosseri’s remarks come as organizations attempt to expand AI-assisted workflows in product development, support, and internal operations. As teams increase usage of chat interfaces, coding assistants, summarization tools, and other model-powered features, the cost of “just running the model” can become a meaningful part of ongoing expenses. A per-engineer cap would be one mechanism to prevent usage from expanding without guardrails.

While the comments point toward tighter controls, the report does not provide specifics on how such caps would be implemented at Meta, including whether limits would be enforced through internal tooling, how budgets would be set, or how frequently they might be refreshed. It also does not spell out whether the idea is already in use in some form or remains a forward-looking prediction.

For Meta, the broader stakes are straightforward: the company relies on AI across content-related products, ranking and recommendation systems, and user-facing experiences. Even when AI supports tasks that are not directly billed to customers, the compute required to run models and the engineering time to integrate them both contribute to cost structure and execution trade-offs.

Industry-wide, the idea of AI budgeting is gaining attention because generative AI workloads can be unpredictable. Different tasks, model choices, and prompt patterns can change token consumption dramatically. Establishing budget-like constraints is one way companies aim to balance experimentation and productivity with cost control, especially when AI is used by many employees rather than just a small research group.

Even so, there are open questions that the reported comments do not answer. The Yahoo Finance item does not disclose whether Mosseri meant a formal policy inside Meta, how any cap would vary by team or project, or what exceptions might exist for high-priority work. It also does not clarify whether the goal is to reduce total model usage, steer teams toward cheaper models, or simply make spending visible and accountable.

Looking ahead, investors and analysts may focus on whether Meta and other large technology companies begin to publish clearer indicates about AI operating costs, efficiency measures, and how internal AI tooling is governed. If per-team or per-engineer budgeting becomes common, it could also shape how companies roll out new AI features, prioritizing those with more predictable usage patterns and controllable cost profiles.