AI token spending is spiraling out of control at major tech companies, and executives are now openly discussing the need for per-engineer spending caps to prevent budgets from reaching unsustainable levels. Meta's head of Instagram, Adam Mosseri, recently warned that the cost of processing AI prompts and responses could soon reach parity with what companies pay their engineers in salary, forcing a fundamental reckoning with how organizations budget for artificial intelligence.

Why Are AI Token Costs Becoming a Crisis?

Token costs refer to the expense of running AI models, measured by the computational resources needed to process user inputs and generate outputs. What started as an experimental expense has become a serious budget concern. Meta shut down an internal AI token spend leaderboard after discovering the company was on track to spend billions of dollars on AI tokens in 2026 alone. The company is not alone in this struggle. Uber exhausted its entire 2026 AI coding budget by April, and Microsoft canceled Claude Code licenses from Anthropic, consolidating its engineers around its own internal Copilot CLI tool instead.

The scale of the problem is staggering. Mosseri explained that at some companies, the annual burn rate of AI tokens for a strong engineer could soon match or exceed their total salary and employment costs, creating an economic paradox where the tool costs as much as the person using it.

What Does Mosseri Think Companies Should Do?

"I think that you can imagine, at least in a year or two, that the burn rate of a strong engineer might be the same as their salary, or their cost of employment. And in that world, you're going to probably need to put in some caps," said Adam Mosseri.

Adam Mosseri, Head of Instagram at Meta

Mosseri's proposal treats AI token budgets like any other constrained resource that companies must manage strategically. He explained that organizations already make difficult allocation decisions across multiple dimensions, and token spending should follow the same logic. Just as companies decide how to deploy limited GPUs, CPUs, storage, and RAM across teams, they will need to decide how to allocate AI token budgets. The key is that per-engineer caps should be proportional to each person's demonstrated ability to use the budget in ways that generate positive return on investment.

How Should Companies Manage AI Token Spending?

• Set Per-Engineer Caps: Establish spending limits proportional to each engineer's track record of generating return on investment, similar to how companies allocate payroll and operating budgets across teams.
• Eliminate Wasteful Experiments: Meta has already begun shutting down activities like token spend leaderboards that incentivize excessive AI usage without creating measurable value.
• Treat Tokens Like Other Resources: Apply the same resource allocation framework used for computing infrastructure and payroll to AI token budgets, recognizing that all three are finite and require strategic deployment.
• Plan for Future Price Competition: Mosseri expects token costs to decline as AI model makers enter a pricing war to attract users, but companies should not rely on this happening quickly.

Currently, Meta does not enforce token caps for any employee, but Mosseri believes such limits could become healthy and necessary within the next year or two. The key insight is that token budgets will need to be managed like payroll, computing infrastructure, and operating expenses, with caps tied to demonstrated ability to use the budget in ways that generate positive returns.

The broader implication is significant: companies are signaling that the era of unlimited AI experimentation may be coming to an end. As token costs continue to rise and organizations face billions in unexpected AI spending, the industry is shifting from asking "Can we afford to use AI?" to "How do we allocate limited AI resources strategically?" This shift mirrors earlier transitions in cloud computing and GPU allocation, where companies learned to treat computational resources as finite and managed them accordingly. For developers and engineering teams, this means the days of running unlimited AI experiments without oversight are likely numbered, and organizations will soon expect the same cost discipline for AI tools as they do for other enterprise infrastructure.