Meta's AI Coding Bill Is Getting So Big, Mosseri Is Talking Token Caps

Adam Mosseri has a new line item to worry about: how many tokens his engineers are burning through in a day. The Instagram chief told colleagues that Meta may need to impose limits on employees' AI token spending within the next year or two, a sign that the free-flowing generosity companies have shown toward AI coding tools is starting to hit budgetary reality.

Mosseri framed the issue in terms familiar to anyone who has run a large engineering org. He compared token budgets directly to compute capacity, headcount labeling budgets, and payroll: finite resources that eventually require guardrails once usage scales past what finance teams initially modeled. Right now, Meta has no such caps in place. Employees can, in theory, spend as much as they want on AI-assisted coding without hitting a wall.

That lack of a ceiling has consequences. Meta reportedly shut down an internal leaderboard that tracked how many tokens each employee was consuming, after the company's overall AI usage put it on pace to spend billions of dollars in 2026. Leaderboards like this are usually built to encourage adoption by making usage visible and a little competitive. Pulling one down suggests the numbers got uncomfortable enough that visibility itself became a liability, either by encouraging more spending or by drawing unwanted attention to how fast the costs were climbing.

Meta isn't alone in confronting this. Uber reportedly blew through its entire full-year AI coding budget by April, just four months into 2026, an overshoot that points to how quickly usage patterns can outpace forecasts once AI coding assistants become embedded in daily engineering workflows. Microsoft, meanwhile, took a different approach to controlling costs and complexity: the company reportedly discontinued internal licenses for Anthropic's Claude Code, pushing its engineers instead toward its own Copilot CLI tool. That move reads less as a pure cost-cutting measure and more as a consolidation play, keeping spend and tooling inside Microsoft's own ecosystem rather than paying for a competitor's product at scale.

Together, these data points describe an industry quietly recalibrating its relationship with AI coding tools. A year or two ago, the priority was maximizing adoption: get every engineer using Copilot, Cursor, Claude Code, or whatever assistant fit the workflow, and worry about the bill later. Now "later" appears to be arriving. When a company the size of Meta is shutting down usage dashboards and a company the size of Uber is running out of runway before summer, it suggests token consumption has moved from a rounding error to a real line on the P&L.

None of this means AI coding tools are becoming less central to how software gets built. If anything, the opposite is true: usage is high enough to strain budgets precisely because adoption has been so successful. What's changing is the governance layer around that usage. Expect more companies to follow Meta's likely path: dashboards, soft limits, and eventually hard caps, treated with the same seriousness as cloud compute budgets or approved headcount.