Instagram's Mosseri: AI token budgets are becoming the next payroll line item

What happened

Instagram head Adam Mosseri said on a recent episode of Lenny's Podcast that companies will eventually need to treat AI token spending the way they treat payroll or other operating expenses — and that engineers could soon face individual caps on how much they spend using AI tools.

Mosseri said Instagram has already begun reining in its own AI costs. "We've managed to get the costs reined in a little bit by shutting down the silly things that we were doing," he said, adding, "It is not that hard to build a token incinerator." He did not specify which projects were cut.

When host Lenny Rachitsky asked about token-consumption leaderboards — the kind that fueled the early "tokenmaxxing" trend — Mosseri called them "a terrible idea." Meta was reportedly one of the largest companies to implement such a leaderboard internally. According to a LinkedIn post by Weave CEO Adam Cohen, Meta ran an internal competition called "Claudeonomics" in which 85,000 employees competed to be the top token consumer, with the leading user burning 281 billion tokens in a single month.

Meta has since reversed course, limiting usage and pushing employees onto its internal tool MetaCode — what Cohen described as "a full 180 in a matter of months."

Why it matters

Mosseri's comments signal a turning point in how large enterprises think about AI costs. The initial phase of AI adoption was characterized by encouragement — leaderboards, gamification, and open-ended access. Now, as bills climb and agentic AI drives more complex (and more expensive) workloads, companies are confronting the same question they face with any resource: how much is too much, and who should be allowed to spend it?

Uber COO Andrew Macdonald also recently raised concerns about rising AI bills, suggesting the issue extends well beyond Meta.

The debate is not just about cost-cutting. Cohen argues that Meta's approach — capping and consolidating onto a single internal tool — misses the point. The real problem, he says, is that companies lack systems to measure which engineers are getting valuable output from their tokens and which are burning them on prompts that never reach production. His recommendation: track output per dollar at the individual level, identify who is getting real value, and coach those who aren't — the same approach you'd take with any employee.

This tension between restriction and optimization will likely define the next phase of enterprise AI adoption. Companies that simply cap spending risk throttling productive engineers alongside wasteful ones. Those that build better measurement systems may find that the problem was never spending itself, but the absence of accountability.

What to watch

• Whether other major tech companies follow Meta's lead in moving from gamified token consumption to per-engineer budgets.
• How Meta's internal MetaCode tool performs compared to external AI coding assistants — and whether engineers resist the consolidation.
• The emergence of tooling that tracks AI output per dollar at the individual or team level, which Cohen and others argue is the missing layer.
• Whether token pricing models from providers like OpenAI and Anthropic shift in response to enterprise demand for more granular cost controls.