Instagram head Adam Mosseri says AI token costs could soon rival engineer salaries, and Meta is learning the hard way that usage incentives without spending controls create exactly the behavior you pay for.
Adam Mosseri did not dress this up as a distant infrastructure problem. Speaking on Lenny's Podcast, in comments Business Insider and AOL picked up this month, the Instagram chief said that within a year or two, "the burn rate of a strong engineer might be the same as their salary or their cost of employment." That should get your attention. It came from the executive running one of Meta's largest product organizations, not from a vendor trying to sell you a dashboard.
Mosseri's point was blunt: tokens are now part of the budget conversation. Not vibes. Not experimentation. Tokens sit beside GPUs, storage, RAM and payroll, because the bill shows up whether the work was useful or not. Instagram, he said, had already "reined in" its AI costs by shutting down "the silly things that we were doing." Asked about internal token leaderboards, he was even clearer. "It's a terrible idea," he said.
He was right.
According to reporting from The Decoder, Meta had an internal system nicknamed Claudeonomics, a nod to Anthropic's Claude, one of the third-party coding tools used inside the company. The leaderboard ranked employees and teams by raw token consumption, with tiers reportedly called Session Immortal and Token Legend. MLQ News reported that Meta employees burned through 73.7 trillion tokens in roughly 30 days after a policy tied AI-driven work results to 2026 performance reviews, with bonuses available for top performers.
The incentive did what incentives do when they're built around the wrong number. It made consumption visible, so people consumed. Employees didn't necessarily create more useful software. Some just ran more prompts.
The leaderboard was the mistake
Look, this is not complicated management science. If you rank people by token usage, some of them will optimize for token usage. The same thing happens when sales teams are paid on calls instead of revenue, or when customer support teams are scored on ticket closures instead of solved problems. You get the metric you ask for. Meta just happened to run the experiment with expensive AI models attached.
Meta's response, according to the same reporting, is a centralized monitoring platform called AI Gateway, which tracks usage and spending across teams in real time and flags abnormal spikes. Formal token budgets and department-level allocations are expected by 2027. The company is also pushing engineers toward MetaCode, its own coding assistant, and away from some outside tools.
The Information has separately reported that Meta restricted engineers in its Applied AI division from using Anthropic's Claude Code and OpenAI's Codex without approval, citing concerns about model distillation. That is a different issue from token costs. Still, the two pressures point in the same direction: less uncontrolled external AI use, more internal visibility, and fewer blank checks handed to teams in the name of productivity.
Even Meta is counting tokens now
Here's what makes Mosseri's warning land. Meta isn't short on capital. Associated Press reported in May that the company raised its 2026 capital expenditure forecast to $125 billion to $145 billion, driven largely by AI data centers and infrastructure. The Verge separately reported that Meta planned to cut about 8,000 jobs in May, around 10% of its workforce, while closing thousands of open roles. Those two facts belong in the same frame.
Meta is willing to spend staggering sums on AI infrastructure. It is also telling employees that inference costs need limits. That's not a contradiction. It is the new accounting.
Building data centers, buying chips and training models are capital decisions made at the top of the company. Letting every engineer burn tokens through Claude, Codex or similar tools is an operating expense that can spread quietly through teams before anyone responsible for the budget has a clean view of it. If you run a smaller company, you don't get Meta's margin for error. You get the bill.
The detail worth sitting with isn't only the 73.7 trillion tokens. It is the mechanism. Meta did not discover that AI tools were useless. It discovered that unmetered AI tools, attached to a bad incentive and a status game, can turn useful software into a cost problem very quickly.
Mosseri's framing is likely to become normal well beyond Instagram. Tokens will be allocated like headcount, GPUs and cloud credits. Teams will ask for more, finance will ask what they got for it, and engineering leaders will need a better answer than adoption went up.
Frankly, that is where the AI productivity story gets real. The first phase was giving everyone access. The next phase is asking whether the work produced is worth the tokens burned to produce it. If Meta has to ask that question while spending up to $145 billion on AI infrastructure, you should assume your company has to ask it too.
Also read: Meta Is Now Rationing AI Tokens Like It Rations Headcount • Chai Discovery Triples Its Valuation to $3.8 Billion in Seven Months • Chamath Palihapitiya Returns to Running a Company With a $135 Million AI Coding Bet
