Instagram head Adam Mosseri says the cost of running AI tools per engineer could soon rival their salary, and Meta is already treating tokens as a budgeted resource rather than a free-for-all.
Speaking on Lenny's Podcast in comments picked up this month by AOL and other outlets, Mosseri 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's not a hypothetical thrown out for effect. It's a projection from the executive who runs one of Meta's largest product organizations, and it lands at a company that just spent months watching its own AI spending spiral.
Mosseri described tokens as having become part of the same budgeting conversation as GPUs, storage, RAM and payroll. Instagram, he said, had already "reined in" its AI costs by shutting down "the silly things that we were doing." Asked about token leaderboards, the internal rankings some teams used to track who was using AI tools the most, Mosseri didn't hedge. "It's a terrible idea," he said.
He had reason to say so. According to reporting from The Decoder, Meta built an internal system nicknamed "Claudeonomics," a nod to Anthropic's Claude, one of the third-party coding tools widely used inside the company. The leaderboard ranked employees and teams by raw token consumption, with tiers reportedly named things like "Session Immortal" and "Token Legend." MLQ News reported that Meta employees burned through 73.7 trillion tokens in roughly 30 days, a spike traceable to a November policy change that made demonstrating "AI-driven work results" a factor in 2026 performance reviews, with bonuses attached for top performers.
The incentive did what incentives do when they're built around the wrong metric. Employees didn't necessarily do more useful work. Some just ran AI tools to climb the rankings.
Meta's fix, per the same reporting, is a centralized monitoring platform called AI Gateway, which tracks usage and spending across teams in real time and flags abnormal consumption spikes. Formal token budgets and department-level allocations are expected by 2027. The company is also steering engineers toward MetaCode, its own coding assistant, and away from external 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, though the token-cost crackdown and the distillation restrictions are distinct issues that happen to be pushing in the same direction: less external AI, more internal control.
A budget line at a company spending $145 billion on AI #
Here's what makes this land differently than a routine cost-cutting memo. Meta isn't short on capital. The company raised its 2026 capital expenditure guidance to a range of $125 billion to $145 billion, most of it funding data centers, custom silicon and training runs for Llama and Meta's superintelligence lab. In April, Meta also cut roughly 8,000 jobs, about 10% of its workforce, with Zuckerberg telling employees at a town hall that the layoffs were a direct consequence of the company's AI infrastructure budget, according to reporting cited by Tom's Hardware.
So Meta will spend $145 billion building AI infrastructure while capping how many tokens an individual engineer can spend using it. That's not a contradiction so much as a signal. Building the compute and paying for inference are two different budget lines, and the second one apparently got out of hand fast enough that a leaderboard had to be shut down.
For founders and infra buyers watching inference economics, the detail worth sitting with isn't the dollar figure. It's the mechanism. Meta didn't hit a token ceiling because AI got less useful. It hit one because a badly designed incentive turned token spend into a status game, and nobody had visibility into the aggregate cost until it was already in the billions. Any company handing engineers unmetered access to Claude, Codex or similar tools without usage tracking is running the same experiment Meta just ran, just without the dashboard to catch it. Mosseri's framing, tokens as a line item to be allocated like GPUs or headcount, is likely to become standard practice well beyond Meta. If a company spending $145 billion on AI infrastructure still needs to cap what individual engineers can burn through, smaller companies with real budget constraints have even less room to treat inference as free.
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