Meta's head of Instagram says token budgets for engineers are becoming as real as payroll, and the confession exposes just how expensive AI coding tools have gotten inside the one company that can supposedly afford anything.
Adam Mosseri didn't sound like an executive bragging about AI adoption. On a recent episode of Lenny's Podcast, the Instagram chief admitted his team had let AI spending run wild, then had to claw it back. "It is not that hard to build a token incinerator," he said, describing tools his engineers built and then quietly shut down. Asked about leaderboards that rank employees by how many tokens they burn, his answer had no hedge in it. "It's a terrible idea."
He had reason to know. Meta built exactly that kind of leaderboard. Internally it was called Claudeonomics, a nod to Anthropic's Claude, one of the outside AI tools Meta engineers had been using heavily. The board ranked the top 250 token consumers with titles like "Token Legend," and according to reporting picked up by outlets including The Decoder and MLQ News, some employees reportedly left AI agents running idle for hours just to climb the rankings. Meta's chief technology officer, Andrew Bosworth, eventually wrote his own memo shutting the whole thing down. "All motion is not progress and token usage alone is not a measure of impact of any kind," he wrote.
The numbers behind the panic #
The numbers behind that memo explain the urgency. Meta employees consumed 73.7 trillion tokens in roughly 30 days, according to internal figures reported by MLQ News, and the company warned a group of about 6,000 employees that AI usage was rising exponentially with little visibility into which teams were actually driving the cost. Meta is now building a centralized dashboard called AI Gateway to track spending in real time, with a formal, team by team token budget framework expected by early 2027. In the meantime, the company is steering engineers away from Claude and toward its own coding assistant, MetaCode, which is cheaper to run because Meta isn't paying a vendor's markup on inference. That's the workaround, not a fix.
Mosseri's framing of the problem is the more interesting part. He described tokens as a line item that now sits next to GPUs, storage, RAM, and headcount when he decides how to allocate resources across Instagram. "I think that you can imagine, at least in a year or two coming, that the burn rate of a strong engineer might be the same as their salary or their cost of employment," he said. That's a striking thing to say out loud. It means Meta is starting to treat an engineer's AI usage as a second paycheck, one that shows up on a different ledger but costs the company just as much.
Meta is not alone #
Amazon told employees last month to stop using AI simply for the sake of using it, after engineers began running agents just to climb internal adoption leaderboards, the same failure mode that hit Meta. Uber burned through its entire 2026 AI coding budget in four months and responded by capping spending at $1,500 per employee per tool, per month. Walmart and Cisco have made similar moves. None of this reads like companies celebrating AI's return on investment. It reads like finance departments discovering, all at once, that inference isn't free just because the model is good.
There's a supply constraint sitting underneath the spending one, too. Google reportedly told Meta around March that it couldn't sell Meta all the Gemini capacity it wanted to buy, delaying some internal projects and forcing more of the rationing Mosseri is now describing in public. That's worth sitting with. Meta is planning to spend up to $135 billion on AI infrastructure this year. That's real money, even for Meta. It has also committed $600 billion to data centers through 2028, yet its own engineers are being told to watch their token spend like a grocery budget. The company building some of the largest data centers on earth still can't give every employee unlimited access to the models running inside them.
Frankly, that gap between capex and internal rationing is the real story here, more than any single leaderboard or podcast quote. If the company spending the most on AI infrastructure in the world still has to cap what its own engineers can burn through, the idea that inference costs will simply fall to zero as models improve looks less like a certainty and more like a bet. Meta itself isn't willing to make it yet. If you run an AI-native engineering org, you're watching the same math play out without Meta's balance sheet to cushion it.
Also read: 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 • StepFun Unveils the First Agentic AI Phone Ahead of Apple and OpenAI