Meta's head of Instagram says token budgets for engineers are becoming as real as payroll, and that confession tells you how expensive AI coding tools have become inside companies 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 said his team had already had to rein in AI costs by shutting down wasteful internal work. "It is not that hard to build a token incinerator," he said. 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 had one. According to The Information reporting cited by Business Insider, a Meta employee built an internal dashboard called Claudeonomics, a nod to Anthropic's Claude, that let staff compete for titles including "Token Legend." Some employees reportedly ran bots in loops to burn tokens and climb the rankings. That is exactly what happens when you turn cost into status.
Meta eventually shut the dashboard down. Business Insider reported in April that the company didn't respond to a request for comment on the episode, but the broader point doesn't need much decoding. If you reward consumption, people consume. Engineers are smart enough to game any metric you put in front of them, especially when the metric is visible, competitive, and detached from shipped work.
The bill is now part of management
The number that should make you stop is 60 trillion. Business Insider reported, citing The Information, that Meta used 60 trillion tokens in 30 days before shutting down the Claudeonomics dashboard in early April. Tokens are the unit AI companies use to price model input and output, but for managers they are becoming something plainer: a budget line with a name attached to it.
Mosseri described tokens as something he now has to allocate alongside GPUs, storage, RAM, and headcount. That is a real change. "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. You don't need to be a CFO to understand the implication. An engineer can now come with a second payroll cost, one paid to Anthropic, OpenAI, Google, or whoever is running the model behind the workflow.
That doesn't mean AI coding tools are useless. Don't confuse cost control with rejection. The hard problem is that agentic coding can burn through tokens quietly, especially when tools keep reading files, rewriting code, testing, failing, and trying again while the user watches a progress bar. A human sees one task. The bill sees every step.
Frankly, the leaderboard was the silly part. The spending problem is not silly at all.
Meta is not the only company pulling back
Uber is already living with the same math. Business Insider reported in May that Uber COO Andrew Macdonald said it was getting harder to justify the money spent on AI tokenmaxxing after CTO Praveen Neppalli Naga told The Information that Uber had blown through its 2026 Claude Code budget by spring. Macdonald said the link between rising token use and more useful consumer features was not clear yet. That is the line every company using these tools has to face.
Amazon had its own version of the problem. At Bloomberg Tech in June, Anthropic president Daniela Amodei noted that Anthropic doesn't run a token leaderboard, while Business Insider reported that Amazon had shut down an informal employee-made one. Dave Treadwell, an Amazon senior vice president, told staff: "Please don't use AI just for the sake of using AI." You can hear the budget meeting behind that sentence.
Even the companies building the rails are rationing access. The Wall Street Journal reported last month that corporate AI users, including Uber, Meta, Microsoft, Salesforce, and DoorDash, are steering employees toward cheaper tools and tracking usage more closely - all in an effort to connect token spending to actual output. Google said it was processing more than 3.2 quadrillion tokens a month, seven times as much as a year earlier, according to the same report. Usage is exploding. Discipline is arriving late.
Meta's own infrastructure spending makes the contrast sharper. AP reported in May that Meta raised its 2026 capital expenditure forecast to between $125 billion and $145 billion, largely because of AI data centers and component costs. Axios reported this month that Meta has already spent more than $200 billion on AI infrastructure and committed another $600 billion through 2028. Business Insider also reported this week that Meta expanded its Hyperion AI data center project in Richland Parish, Louisiana, to more than $50 billion.
It is real money. And yet Mosseri is still talking about caps.
That gap is the story. Meta can spend at a scale almost no other company can copy, and it still has to ask whether every engineer should get unlimited access to expensive frontier models. If you're running an AI-native engineering org, you don't get to ignore that question because the demo looks good. You have to decide which work deserves the best model, which work can use a cheaper one, and which work shouldn't be handed to an agent at all.
That's the job now. The optimistic version is that model prices fall, open models improve, and internal tools like Meta's own coding assistants absorb more of the load. Mosseri said he expects a pricing war among frontier model providers, even as total usage rises. That may happen. But budgets are made in the present tense, not in a future pricing curve.
For now, Meta's lesson is blunt: AI adoption is no longer the trophy. Useful output is. The company built a scoreboard for burning tokens, watched people behave exactly as the scoreboard encouraged, and then had to turn the lights off. If Meta has to ration the machine, you probably do too.
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, and StepFun Unveils the First Agentic AI Phone Ahead of Apple and OpenAI
