Meta caps AI token use as internal costs hit billions

Meta is imposing token usage limits on roughly 6,000 employees after internal AI costs spiraled to billions of dollars, exposing a widening gap between AI adoption and its economics.

"Nobody should be using AI just for the sake of using AI," Meta Chief Technology Officer Andrew Bosworth wrote in an April memo obtained by The Information, adding that "token usage itself is not a measure of impact."

Employees consumed 73.7 trillion tokens in a single 30-day period, driven by a phenomenon called "tokenmaxxing" — workers deliberately running multiple AI tasks simultaneously to climb an internal leaderboard called "Claudeonomics" that ranked the top 250 users by token consumption. Meta took the leaderboard offline after the surge.

The cost crisis at one of AI's biggest spenders — Meta allocated $145 billion in annual capital expenditure partly for AI infrastructure — raises a fundamental question for the industry: if the companies building AI cannot afford their own token bills, what does that mean for the profit margins of model providers like OpenAI and Anthropic?

The Tokenmaxxing Problem

The internal cost explosion traces back to a November policy shift in which Meta told employees that demonstrating "AI-driven work results" would be a core performance requirement for 2026, with top performers receiving bonuses. The incentive backfired. Instead of using AI selectively, some employees began competing on the "Claudeonomics" leaderboard, which tracked token consumption. One internal dataset showed consumption rose from 60.2 trillion tokens in a 30-day window to 73.7 trillion before the company pulled the ranking.

Meta is now building a central dashboard called "AI Gateway" to monitor company-wide AI usage and spending in real time, with automatic alerts for abnormal consumption spikes. The company plans to roll out the tool to a broader employee base in the coming weeks and implement structured token budget allocations by 2027. It is also pushing workers toward its internal coding assistant, MetaCode, to reduce reliance on Anthropic's Claude, which has become the primary coding tool for Meta engineers.

Industry-Wide Cost Squeeze

Meta is not alone. Amazon shut down an internal AI leaderboard last month after employees performed unnecessary operations to boost their scores, driving up compute costs. Uber and ServiceNow exhausted their full-year Anthropic tool budgets within the first few months of 2026, according to The Information. ServiceNow now monitors daily per-employee usage to track and contain costs. Venture capital firms are also setting AI usage caps for their teams, with daily token bills reaching thousands of dollars.

The spending discipline is rippling through the broader AI market. The LLM Token Spending Index, which tracks the average price paid per 1 million tokens across the market, fell for seven consecutive trading days through June 11 — the longest losing streak since January. The index had more than doubled since December before peaking in May and reversing sharply.

OpenAI is considering cutting token prices to win enterprise clients ahead of its confidential IPO filing this week, the Wall Street Journal reported. Chief Executive Officer Sam Altman has called AI usage costs "a huge problem" and said the company would "help people get more value for less spending." Any price cuts, while competitive, would directly pressure the margins of both OpenAI and Anthropic, which are each losing billions of dollars on the computing power required to run their AI systems.

What This Means for Investors

The shift from a "more tokens is better" growth narrative to a cost-constrained reality has implications across the AI value chain. Morgan Stanley has described the current pullback in token pricing as a "speed bump," while Citadel has argued that the binding constraint on AI adoption has moved from model capability to cost and scarcity, with users accelerating their shift to cheaper models.

For model providers like OpenAI and Anthropic, the pressure is twofold: their largest customers are capping usage even as the providers themselves face the need to cut prices to maintain market share. For hardware suppliers like Nvidia, whose data center revenue depends on expanding AI compute demand, a sustained slowdown in token consumption growth could challenge the capital expenditure expectations baked into current valuations. Meta shares, trading at roughly 22 times forward earnings, have yet to reflect the full cost of the internal AI spending the company is now trying to contain.

This article is for informational purposes only and does not constitute investment advice.