Meta is applying the brakes to its artificial intelligence ambitions. After months of aggressively encouraging staff to utilize AI tools, the social media behemoth is now moving to restrict internal AI token consumption and address a crisis triggered by a radical reorganization, with runaway costs and collapsing employee morale simultaneously pressuring management.
According to an internal memo reviewed by The Information, Meta informed approximately 6,000 employees this week that its internal AI usage alone is projected to cost tens of billions of dollars by 2026. The company plans to formally implement a token management system centered on budgets and quotas in 2027. Concurrently, Chief Executive Mark Zuckerberg acknowledged in a separate internal memo that the company "made mistakes" in using AI to drive team restructuring and pledged to provide "meaningful roles" for affected employees.
The release of these two memos highlights the dual pressures Meta faces in its AI transition. On one front, internal AI costs are rising exponentially. On another, employee discontent stemming from the aggressive restructuring has reached a breaking point, with one worker publicly venting frustration using profanity during a live, company-wide meeting attended by thousands, laying bare the internal discord.
These developments have drawn market attention. AI researcher Gary Marcus noted that "tokenmaxxing is giving way to tokenminimizing," predicting this trend will result in third-quarter revenue for Anthropic and OpenAI falling short of their second-quarter performance. For Meta, finding a balance between cost control and retaining AI talent has become its most pressing management challenge.
Unchecked Internal Usage Leads to Astronomical Token Consumption
Meta's internal AI usage has expanded at a pace far exceeding expectations. In April, data from an internal leaderboard dubbed "Claudeonomics" showed employees consumed 60.2 trillion tokens over a 30-day period, a figure that later climbed to 73.7 trillion. Named after Anthropic's flagship product, the leaderboard tracked AI usage among over 85,000 employees, ranking the top 250 "super users." The single highest user consumed 281 billion tokens in 30 days, which, based on Anthropic's public pricing, could translate to a cost in the millions of dollars.
This leaderboard fostered a phenomenon known as "tokenmaxxing," where employees competed to inflate token usage to showcase their AI proficiency. Some even instructed AI agents to run multiple tasks in parallel, artificially amplifying consumption. Employees received gamified incentives through badges like Bronze, Silver, Gold, Platinum, and Jade, as well as titles such as "Session Immortal" and "Token Legend."
Meta's Chief Technology Officer, Andrew Bosworth, warned in April that "no one should be using AI for the sake of using AI," emphasizing that "token usage is not a measure of impact in any sense." The company subsequently shut down the Claudeonomics leaderboard. The recent internal memo further reveals that Meta is building a central dashboard called "AI Gateway" to monitor employee AI usage and spending in real-time. It will also implement automated alerts for anomalous consumption and track current costs to forecast future expenditures, aiding in computing resource planning and supplier negotiations.
Shifting to In-House Tools to Reduce External Vendor Reliance
Another path to cost control involves steering employees toward Meta's internally developed AI tools. The memo indicates Meta plans to guide staff away from third-party AI programming tools, particularly Anthropic's Claude, and toward the company's own coding assistant, MetaCode (formerly Devmate).
Reportedly, Meta's newly formed Applied AI Engineering (AAI) department has assigned engineers specifically to enhance MetaCode's capabilities. Methods include generating high-quality reinforcement learning data by having MetaCode repeatedly solve programming challenge questions to train its coding response abilities. The company stated it will still allow employee access to third-party AI models.
Meta faces dual financial pressures. On one hand, the company plans capital expenditures as high as $145 billion this year, partly to expand data centers, AI chips, and talent. On the other, investors are continuously pressuring the company to show returns on its massive AI investments. Meta has already introduced paid subscription tiers on Facebook, Instagram, and WhatsApp and signaled plans to charge businesses using its AI commercial agents. In this context, the strategic value of reducing internal operational costs is becoming increasingly prominent. Notably, Meta is not alone; reports indicate Uber and ServiceNow exhausted their annual budgets for Anthropic tools within the first few months of 2026, and several venture capital firms have also imposed caps on employee AI usage due to daily token costs reaching thousands of dollars.
Forced Reassignments Ignite Internal Crisis, Prompting Public Employee Protest
Beyond cost issues, a larger internal crisis at Meta stems from the AI-driven organizational restructuring. The Applied AI department, established in March 2026 with about 6,500 engineers and product managers, saw a significant number of employees forcibly reassigned with little warning. In May, Meta cited its AI transition as the reason for cutting approximately 8,000 jobs, with another 7,000 people moved into new AI-related projects.
The primary work for many of these reassigned engineers now involves generating puzzles, writing programming challenges, and completing model test evaluations to provide data for AI model training. For engineers accustomed to product development and feature launches, this shift is widely perceived as a career demotion. One current employee described the situation:
"You suddenly have no life purpose, you barely talk to anyone, and you just repeat these tasks mechanically every week."
Another employee stated bluntly:
"Most people find this kind of work suffocating."
An excessively flat organizational structure has further exacerbated tensions. Reports indicate that in some Applied AI teams, each manager oversees around 50 direct reports, leading to a lack of support for employees, unclear promotion pathways, and difficulty gaining managerial visibility. The pent-up frustration erupted during a live all-hands meeting this week when an attendee, losing emotional control, interrupted a speaker with profanity, demanding that those present relay criticism to a specific AI executive, calling the individual "an asshole." Previously, over 1,600 Meta employees had signed a petition calling for the halt of an internal project collecting AI training data by recording U.S. employees' mouse clicks, keystrokes, and screen activity. Meta later scaled back the project slightly under pressure.
Zuckerberg Acknowledges Errors as Management Scrambles for Repairs
Faced with the escalating internal crisis, Meta's top executives have made public statements. Instagram's Chief Product Officer, Chris Cox, described the past few months at the all-hands meeting as a "hard" and "brutal" period, comparing employees' situation to "running a marathon in a hailstorm, having your teammates swapped out mid-race, while someone is filming the whole thing." He offered an unusually measured assessment of AI itself:
"It's not a god, and it's not a demon. It's not as good as you think, and it's not as bad as you think."
Zuckerberg's language in the internal memo was more direct: "Given the complexity of these adjustments, we made mistakes." He promised to provide "as much stability as possible," announced a large-scale hackathon for July, and began adjusting the Applied AI department's management structure. According to Reuters, Zuckerberg also indicated he does not expect further company-wide layoffs this year.
Analysts note that these statements show Meta's leadership recognizes that this round of restructuring is posing a substantive threat to its talent pool. Engineering talent is the scarcest resource in the current AI race. If core employees continue to feel marginalized, the risk of their departure could create significant repercussions at critical competitive junctures. Whether the current repair measures can genuinely stabilize morale remains to be seen.
Comments