Part 7 · Analysis
Inside Meta's Trillion-Token Obsession
An internal AI leaderboard reportedly reached 73.7 trillion tokens in 30 days. The figure raises a deeper question about AI spending, incentives and real productivity.
Harbour Creative Team9 min read
Harbour Creative Team9 min read
The Battle for AI Power — Part Seven
Reporting reviewed through 12 November 2026.
A leaderboard can tell you who is winning a game. It can also change the way people play it.
Inside Meta, an employee-created dashboard called Claudeonomics offered a peculiar measure of technological enthusiasm: the volume of artificial intelligence tokens consumed by staff. Employees reportedly competed to appear among the company's biggest users, collecting titles that sounded more like achievements in an online game than measures of workplace productivity. In one reported 30-day period, the dashboard reached 73.7 trillion tokens.
The number is remarkable. But its meaning is less straightforward than it first appears. Tokens measure units of text or other model input and output; they are not a direct measure of work completed, time saved, customer value or money spent. A company can use more tokens because its AI tools are genuinely more useful. It can also consume more because employees repeat tasks, run lengthy automated jobs, or treat visible usage as evidence of ambition.
That tension turns Meta's internal experiment into a useful study of the AI industry's newest management problem. Companies are encouraging their employees to make AI central to their work, often before they have settled on sensible ways to measure the returns.
How the leaderboard worked
The Information reported that Claudeonomics drew on usage information from more than 85,000 Meta employees and displayed the 250 most intensive users. Staff could pursue badges including 'Session Immortal' and 'Token Legend.' The practice became known as tokenmaxxing: an effort to demonstrate unusually high AI usage, sometimes for status rather than demonstrable output.
According to The Information, an earlier snapshot showed 60.2 trillion tokens consumed across a 30-day period. About a week later, the figure had reached 73.7 trillion for the preceding 30 days. These were rolling 30-day totals, not necessarily counts of tokens used during a single named calendar month. The dashboard was subsequently taken offline.
A Meta spokesperson told The Information that the employee who created the leaderboard had removed it, rather than the company directing its removal. That distinction is easy to miss, but material. Similarly, the reported existence of a ranking does not prove that Meta officially used token counts to evaluate employees. The company said its performance assessments focused on the impact people delivered and that it did not advocate treating individual token use as the main productivity metric.
The difficulty was the signal sent by the broader culture. When organisations call for employees to become more AI-native, celebrate ambitious experimentation and reward visible output, employees naturally seek proof that they are participating. A public ranking offers proof that is instantly understandable, even if it measures the wrong thing.
Seventy-three trillion tokens is a lot. It is not a bill.
A token is a unit processed by an AI model. It might be part of a prompt, part of an answer, a chunk of a document retrieved by an agent or one of many steps in an automated coding workflow. Different models have different prices; input and output tokens are often priced separately; cached input can be cheaper; internal systems may carry costs that do not resemble published API prices.
One tempting calculation is to multiply 73.7 trillion tokens by $3 per million tokens. The result is about $221.1 million for a 30-day period. Multiplying that by 12 gives roughly $2.65 billion. The arithmetic is correct as a hypothetical scenario. It is not proof that Meta paid $221 million that month or spends $2.65 billion each year on this activity.
There are at least five reasons the real cost could be substantially different. The token count may mix models with sharply different prices. It may include input, output and cached tokens in unknown proportions. Some traffic may be covered by enterprise arrangements rather than standard listed rates. Internal infrastructure has its own economics. And an annualised figure assumes the same workload, prices and operating pattern continue for twelve months.
The Information's reporting raised questions about expensive token consumption, but Meta has not published a public invoice reconciling this leaderboard total to a specific dollar cost. Our headline therefore should not present an illustrative calculation as an audited expense.
The broader financial issue is real even without that dramatic estimate. Meta's own September 2026 quarterly filing explicitly cited third-party AI token costs among factors contributing to higher expenses. It does not isolate those costs in a way that would allow readers to identify the amount attributable to the leaderboard.
Why employees began chasing tokens
In any organisation, people respond to what they believe management values. If leaders say that using AI is important, staff need to decide what effective adoption looks like. A developer might ask an assistant to review code, an analyst might automate repetitive reporting, and a team might deploy an agent to work through a large document archive.
Those can be sensible applications. But when the number of tokens becomes visible and comparable across employees, the incentive changes. A long-running agent can produce more tokens than a short, decisive answer. Multiple overlapping agents can consume resources faster than a carefully chosen workflow. Usage can rise without creating a corresponding increase in quality.
The Information reported instances in which some Meta workers sought to inflate consumption by sending AI agents through numerous tasks in parallel. That is an account of behaviour reported within the company, not evidence that every high-usage employee was wasting resources. Some legitimate engineering workloads are enormous. The challenge is distinguishing them from activity whose main achievement is the consumption itself.
Meta chief technology officer Andrew Bosworth reportedly pushed back against wasteful use. The episode illustrates how quickly an informal internal competition can turn a procurement and infrastructure decision into a cultural phenomenon.
The difference between adoption and productivity
A token leaderboard is appealing because it is objective in a narrow sense. A system can count usage. It cannot, by counting alone, show whether the resulting work was worthwhile.
Consider two employees producing the same software fix. One uses an AI assistant for a brief targeted review, validates the result and finishes in an hour. Another orchestrates several agents, generates hundreds of alternative versions, and needs substantial time to inspect their output. The second employee may consume far more tokens. It does not follow that the second employee created more value.
The reverse can be true too. A complex task may need extensive model reasoning or repeated tests, and its benefits may be greater than the cost. Some AI work cannot be judged fairly by short-term savings. The point is that usage alone cannot settle the question.
A more informative evaluation would look at completed work, quality, error rates, rework, time saved, security risks, and the cost of reaching a reliable result. In software development, that might mean changes that pass tests and remain stable in production. In customer support, it might mean genuinely resolved cases rather than automatically generated responses. In research, it might mean evidence that can be checked rather than pages of confident prose.
This kind of measurement is harder than drawing up a ranking. It asks companies to decide which results matter, and to confront the possibility that some impressive demonstrations of AI capability are not yet productive enough to justify their cost.
The financial backdrop: Zuckerberg's enormous wager
Meta's internal token competition unfolded inside a company making extraordinary investments in AI infrastructure. In its third-quarter 2026 results, Meta reported revenue of $64.6 billion, up 26% from the year-earlier quarter. Capital expenditures, including principal payments on finance leases, were $34.91 billion for the quarter. Its September 2026 filing anticipated approximately $135 billion to $145 billion in capital expenditures across 2026 to support AI initiatives and the core business.
That spending is not a measurement of Claudeonomics or employee token use. Much of it supports data centres, servers, networking and other infrastructure on which Meta's consumer services and AI ambitions depend. Nevertheless, it explains why the economics of AI are no longer a small line item or an isolated engineering concern.
The same quarter's operating margin was 20%, compared with 40% a year earlier. It would be misleading to attribute that decline entirely to AI: Meta recorded a $9.89 billion legal charge connected to an agreement with U.S. state attorneys general. Excluding that exceptional charge changes the comparison materially. Meta's advertising business also remained substantial and growing. That is precisely why its position is so unusual: a hugely profitable existing business is funding a costly attempt to define the next computing platform.
Zuckerberg has framed Meta's long-term efforts around personal superintelligence and new model development. The company's October results referred to Muse and Muse Wonder as milestones. The challenge now is to make those investments productive for employees, compelling for billions of users, and eventually defensible to shareholders.
What the vanished leaderboard reveals
The removal of Claudeonomics did not end the broader question. In fact, the most revealing part of the story may be the gap between the dashboard's appeal and its usefulness. A ranking that made AI adoption feel exciting also made it possible to confuse consumption with contribution.
For a company with Meta's resources, experimentation on a large scale can be rational. Failed trials teach engineers what works, and not every valuable innovation can be measured instantly. But there is a difference between paying for experimentation and rewarding avoidable waste. The greater the spending, the more important that difference becomes.
Other companies face a similar decision, even if their token totals are far smaller. As coding assistants and autonomous agents become ordinary workplace tools, leaders must ask who authorises long-running tasks, what budgets apply, which models are appropriate for which jobs, and how results are independently checked. Without those answers, making AI universal can easily become a spending policy disguised as a productivity policy.
The real question is what all that computing produces
Meta's reported 73.7 trillion tokens should provoke curiosity, not a reflexive conclusion that the company wasted hundreds of millions of dollars. The reported total is evidence of extraordinary consumption, and the internal competition is evidence of the incentives that can grow around a visible metric. Neither alone establishes the size of the bill or the value created.
The next phase of the AI race will be decided partly by hardware and models, but increasingly by how effectively organisations use them. That is a less glamorous contest than a leaderboard. It demands careful engineering, realistic budgets, quality controls and managers willing to reward useful work rather than impressive numbers.
There is an irony in the story of Claudeonomics. AI is supposed to help organisations do more with their resources. Inside one of the companies investing most heavily in that promise, workers briefly celebrated the act of consuming the resource itself.
The number was 73.7 trillion. The question worth asking is still much simpler: what did it achieve?
Sources and further reading
- The Information: Meta employees compete for 'Token Legend' status — original reporting on dashboard and incentives.
- The Information: The tokenmaxxing tide may already be turning — reported 73.7 trillion rolling total, dashboard removal, Meta response.
- The Information: Meta shutters internal AI token leaderboard — spokesperson statement about the removal.
- The Information: Meta moves to curb employee AI usage — Bosworth and cost-control reporting.
- Meta third-quarter 2026 earnings release — company disclosures on results and expenditures.
- Meta September 2026 Form 10-Q — third-party AI token expenses and capex outlook.
About this article
This article provides analysis and interpretation of available information. Sources and supporting material, where applicable, are linked within the article.
Found an error? Contact the editorial desk. See our editorial standards and corrections policy.
From the Harbour Creative Blogs archive
Keep exploring this story
Related reporting, people and places from our published coverage. These are editorial connections, not a popularity ranking.
Company
Nvidia
The chipmaker at the centre of global AI investment: its revenue, products, leadership, customers and commercial relationships.
Company
Meta
The owner of Facebook, Instagram and WhatsApp, and a major investor in AI and computing infrastructure.
Person
Mark Zuckerberg
Meta's founder and chief executive, associated with its social platforms and AI investment.


