Token-maximization turns AI adoption into a costly contest

How token-maximization works

Tokens are the small pieces of text that models process when receiving a prompt and generating a response. OpenAI, Anthropic and Google generally bill companies according to the number of tokens they use.

That makes tokens convenient for tracking costs. The mistake is treating volume as a measure of AI adoption, employee interest or performance.

Employees may deliberately consume more tokens to rise in a leaderboard, meet a formal target or prove they are using AI actively. Against the backdrop of billions of dollars in corporate AI investment, a seemingly harmless metric can create substantial expenses.

From Meta’s leaderboard to corporate policy

The term’s exact origin is unclear, but many people first encountered it after reports about Claudeonomics at Meta. An employee created the leaderboard to rank colleagues by token consumption.

The list showed the 250 employees with the highest usage. Workers were expected to demonstrate that AI was improving their productivity, or sometimes merely to show that they were using it. The system disappeared quickly after becoming public.

Meta’s episode could have remained an internal joke. Instead, it exposed a broader weakness: companies often choose the metric that is easiest to count before proving that it tracks anything they actually value.

Reports described similar behavior elsewhere:

Nvidia CEO Jensen Huang also supported the approach. He said he would be “deeply concerned” if an engineer earning $500 000 a year had not used at least $250 000 worth of tokens. Nvidia is itself among the main beneficiaries of rising AI spending.

The pattern is not limited to large technology companies. Small businesses usually apply it less visibly, but employees may still be asked to use AI as often as possible to justify labels such as AI-first and AI-native.

Companies also include AI proficiency in job descriptions when it is not obviously necessary for a particular role, then use it as an evaluation criterion.

In roughly a year, token-maximization moved from a strategy label to a warning about inefficient AI use.

The bill for looking productive

AI computing is expensive, especially when companies assign autonomous work to AI agents. Estimates put the cost of deploying agents at 10–100 times that of ordinary chatbots.

2026 has brought wider adoption of AI agents, and many companies are therefore consuming tokens at high speed. Uber is one prominent example: according to reports, it spent its entire annual AI budget in just four months of 2026.

Token-maximization is particularly costly when it is easy to game. It can increase the amount of work being performed without increasing the result.

There is also a security cost. Every deployed AI agent or chatbot creates additional potential problems with permissions and authentication. Meaningless agent activity launched to improve a ranking increases the chance that a poorly implemented system will cause trouble.

The alternative is “value-maximization”: teams using AI report on progress and are held responsible for it. Companies should measure outcomes AI is supposed to affect:

The question companies are avoiding is not how much AI they can make employees use, but whether the work would have been done better without it. I think that distinction will become unavoidable as shareholders, investors and customers demand concrete results rather than evidence that a company has spent heavily on AI.

Token counts are attractive because they produce an immediate number. But a number that measures activity without value turns adoption into theatre — and gives every inefficient system a reason to run longer.