If you have not heard the term tokenmaxxing, you will. It describes the practice of maximising AI token consumption inside an organisation, not because the work requires it but because usage itself became a measure of productivity, status and corporate commitment to the AI revolution.
It started as enthusiasm. It ended as a line item that nobody could justify.
This is a token
A token is the basic unit of AI usage. Every time you send a prompt to ChatGPT, Claude or any other large language model, the input is broken into tokens, roughly three-quarters of a word each. The model processes the tokens and generates a response, which also consists of tokens. The provider charges by the token, or bundles them into a subscription.
At individual scale, the cost is modest. At enterprise scale, with thousands of employees using AI tools across coding, writing, analysis and internal workflows, token consumption becomes a significant budget line.
How all happened
The culture emerged from a reasonable starting point. Companies wanted employees to adopt AI tools. Managers wanted to see that expensive subscriptions were being used. Some companies, including Amazon, built internal leaderboards that tracked and ranked employees by AI usage.
The incentive structure was predictable. If usage is measured and rewarded, people use more. If the leaderboard is visible, people compete to be at the top. If being at the top signals that you are an engaged, AI-forward employee, the incentive to maximise tokens becomes self-reinforcing.
Amazon built Kiro Rank, which scored and ranked employees based on how much they used its in-house AI coding tool, Kiro. Employees began running unnecessary tasks to inflate their scores, increasing the company's compute costs without producing additional value. Amazon shut the leaderboard down in late May after discovering the gaming behaviour.
Cost problem: Ooh ouch #1
Uber burned through its entire 2026 AI token budget in the first four months of the year, partly driven by heavy use of Claude Code. The company introduced monthly caps of $1,500 per employee. Microsoft cancelled Claude Code subscriptions for employees in several product divisions after internal costs escalated.
An engineer at a large corporation may consume 210 billion tokens per week. At enterprise scale, companies are spending more than $150,000 per month on AI operations. The price per token has fallen, but total spending has risen because usage is growing faster than efficiency gains.
This is the tokenmaxxing paradox. Cheaper AI does not reduce the bill. It increases consumption, which increases the bill. The cost curve bends the wrong way.
Oooh, ouch#2
Tokenmaxxing revealed something uncomfortable about how companies adopted AI. Usage was treated as a proxy for productivity. Leaderboards rewarded activity rather than output. Nobody asked whether the tokens consumed produced work that justified the cos
Uber's COO Andrew Macdonald said on a podcast that the technical research team was using AI coding services but that "the productivity gains have not been clearly apparent." Management has begun discussing how to manage token costs.
That admission, from one of the most AI-forward companies in the world, suggests the problem is not isolated. If Uber, Amazon and Microsoft are struggling to demonstrate return on AI spending, smaller companies with fewer resources to measure impact are likely in worse shape.
Tokenmaxxing era is ending
Leaderboards are being dismantled. Budgets are being capped. Companies are shifting from measuring usage to measuring output, which is harder but more honest.
Nature Machine Intelligence published an editorial in May calling on companies and researchers to "stop tokenmaxxing and deploy AI sensibly instead." The piece warned that the frenzy to embed AI into every workflow was creating waste, not value.
The correction was inevitable. AI tools are useful. Treating their consumption as a performance metric is not. The companies that survive the tokenmaxxing hangover will be the ones that figured out what the tokens were actually producing, rather than how many they could burn through before someone noticed the bill
