Why Companies Cannot Turn AI into Real Value: Silicon Valley's Newest Craze and the Tokenmaxxing Trap

When American factories switched to electricity at the start of the twentieth century, something strange happened: productivity did not rise the way everyone expected. Factories swapped gas lamps for electric bulbs and steam engines for electric motors, and output stayed flat, because the layout of the factory still belonged to the age of steam. The real transformation arrived twenty years later, when manufacturers redesigned the whole building around what electricity made possible.

We are standing at the same threshold with AI. Over the past months a debate that started in Silicon Valley and reached the pages of Forbes and the Wall Street Journal describes exactly how we are crossing it.

Its name: tokenmaxxing.

A token is the smallest unit an AI model uses to read and write text, roughly three quarters of a word. When you type a question, the model breaks it into tokens; when it answers, it strings tokens together again. It works like a taxi meter: the counter runs constantly, turning for every word read and every word written, and the invoice at the end of the month is the sum of that counter.

Tokenmaxxing means pushing that consumption to its limit. Writing as many prompts as possible, holding conversations as long as possible, handing off as many tasks as possible to AI. And tracking it not as a personal habit but as a corporate performance indicator.

The "-maxxing" suffix comes from Gen Z internet culture. Like looksmaxxing or sleepmaxxing: optimising something to its absolute limit. Tokenmaxxing is precisely that, pushing AI usage as far as it will go.

The logic looks simple. Whoever uses more, learns more. Whoever experiments more, discovers more. But is that actually true?

This story started at Meta. Last April the publication The Information revealed that an employee had set up a secret leaderboard inside the company. It was called Claudeonomics, after Anthropic's Claude model. More than 85,000 Meta employees were ranked by the volume of tokens they consumed. Those at the top earned titles such as "Token Legend" and "Session Immortal".

The numbers were staggering. Meta employees consumed roughly 60 trillion tokens in 30 days, and the person in first place spent 281 billion tokens on their own, many times more than an ordinary user could get through in a lifetime.

Meta was not alone. At Google I/O in May 2026, CEO Sundar Pichai announced that the company was processing 3.2 quadrillion tokens a month. A year earlier that figure had been 480 trillion. Midway through his talk, Pichai said:

"Some might call this tokenmaxxing. They are probably right."

Uber's disclosure landed loudly: the company burned through its entire 2026 AI tooling budget in four months.

Nvidia CEO Jensen Huang raised the bar further:

"At Nvidia we have engineers on 500,000 dollars a year. At the end of the year I am going to ask them how many dollars of tokens they spent. If the answer is something like 5,000 dollars, that would drive me crazy. If an engineer of that calibre has not consumed at least 250,000 dollars of tokens, I would be seriously worried."

Investors read these sentences as a sign of success. Some companies turned tokenmaxxing into a performance review criterion.

I am not an economist, but I know one old and powerful rule from economics: the moment a measure becomes a target, it stops being a good measure. It is called Goodhart's Law. And tokenmaxxing has become a textbook-clean example of it.

Here is a short summary of what followed.

At Amazon, employees began having AI agents perform meaningless tasks purely to inflate their numbers. One engineer wrote looping bots that called the AI over Slack. Their only purpose was to keep the counter spinning. Amazon shut the leaderboard down soon after. Senior Vice President Dave Treadwell told employees:

"Don't use AI just to use AI."

Meta removed Claudeonomics too. Microsoft cancelled some Claude Code subscriptions. Uber's operations lead Andrew Macdonald openly admitted they could find no meaningful connection between token spend and company productivity.

At the end of May 2026 Fortune ran the headline: "Tokenmaxxing is over."

Let us look at the numbers a little more closely, because the answer here will feel very familiar to anyone in education.

Stanford's Digital Economy Lab published a study in 2026. The researchers showed that token consumption for the same software task can vary by up to 30 times between models and between attempts. And spending more tokens does not mean a more accurate result. Accuracy usually peaks somewhere in the middle, then starts to fall.

Data from the software analytics company Jellyfish points the same way. The engineers who use AI most cost roughly 10 times more than those who use it least, while producing only twice the output.

Let us pause here, because this pattern is not unfamiliar to us at all.

Measuring a training programme's success by hours attended. Judging a student's development by pages read. Assessing a teacher's impact by the number of activities run. All the same trap. Activity is not an outcome, and sometimes it is not even a symptom of one.

The investor Michael Burry, the "Big Short" who saw the 2008 crisis coming, described tokenmaxxing this way:

"Quota-driven, leaderboard-driven, management-mandated overconsumption."

It matters to understand both sides of the argument.

Those defending tokenmaxxing say:

AI adoption is existential. If employees do not use the technology, the company falls behind.

There is no discovery without consumption. We cannot learn what AI is capable of without pushing it to its limits.

Token consumption is at least a measurable indicator. Better than measuring nothing.

Those against point somewhere else:

Activity is not output. Measuring engineers by token consumption is as wrong as measuring a sales team by the number of calls made.

Tokenmaxxing serves the AI companies. Anthropic's revenue doubled and OpenAI's usage figures grew fivefold. The question of who the consumer is consuming for is still standing.

When a measure becomes a target, people find a way to game the system.

HubSpot CEO Yamini Rangan summed up the debate in a single sentence:

"Outcome maxxing matters far more than token maxxing."

Now the real question: how much value is AI use actually producing for companies?

McKinsey's 2025 State of AI report paints a striking picture. The share of companies producing a measurable profit impact (EBIT) from AI is only around six percent. The clearest thing these "high performers" have in common is not that they consume more AI; it is that they have fundamentally redesigned individual workflows.

According to McKinsey's data, 55 percent of high performers have redesigned their workflows in this way. Among the rest, the figure sits at just 20 percent.

MIT's NANDA research group makes it starker still in The GenAI Divide. Only 5 percent of corporate AI pilots produced a measurable financial impact. The other 95 percent stalled at the experiment stage, against total corporate AI investment of some 30 to 40 billion dollars.

So companies are spending money. Employees are consuming tokens. Leaderboards keep spinning. But value? Value is appearing in very few places, in very few companies.

What does the rise and fall of tokenmaxxing teach us? Not simply that a Silicon Valley fashion has ended, but which questions will shape the next phase of our relationship with AI.

Outcomes will be measured, not activity. The "activity" metrics that track AI usage will give way to concrete result metrics. Not "we entered 50,000 prompts this week", but teams who can say "our weekly reporting went from six hours to an hour and a half, and we moved the time we gained into customer analysis".

The right model will be chosen for the right job. The AI companies themselves are turning in this direction. New-generation AI systems first classify the incoming question: a light, cheap model for a simple query, a powerful, expensive one for complex analysis. Using the strongest and most expensive model to draft a simple email is no longer a deliberate choice; it is waste.

AI literacy will move beyond "how to use it". Learning to write prompts is necessary but not sufficient. The real skill is being able to evaluate AI output critically, knowing when to use it and when not to, and taking responsibility for the results.

Whoever redesigns their workflows will win. Just like the manufacturers a century ago who saw no productivity gain until they redesigned their factories around electricity, and then rose to lead once they had. Adding AI to existing processes is not enough. The real question is: "If we designed this work from scratch, with AI, how would we do it?"

Using AI a great deal is not a strategy. Using AI well is a strategy.

The tokenmaxxing story is a new version of a very old one. Gartner's Hype Cycle says a similar pattern repeats with every wave of new technology: a trigger, inflated expectations, disillusionment, then enlightenment and productivity. With AI we are still at the peak of expectations. Excitement and overuse are running side by side. But the first signs of disillusionment are showing too. The real question is whether we can turn that disillusionment into maturity.

The answer to that lies not with the AI companies but in our hands. Educators, managers, employees, students. Learning to see AI not as an instrument of display but as a partner that produces meaning. Being able to tell activity from outcome, usage from value, and spectacle from genuine transformation.

That is the part that takes both skill and courage.