An employee at the company Slash reportedly spent $80,000 of company AI credits “vibe-coding” a video game called “brainrot shooter.” Various accounts describe it as not a good game; however, Slash’s response was to ask the internet to play it anyway so the company could write the whole episode off as a “marketing expense”.
This follows a sequence of companies paying for the (rather steep) unexpected costs of AI in the workplace. But, as we are all beginning to find out, these are not minor mistakes by reckless individuals. What is playing out here shows what happens after companies push employees to use AI so much that the AI itself becomes the full-time job.
Welcome to the Tokenmaxxing Office
The term “tokenmaxxing” is one people in the AI world are beginning to become more and more familiar with. In simple terms, it means burning through AI tokens (the metered units of text an AI model processes) because the volume itself became a metric that matters.
For most of 2026, that metric was treated as a proxy for innovation. Meta, Amazon, JPMorgan, and Disney all reportedly built internal leaderboards ranking employees by how many tokens they consumed.
At the time, the logic behind these initiatives made sense: ‘if more AI signals more productivity, we should reward the people who use it the most’.
This logic also assumed employees would use that incentive to do better at work. Yet, as we keep on seeing, it does not. Employees are using it to do the easiest work imaginable, as often as possible, until the number goes up.
A Field Guide to the Office Tokenmaxxer: Cache Wizard Status
Meta’s Tokenmaxxing leaderboard reportedly came with its own internal honours system, including a title awarded to the platform’s top consumer: Cache Wizard.
The leaderboard’s biggest user averaged 281 billion tokens a month, at a compute cost in the hundreds of thousands of dollars, and there is no public record of what any of that consumption actually produced.
Accenture has been quietly trying to walk the same habit back. According to 404 Media, the firm’s head of AI strategy told staff internally that the company’s heaviest token users were not its engineers, after discovering employees were burning credits on tasks as mundane as converting PDFs into PowerPoint slides.
The most candid account, though, comes buried inside Glean’s Work AI Institute survey of 6,000 workers. One engineer explained that he asks AI tools questions whose answers already sit in the company’s documentation, because the model processes the request slowly and burns plenty of tokens along the way. He prototypes features he has no intention of building, then deletes them. He is conscious, in his own words, of not wanting to be seen as someone who “uses too little AI.” Every one of those habits is a deliberate waste, chosen because the alternative looked worse on a dashboard than wasting money did.
The Inevitability of Tokenmaxxing
There is a reasonable defence of usage tracking, and it is worth taking seriously before dismissing the whole exercise as corporate theatre.
Todd Olson, Pendo’s chief executive, made the point plainly in a recent interview. He explained that an employee spending zero tokens is, almost by definition, getting zero value from the tools the company paid for. Some signal beats no signal at all, and a business with thousands of employees needs some way to know whether an expensive rollout is being used.
The trouble is what happens the moment that signal becomes the target rather than a proxy for one. Amazon found this out when staff began deploying AI agents to run pointless tasks purely to climb its internal rankings, then quietly shut the leaderboard down. Meta also did the same. Both companies discovered that once token count became a target, the only thing it reliably tracked was how willing an employee was to waste the company’s money in full view of management.
Unpaid Labour, By Design
Michael Burry, the investor who shorted the 2008 mortgage market, has been arguing for months that tokenmaxxing is not really about productivity at all. In his Substack posts, he explains:
“Tokenmaxxing is not merely heavy AI use, and it is certainly not sustainable AI use. It is quota-driven, leaderboard-driven, management-mandated overconsumption.”
There are also other costs. Every prompt an employee fires off, including the pointless ones, generates usage data and interaction traces that flow back to the model providers and the companies building proprietary tools on top of them. Employees told to use AI as much as possible were not just padding a dashboard. Many were unknowingly producing the training signal and usage volume that made the AI industry’s own growth numbers look stronger heading into a wave of IPOs. None of them were paid for that work as labour. It showed up on a leaderboard instead.
Tokenmaxxing Is Funny, Until It’s Not
There is something genuinely funny about a company writing off a video game nobody wanted to play as a marketing expense. But, in the end, somebody will always end up paying for the bit.
The compute those leaderboards ran on didn’t come from nowhere, it came from budgets that could have gone toward headcount, toward infrastructure, toward literally anything with an output. The employees performing usage weren’t having fun; they were doing invisible, unpaid work to protect themselves from a dashboard. And every one of those pointless prompts fed into the usage numbers the AI industry was pointing to as proof of real demand, right as it walked into a wave of IPOs.
The same companies that built these leaderboards are the ones telling investors, and laid-off staff, that AI is why headcount had to shrink. Meta cut about 8,000 jobs in May, while moving 7,000 more employees into AI-focused roles the company said would make them more productive. Cisco cut nearly 4,000 jobs the same month. Cloudflare cut 20% of its staff, about 1,100 people, despite reporting its highest-revenue quarter on record. Oracle has cut 21,000 jobs this year alone, with more reportedly on the way.
None of those companies is telling employees to burn money proving they use AI enough. That’s an option reserved for the ones still on the payroll. The joke was never free. Somebody was always going to be handed the bill, it just turned out not to be the people who ran up the tab.
Author: Grace Sharp
See Also:
Y Combinator Built Sam Altman’s Empire, Now He’s Tokenmaxxing It
AI Risk Is Moving Faster Than Most Businesses Can Handle
Company Blew $500M on Claude Because Nobody Set a Spending Limit
