Man vs. Token
So. AI was going to replace everyone, remember? Developers were supposed to be the first. Then designers and illustrators. Next up were project managers and agile folks. After that… well, technically you, reading this right now. It was just a matter of time.
And here we are: the bill has arrived. Not the apocalypse bill for human jobs, but the actual bill, the invoice, the little buddy bill. And it’s saltier than the salary of the employee who was laid off in a fifteen-minute call with the camera off.
The paradox no one put in the board presentation
Microsoft has started canceling most of its Claude Code licenses. Not because the tool was bad, quite the contrary; it was too good. They canceled it because engineers adopted it with such enthusiasm that the scale of use became a real financial problem. Microsoft encouraged people to use it, they used it too much, and now they’re backpedaling on the decision.
Uber burned through all its AI tool budget for 2026 in just four months. Four fuc*ing months! The year had barely begun when the company was already out of budget for the very thing they incentivized via internal leaderboards ranking which teams used the most AI.
(This is real. We’re all living in a corporate fanfic of LinkedDisney.)
Amazon is encouraging its employees to “tokenmaxx”—that is, to use as many tokens as possible in their day-to-day.
Meta had a dashboard called “Claudeonomics” to track who was using the model the most.
Bryan Catanzaro, VP of deep learning at Nvidia, said in an interview that, for his team, “the cost of computation far exceeds the cost of employees.”
Wait. What?
The VP of one of the largest chip companies on the planet—the same company that profits absurdly from the AI race—is saying that AI costs more than paying people?
Follow the thread and I’ll explain along the way.
Tokenmaxxing: the productivity that incurs interest
Here’s the structural problem, and it has a cruel elegance that dear old Uncle Nassim Taleb would love to dissect in a new 1,000-page book (for Shiva, no).
When the cost per token decreases, companies don’t save money. They use more. It’s the classic rebound effect: the cheaper gas gets, the more people drive, and total spending increases. Only here, the gas is tokens and the car is an AI agent you set to run on full autonomy while you sleep.
Gartner predicts that by 2030, the cost of inference in a large model will drop nearly 90% from 2025. That sounds great on the PowerPoint slide. The problem is the next slide: agent-based models consume way more tokens per task than standard models; total consumption is expected to grow 24 times by 2030, according to Goldman Sachs, and AI providers won’t fully pass on the cost savings to the customer.
Loose translation: the customer is going to pay more, not less. They’ll just receive a bill with more line items so they don’t notice right away.
Gartner’s senior analyst, Will Sommer, summarized it all with a quote that deserves to be framed in every innovation room in the world:
“Chief Product Officers should not confuse the deflation of commoditized tokens with the democratization of frontier reasoning.”
Nice phrase. Even nicer when translated literally: you’re paying a premium for the hard part, and the cheap part you think you’ll gain won’t make up for it.
The irony every laid-off worker deserved to see
Now comes the karma part.
In 2023 and 2024, a number of tech companies made mass layoffs with the justification—sometimes explicit, sometimes implicit—that AI would make those jobs unnecessary or much more efficient.
Microsoft, Google, Meta, Amazon: tens of thousands of people let go in a period when those companies’ stocks were rising and their margins were perfectly healthy.
The narrative was seductive: why pay a salary with taxes, benefits, vacation, and the existential risk of the person asking for a raise when you could just pay for tokens?
The response now in 2026: because the token doesn’t take vacations.
But neither does it stop spending while you sleep. Because the agent runs in loops and each iteration costs money. Because when you put everyone on a leaderboard to “use more AI” and people take it seriously, you find out that encouraging unchecked consumption is just as smart as it sounds.
The laid-off human was predictable. Costing a fixed monthly amount. Sometimes they complained. Sometimes they were late. Sometimes they spilled coffee on their laptop (that wasn’t me). But when they logged off at 6 PM, they stopped. The AI agent doesn’t stop. And every prompt it generates autonomously has a price.
Jensen Huang, CEO of Nvidia, recently stated that he envisions a day when 100 AI agents will work alongside each employee in the company. Great. Has anyone calculated the bill for those 100 agents running in parallel under the current pricing model?
(Someone has calculated it. That’s why we’re here.)
What happens when hype hits reality
Using AI isn’t the problem. After all, I’m a confessed fan of the subject and I believe AI helps solve a lot of things. What I’m saying is something more disturbing: it works; the problem is the bill.
The difference between promise and delivery isn’t technological. It’s economic. The tools do what they say they do. The business model that has been built around them—fire humans, replace them with tokens, scale infinitely—has hit a wall called “this costs real money.”
On the cancellation of Claude’s usage by Microsoft: Anthropic declined to comment when Fortune magazine asked. Microsoft didn’t comment either, and now they’re rolling out Copilot for employees.
What’s left is a slightly comical scenario of companies that created internal leaderboards to incentivize maximum AI use and are now trying to figure out how to disincentivize that use without appearing to completely contradict themselves.
High-level corporate sport. Famous management geared towards competitions.
The 100x org and the most honest CEO in Silicon Valley
On the same day Fortune published the cost problem at Microsoft, ClickUp’s CEO, Zeb Evans, posted on X one of the most honest—or perhaps disturbing, depending on your point of view—statements a tech CEO has ever made in public:
“Today we reduced headcount by 22%. The business is in the best moment in its history. I made this decision and take full responsibility. I did this because the way to operate with the highest level of productivity is changing.”
Twenty-two percent of ClickUp. In a company valued at $4 billion. That wasn’t going through a financial crisis. That is doing very well, thank you.
The justification is what Evans calls a “100x org”—a company where the best engineers don’t write code, but orchestrate and revise AI agents, multiplying their productivity many times over. For those who remain and achieve what he calls “100x impact,” the company is creating salary brackets that reach $1 million per year in cash.
One million. In cash. Per year.
(Note the new performance criterion for your next 360 evaluation.)
There’s something almost admirable in the brutality of the logic: Evans didn’t use the usual euphemism of “restructuring to better serve our customers” or “strategic resource adjustment.” He outright said that the positions of many people have become structurally obsolete—not because they were bad, but because the model has changed. At least he’s honest.
The problem is the flip side of that coin.
If Microsoft is canceling AI licenses because it’s too expensive, and ClickUp is laying off 22% to redirect resources to those who remain, there’s a math here that doesn’t add up with optimism: someone is paying this bill. It’s either the token or the human. For now, it seems the bill is coming to both.
Evans closed the post with a phrase that will circulate in many tech boards in the coming weeks: “Almost every company will make changes like these. Those that do so proactively will define what comes next.”
Maybe someone will figure out how to properly leverage AI’s potential (no, don’t look at me).
I have no solution for this. That’s not my job here. I just came to confuse you.
But if you were laid off in 2024 with the veiled promise that AI would do your job better and cheaper, and you’re watching this 2026 story where the tokens bill costs more than your salary, at least you can sleep at night without accumulating inference costs.
For now. Neuralink is just around the corner.